PMCID
stringclasses 8
values | Title
stringclasses 8
values | FullText
stringclasses 8
values |
|---|---|---|
PMC2872605
|
Kinase-Dead BRAF and Oncogenic RAS Cooperate to Drive Tumor Progression through CRAF
|
We describe a mechanism of tumorigenesis mediated by kinase-dead BRAF in the presence of oncogenic RAS. We show that drugs that selectively inhibit BRAF drive RAS-dependent BRAF binding to CRAF, CRAF activation, and MEK–ERK signaling. This does not occur when oncogenic BRAF is inhibited, demonstrating that BRAF inhibition per se does not drive pathway activation; it only occurs when BRAF is inhibited in the presence of oncogenic RAS. Kinase-dead BRAF mimics the effects of the BRAF-selective drugs and kinase-dead Braf and oncogenic Ras cooperate to induce melanoma in mice. Our data reveal another paradigm of BRAF-mediated signaling that promotes tumor progression. They highlight the importance of understanding pathway signaling in clinical practice and of genotyping tumors prior to administering BRAF-selective drugs, to identify patients who are likely to respond and also to identify patients who may experience adverse effects.The RAS–ERK (extracellular-signal regulated protein kinase) MAPK (mitogen-activated protein kinase) signaling pathway regulates cell responses to environmental cues (Marshall, 1995) and plays an important role in human cancer (Gray-Schopfer et al., 2007). The pathway comprises the RAS small guanine-nucleotide binding protein and the protein kinases RAF, MEK (mitogen and extracellular-regulated protein kinase kinase), and ERK. RAS is attached to the inner face of the plasma membrane and is activated downstream of growth factor, cytokine, and hormone receptors. Active RAS recruits RAF to the membrane for activation through a complex process involving changes in phosphorylation and binding to other enzymes and scaffold proteins (Kolch, 2000). RAF phosphorylates and activates MEK, which phosphorylates and activates ERK. The complexity of this pathway is increased by the multiplicity of its components. There are three RAS (HRAS, NRAS, and KRAS), three RAF (ARAF, BRAF, and CRAF), two MEK (MEK1 and MEK2), and two ERK (ERK1 and ERK2) genes that encode proteins with nonredundant functions. Furthermore, the pathway is not linear. BRAF binds to and activates CRAF in a RAS-dependent manner that appears to require CRAF transphosphorylation by BRAF (Garnett et al., 2005; Rushworth et al., 2006; Weber et al., 2001), providing subtle pathway regulation that is not fully understood. ERK phosphorylates many substrates and the duration and intensity of its activity affects how cells respond to extracellular signals (Marshall, 1995). Thus, the pathway must be carefully controlled to ensure appropriate responses to environmental cues. In normal cells, outcomes include survival, proliferation, senescence, and differentiation, but in cancer the constitutive pathway activation favors proliferation and survival. RAS–ERK signaling is particularly important in melanoma. Somatic mutations occur in BRAF, NRAS, and KRAS in 43%, 20%, and 2% of melanomas respectively (www.sanger.ac.uk/genetics/CGP/cosmic/). The mutations in RAS trap it in a GTP-bound, active conformation and mostly involve glycine 12 (G12), glycine 13 (G13), and glutamine 61 (Q61). A glutamic acid substitution for the valine at position 600 (BRAF) accounts for over 90% of the mutations in BRAF in cancer. However, over 100 other rare mutations have been described, most of which cluster to the glycine-rich loop and activation segment in the kinase domain. These regions normally trap BRAF in an inactive conformation by forming an atypical intramolecular interaction, and it is thought that the mutations disrupt this interaction, thereby allowing the active conformation to prevail (Wan et al., 2004). Functional studies have shown that most of the mutations in BRAF are activating and enhance its ability to directly phosphorylate MEK (Wan et al., 2004; Garnett and Marais, 2004). Curiously however, some mutants have impaired activity and although they cannot directly phosphorylate MEK, they appear to retain sufficient activity to bind to and transphosphorylate and activate CRAF in a RAS-independent manner (Garnett et al., 2005), allowing these mutants to activate the pathway indirectly through CRAF. More puzzling are mutations that occur at aspartic acid 594 (D594). The carboxy oxygen of this highly conserved residue (the “D” of the DFG motif) plays a critical role in chelating Mg and stabilizing ATP binding in the catalytic site (Johnson et al., 1998). As in other kinases, mutation of this residue causes inactivation and thus cancer mutants such as BRAF cannot phosphorylate MEK, activate CRAF, or stimulate cell signaling (Ikenoue et al., 2003; Wan et al., 2004). These mutants therefore appear catalytically and biologically inactive and yet 34 have been found in human cancer (www.sanger.ac.uk/genetics/CGP/cosmic/). Furthermore, while BRAF mutations (over 10,000 described) occur in a mutually exclusive manner with RAS mutations, four of the 34 kinase-dead mutants are coincident with RAS mutations, a highly significant enrichment (p < 10; Fisher's Exact Test) that suggests functional interaction. It has been shown that BRAF is 500-fold activated, can stimulates constitutive MEK–ERK signaling in cells (Gray-Schopfer et al., 2007) and induce melanoma in mice (Dankort et al., 2009; Dhomen et al., 2009), showing that it can be a founder mutation in melanoma. Importantly, BRAF inhibition blocks melanoma cell proliferation and induces apoptosis in vitro and blocks melanoma xenograft growth in vivo (see Gray-Schopfer et al., 2007). These data validate BRAF as a driver of melanomagenesis and as a therapeutic target in melanoma, so drugs to target this pathway have been developed. The first to be tested clinically were the multi-kinase inhibitor sorafenib and the MEK inhibitor PD184352 (CI1040). Disappointingly, both failed to produce objective responses in patients, either because they were not sufficiently potent, or because they caused unacceptable toxicity (Halilovic and Solit, 2008). Recently, more potent and selective BRAF inhibitors have been described. For example, the triarylimidazole SB590885 and the difluorophenylsulfonamine PLX4720 display excellent selectivity for BRAF in vitro and preferentially inhibit BRAF mutant cancer cell proliferation (King et al., 2006; Tsai et al., 2008). More importantly, BRAF-selective drugs have recently entered the clinic and are producing excellent responses in patients with BRAF mutant melanoma (Flaherty et al., 2009; Schwartz et al., 2009). The aim of this study was to better understand the responses that melanoma cells make to BRAF-selective inhibitors and thereby to provide a molecular basis for the design of clinical trials using BRAF drugs. We also wished to examine if kinase-dead BRAF and oncogenic RAS functionally interact in vivo. We selected four drugs for our studies (Figures S1A–S1D). Sorafenib is a class II (inactive conformation binder) drug (Wan et al., 2004) that inhibits BRAF at 40 nM, CRAF at 13 nM, and several other kinases in the low nM range (Wilhelm et al., 2004). It is the least-selective drug that we used. PLX4720 is a class I (active conformation binder) inhibitor that is highly selective and inhibits BRAF at 13 nM (Tsai et al., 2008). 885-A (Figure S1C) is a close analog of the class I inhibitor SB590885 (King et al., 2006) that is also highly selective for BRAF. It inhibits BRAF at 2 nM (Figure S1E), is ineffective against a panel of 64 other protein kinases (Table S1), and preferentially blocks BRAF mutant cancer cell proliferation (Figure S1F). Finally, we also used the potent and selective MEK inhibitor PD184352 (Sebolt-Leopold et al., 1999). As expected, all four drugs blocked ERK activity in BRAF mutant A375 melanoma cells (Figure 1A; see Table S2). Similarly, all four drugs inhibited ERK in SkMel24, SkMel28, D25, and WM266.4 cells, another four lines that express mutant BRAF (Figure S1G). We also tested the drugs in D04, MM415, MM485, and WM852 NRAS mutant cells (Table S2). As expected, PD184352 and sorafenib inhibited ERK in all of these lines (Figure 1A). Surprisingly, however, PLX4720 and 885-A caused an unexpected increase in ERK activity in the NRAS mutant cells (Figure 1A). NRAS or CRAF depletion by RNA interference (RNAi) blocked MEK/ERK activation by PLX4720 and 885-A in NRAS mutant cells (Figure 1B and 1C) and we show that 885-A activated CRAF in these cells (Figure 1D). We previously reported that oncogenic RAS requires CRAF but not BRAF to activate MEK (Dumaz et al., 2006) and consistent with this, BRAF is inactive in NRAS mutant cells (Figure 1E). These data therefore present an intriguing paradox. BRAF is not active and is not required for MEK/ERK activation in RAS mutant cells. Nevertheless, BRAF inhibitors hyperactivate CRAF and MEK in these cells, so we studied the underlying mechanism(s). Wild-type BRAF binds to CRAF in a RAS-dependent manner and although this binding is weak, it leads to CRAF activation (Garnett et al., 2005). Since RAS and CRAF are required for ERK activation by PLX4720 and 885-A, we investigated if these drugs induce BRAF binding to CRAF. Endogenous BRAF was immunoprecipitated from melanoma cells and western blotted for endogenous CRAF. We show that CRAF did not bind to BRAF in untreated or PD184352 treated WM852, D04, MM415, or MM485 cells (Figure 2A), demonstrating that MEK inhibition does not induce binding. In contrast, sorafenib and 885-A induced strong binding of BRAF to CRAF in all four lines (Figure 2A). We also performed the experiment in the inverse manner, immunoprecipitating CRAF and showing that BRAF binding was strongly induced by sorafenib and 885-A (Figure 2A). Curiously, PLX4720 did not appear to induce BRAF binding to CRAF, but previous studies have shown that ERK phosphorylates BRAF in a negative-feedback loop that destabilizes its binding to CRAF (Rushworth et al., 2006). We show that PD184352 stabilizes BRAF binding to CRAF in the presence of PLX4720 (Figure 2B), demonstrating that PLX4720 does induce binding, albeit less strongly than the other drugs. In addition to inducing BRAF binding to CRAF in NRAS mutant cells, 885-A and sorafenib also induce this binding in WM1791c melanoma cells and in SW620 and HCT116 colorectal carcinoma cells (Figure 2C), all of which express mutant KRAS (Table S2). Importantly, no strong binding of BRAF to CRAF was seen in A375 cells even in the presence of PD184352 and the drugs did not induce strong BRAF binding to CRAF in two other BRAF mutant melanoma cell lines (Figure 2D and Figure S2). Thus, sorafenib, 885-A and PLX4720 all induced BRAF binding to CRAF in NRAS or KRAS mutant cells, but not in BRAF mutant cells, showing that BRAF inhibition per se did not induce this binding; it only occurred when BRAF was inhibited in the presence of oncogenic RAS. To confirm the essential role of RAS, we show that a CRAF mutant (CRAF) that cannot bind to RAS (Fabian et al., 1994) did not bind to BRAF (Figure 3A and Figure S3A) and the corresponding mutant of BRAF (BRAF) did not bind to CRAF (Figure 3B and see Figure S3B). We also prepared membrane/cytosol fractionations of RAS mutant cells and show that under normal conditions over 40% of CRAF is in the membrane, whereas BRAF is largely cytosolic (Figure 3C). Notably, 885-A treatment leads to strong recruitment of BRAF to the membrane fraction, whereas CRAF is only weakly affected (Figure 3C). We also show that under normal conditions, EGF did not induce BRAF binding to CRAF in PMWK cells, a line that is wild-type for BRAF and RAS (Table S2). However, in the presence of 885-A, EGF induced robust binding of BRAF to CRAF in PMWK cells and this resulted in sustained pathway activation (Figure 3D). This shows that BRAF binding to CRAF is induced in the presence of both oncogenic RAS and activated wild-type RAS. We note that sorafenib and 885-A induce a mobility shift in BRAF in SDS-gels (Figure 2A). BRAF also undergoes a mobility shift in PLX4720 treated cells in the presence of PD184352 (Figure 2B). This mobility shift is reduced when immunoprecipitated BRAF is treated with calf intestinal alkaline phosphatase (CIP; Figure 3E) and PD184352 pretreatment reduced, but did not ablate the magnitude of the shift induced by 885-A (Figure 3F). Importantly, in vitro CIP treatment and cell pretreatment with PD184352 did not prevent BRAF binding to CRAF (Figures 3E and 3F). Together, these data suggest that the BRAF bound to CRAF is hyperphosphorylated through MEK–ERK-dependent and MEK–ERK-independent mechanisms, but that this phosphorylation is not required for BRAF binding to CRAF. To test directly if BRAF binding to CRAF is driven by 885-A binding to BRAF, we mutated the so-called “gatekeeper threonine” (T529) of BRAF to asparagine (T529N). Since BRAF is not active in RAS mutant melanoma cells (Figure 1E), we measured BRAF activity using transient expression in COS cells (Wan et al., 2004). The results show that BRAF is still activated by HRAS, NRAS and KRAS (Figure 4A and Figure S4A). Importantly, BRAF is ∼170-fold less sensitive to 885-A than wild-type BRAF (17 nM versus 2869 nM; Figure 4B) and 885-A did not stimulate its binding to CRAF (Figure 4C), proving that drug binding to BRAF drives BRAF binding to CRAF. Next, we expressed a kinase-dead version of BRAF (BRAF) in D04 cells and show that it forms a constitutive complex with CRAF (Figure 4D) and that it activates MEK constitutively (Figure 4E, compare lanes 1, 4, and 7). Notably, 885-A does not further enhance MEK activation driven by BRAF (Figure 4E, compare lanes 4, 6 to 7, 9), presumably because it cannot further inhibit this already inactive kinase. Two other kinase-dead BRAF mutants, the classical catalytic lysine mutant (BRAF), and BRAF, a mutant found in human cancer (Wan et al., 2004), also activate MEK in D04 cells (Figure 4F). Thus, it is BRAF inhibition and not drug binding that drives BRAF binding to CRAF. This experiment also shows that MEK activation driven by kinase-dead BRAF is inhibited by sorafenib (Figures 4E and 4F). Indeed, cell responses to sorafenib appear to be paradoxical. We show that although sorafenib inhibits ERK (Figure 1A), it induces BRAF binding to CRAF (Figure 2A), CRAF activation (Figure 4G) and CRAF phosphorylation on S338 (Figure 4G, inset), a critical event in CRAF activation (Mason et al., 1999). To test directly the role of CRAF in cells when BRAF is inhibited, we mutated its gatekeeper threonine to asparagine (CRAF). Notably, CRAF still binds to BRAF in sorafenib and 885-A treated cells (Figure 4H), demonstrating that drug binding to CRAF is not required for BRAF binding to CRAF. More importantly, in the presence of CRAF, sorafenib activates rather than inhibits the pathway (Figure 4H, compare lanes 3 and 7). We therefore posit that sorafenib induces paradoxical activation of CRAF because it inhibits BRAF and drives CRAF activation, but simultaneously binds to and inhibits CRAF. In agreement with this model, we show that two other pan-RAF inhibitors, ZM336372 and RAF265 also induce BRAF binding to CRAF, but without activating ERK (see Figure S4B). Our data establish that inhibition of BRAF in the presence of oncogenic RAS hyperactivates CRAF, MEK, and ERK. To investigate the consequences of this in vivo, we used conditionally targeted alleles of oncogenic Kras (Kras) and kinase-dead Braf (Braf) in transgenic mice. These alleles use Cre-recombinase/LoxP-Stop-LoxP (LSL) technology to regulate inducible expression of mutant proteins from the endogenous mouse genes to ensure normal levels of protein expression. The Kras allele has been described (Jackson et al., 2001), and we recently developed the Braf allele. Briefly, exon 15 of endogenous Braf was targeted to mutate D594 to alanine (D594A; see Figure 5A). To prevent expression of Braf in all cells, an LSL cassette was inserted between exon 14 and the mutated exon 15. This contains a minigene for exons 15–18 of Braf, a transcription terminator and a Neo selection marker to ensure that only Braf is expressed. Removal of the LSL cassette by Cre-recombinase reveals the mutated exon 15 and Braf is expressed. These mice were crossed to Tyr::CreERT2 mice (Yajima et al., 2006), in which the tyrosinase promoter is used to express tamoxifen-activated Cre-recombinase (CreERT2) in the melanocytes. Since CreERT2 is activated by tamoxifen, this approach provides exquisite spatial and temporal control over Kras and Braf expression. Kras, Braf, and Tyr::CreERT2 mice were crossed to generate Kras;Tyr::CreERT2, Braf;Tyr::CreERT2, or Kras;Braf;Tyr::CreERT2 mice. In all cases, the conditionally targeted alleles were balanced over a corresponding wild-type allele. Mice were treated with tamoxifen at 2–3 months of age to induce mutant protein expression. We have recently shown that in this model, Braf induces skin hyperpigmentation, nevus formation, and melanoma (Dhomen et al., 2009). In contrast, Braf did not induce skin hyperpigmentation, nevi (data not shown) or tumors (Figure 5C). Kras induced weak tail darkening after 5–6 months (Figure 5B) but did not induce either nevi (data not shown) or tumors (Figure 5C). However, when Braf and Kras were combined, they induced a conspicuous skin phenotype. Within 2–3 months the ears (data not shown), tails (Figure 5B), and paws (Figure 5D) darkened visibly. The mice did not develop nevi, but within 6 months, they all developed large, rapidly growing oligo-pigmented tumors (Figures 5C and 5E). The tumors displayed evidence of ulceration (Figure 5F) and were composed largely of spindle cells that exhibit features of malignancy, including cellular atypia, nuclear pleomorphism, and conspicuous nucleoli (Figure 5G). They were highly proliferative as evidenced by large numbers of mitotic figures in the superficial and deep aspects of the lesions (∼6 mitosis/10HPF; Figure 5H) and positive staining for Ki67 throughout (Figure 5I). The tumors were strongly and diffusely positive for S100 (Figure 6A) and expressed the melanocyte markers tyrosinase, Dct, Pax3, and silver (Figure 6B), consistent with a diagnosis of melanoma. Genomic DNA analysis of the tumors and cell lines derived from them confirmed that Braf had been recombined to Braf (Figure 6C). However, for technical reasons we could not detect Kras recombination (data not shown), so used RT-PCR to amplify and sequence Kras mRNA. We show that only wild-type Kras is expressed in the kidneys, whereas the tumors expressed both wild-type Kras and Kras (Figure 6D). Importantly, we show constitutive binding of Braf to Craf in cells from the Kras/Braf tumors (Figure 6E). As a control, we used cells from melanoma induced by Kras overexpression. Briefly, when Kras was overexpressed in melanocytes in mice using the β-actin promoter (β-actin:LSL:Kras; Meuwissen et al., 2001), it induced rapid onset melanoma (median time to onset 2 months, 100% penetrance within 3 months) in the absence of Braf (manuscript submitted). Importantly, in cells from these tumors, Braf does not bind to Craf (Figure 6E). Thus, it is only kinase-dead Braf and not wild-type Braf that binds to Craf in the presence of oncogenic Kras. In this study, we show that inhibition of BRAF by chemical or genetic means in the presence of oncogenic or growth-factor activated RAS induces BRAF binding to CRAF, leading to CRAF hyperactivation and consequently elevated MEK and ERK signaling. The mechanism we describe is another paradigm of RAF activation downstream of RAS and based on our findings, we propose the following mechanism by which this occurs. We posit that in RAS mutant cells, BRAF maintains itself in an inactive conformation through its own kinase activity, either through auto-phosphorylation, or by phosphorylating a partner protein that then keeps it inactive (Figure 7A). We are currently using mass-spectrometry and mutagenic approaches to elucidate the underlying mechanism. We propose that when BRAF is inhibited, it escapes this auto-inhibited state and is recruited to the plasma membrane by RAS, where it forms a stable complex with CRAF. Critically, we posit that because it is inhibited, BRAF does not directly phosphorylate MEK, but rather it acts as a scaffold whose function is to enhance CRAF activation, thereby allowing CRAF to hyperactivate the pathway (Figure 7B). We do not know the stoichiometry of the components in these complexes, but since BRAF and CRAF must both bind to RAS for complex formation, it seems likely that at least two RAS proteins are needed to stimulate formation of the complex (Figure 7B). It is unclear why PLX4720 only induces weak binding of BRAF to CRAF, but this may stem from its unique property of displacing the α-C helix of BRAF when it binds (Tsai et al., 2008) and suggests that this helix is important for BRAF binding to CRAF, something that will only be resolved when the BRAF:CRAF crystal structure is solved. We have attempted to identify other proteins that may be required to stabilize the BRAF–CRAF complexes. Our unpublished mutagenesis data suggests that 14-3-3 is required to stabilize these drug-induced complexes (data not shown) and this is consistent with previous observations demonstrating that 14-3-3 mediates BRAF binding to CRAF (Garnett et al., 2005; Rushworth et al., 2006). Although this appears to contradict our observation that dephosphorylation does not disrupt the complex, because 14-3-3 binds to BRAF and CRAF in a phosphorylation-dependent manner, we presume that 14-3-3 protects these sites from dephosphorylation. We have also used RNAi to examine the potential role of other proteins implicated in BRAF-CRAF complex formation or pathway activation, including the scaffold proteins KSR, Sprouty2 and RKTG and the small G protein RHEB, but our preliminary results have not revealed obvious roles for these proteins. Our studies have parallels to the recently described heterodimers between DRAF and KSR in Drosophila (Rajakulendran et al., 2009). Notably, flies have only one RAF isoform and it appears to be an ortholog of BRAF rather than ARAF or CRAF. Our inability to demonstrate an obvious role for KSR in mediating BRAF binding to CRAF or CRAF activation by BRAF suggests that the mechanism underlying dimerization here may be different from those described in flies, but clearly additional studies are required to investigate further the role of scaffold proteins in mediating the phenomena we report. In contrast to the BRAF-selective inhibitors, the pan-RAF inhibitors appear to induce paradoxical activation of CRAF. They induce BRAF binding to CRAF and CRAF activation, but do not activate MEK–ERK signaling. We posit that this is because these agents target both BRAF and CRAF. Thus, although their inhibition of BRAF will stimulate CRAF activation, they will simultaneously inhibit CRAF (Figure 7C). This model is supported by our observation that CRAF converts sorafenib from a pathway inhibitor to a pathway activator and we argue that the paradoxical activation of CRAF by these inhibitors is mediated by BRAF, rather than disrupted feedback inhibition as previously suggested (Hall-Jackson et al., 1999). Recently, paradoxical activation of PKB/AKT and PKCɛ was also described (Cameron et al., 2009; Okuzumi et al., 2009). While ATP-competitive inhibition can block kinase function, they do not block the upstream events that activate the target kinase. For instance, PKB/AKT inhibitors block the function of this kinase, but occupation of the ATP-pocket by these inhibitors was sufficient to induce the priming phosphorylation usually required for its full activation (Okuzumi et al., 2009). Inhibitor binding to PKCɛ has been shown to have a similar effect (Cameron et al., 2009). Importantly, the paradoxical activation of PKB/AKT and PKCɛ did not result in pathway activation because of the continued presence of the inhibitors (Frye and Johnson, 2009). In contrast, although BRAF inhibitors also block BRAF kinase activity, this relieves auto-inhibition and results in BRAF hyperphosphorylation, BRAF binding to CRAF, pathway activation and oncogenesis, all presumably because BRAF can heterodimerize with CRAF. Our study also highlights the critical difference between BRAF-selective and pan-RAF drugs. Whereas BRAF-selective drugs cause pathway activation in a RAS-dependent manner, this does not occur with pan-RAF drugs. Our results provide important insight into the genetics of human cancer. Excluding V600 mutants, D594 mutants are the third most common in BRAF in cancer (34 out of 443 cases or ∼7.7%; www.sanger.ac.uk/genetics/CGP/cosmic/). Furthermore, as mentioned in the Introduction, while BRAF and RAS mutations are generally mutually exclusive, 4 of the 34 (11.8%) tumors with D594 mutations also have mutations in RAS. This is a highly significant enrichment for the coincidence of these mutations (p < 10; Fisher's Exact Test) and suggests a functional interaction. We now provide strong circumstantial evidence of such an interaction using transgenic mice. By themselves, Braf and Kras do not induce melanoma, but they cooperate to induce rapid onset melanoma. This highly significant result (p < 0.0002) provides a rational explanation for the coincidence of these mutations in human cancer. Furthermore, we show that the BRAF inhibitors also hyperactivate this pathway in growth factor stimulated cells, providing an explanation of why kinase dead BRAF mutations are not always coincident with RAS mutations; presumably in some tumors the cooperating mutation is upstream of RAS. Our results also suggest several potential mechanisms by which resistance to RAF targeting drugs could develop in patients. BRAF mutant tumors could become resistant to BRAF-selective drugs, if they acquire a mutation in RAS or an upstream component that activates RAS, or if the drugs select a population of cells harboring pre-existing mutations in RAS. Theoretically this would cause BRAF-mediated CRAF activation, which may not only induce resistance, but could potentially promote tumor growth. In line with this, increased expression of CRAF can mediate acquired resistance to pan-RAF drugs in BRAF mutant cancer cells in vitro (Montagut et al., 2008), establishing that CRAF can mediate resistance under some circumstances. Our in vitro studies also suggest that a potential mechanism of resistance in patients with RAS mutant tumors being treated with pan-RAF drugs is acquisition (or selection for cells with pre-existing mutations) of a CRAF mutation such as a gatekeeper mutant that prevents drug binding. Again this would potentially result in BRAF-mediated activation of CRAF (Figure 7D) and possibly accelerated tumor growth. Although our studies are restricted to cell lines and transgenic mice, they do have important immediate clinical implications. They strongly argue that BRAF-selective inhibitors should not be administered to patients with RAS mutant tumors, because long-term use could accelerate tumor growth. Intriguingly, 10%–15% of patients treated with BRAF-selective drugs develop squamous cell carcinoma (SCC)(Flaherty et al., 2009; Schwartz et al., 2009). Although MEK–ERK signaling has not yet been implicated in this response, 22% of SCCs harbour oncogenic mutations in RAS (9% HRAS, 8% NRAS, 5% KRAS: www.sanger.ac.uk/genetics/CGP/cosmic/), raising the intriguing possibility that the BRAF-selective drugs act as tumor promoters in premalignant skin cells harboring existing mutations in RAS and/or activation of upstream components that activate RAS. While sorafenib is equipotent for wild-type and BRAF (Wilhelm et al., 2004), the BRAF inhibitors we used are approximately 10-fold more active against BRAF (King et al., 2006; Tsai et al., 2008). Nevertheless, our data establish that they target wild-type BRAF in RAS mutant cells. The problem of mutant v.s. wild-type protein specificity is likely to be difficult to resolve, because whereas full inhibition of BRAF may be necessary for clinical response in BRAF mutant tumors, activation of only a small proportion of wild-type BRAF could be sufficient to activate the pathway in RAS mutant cells. Thus, to achieve efficacy against BRAF but avoid activation of wild-type BRAF in RAS mutant cells, the drugs will need to be exquisitely selective for the mutant protein. Alternatively, pan-RAF drugs may be effective because they will target both BRAF and CRAF activated by BRAF in RAS mutant tumors. Furthermore, our data suggest that CRAF or MEK selective drugs should be used in RAS mutant tumors, because they do not induce BRAF-CRAF complexes and will not activate the pathway if the tumors acquire mutations such as CRAF that block drug binding. Perhaps RAF and MEK inhibitors should be combined to provide the best responses and prevent emergence of resistance, but these issues need to be balanced against the urgency of the clinical problem being addressed. In summary, we show that inhibition of BRAF in RAS mutant cancer cells leads to MEK hyperactivation through CRAF. We have elucidated another mechanism by which BRAF activates MEK–ERK signaling, not only to drive tumorigenesis and tumor progression, but also potentially to allow development of de novo or acquired resistance to RAF-targeted therapies. Clearly, BRAF is a remarkably versatile oncogene that can promote MEK–ERK activation and tumor progression through several mechanisms and these will require different therapeutic strategies for effective disease management. Notably, many of the mutations that occur in other kinases in cancer are also predicted to cause inactivation (www.sanger.ac.uk/genetics/CGP/cosmic/). Our data raise the possibility that these could also act as idiosyncratic gain-of-function mutations that drive tumorigenesis. This study also raises important clinical questions and highlights the importance of fully understanding how signaling networks function to fully comprehend how patients may respond to targeted drugs. They also highlight the importance of genetic screening for patients, not only to identify those who are likely to respond, but to exclude those who could experience adverse effects and thereby ensure successful implementation of personalized medicine. Expression vectors for epitope-tagged BRAF and CRAF have been described (Wan et al., 2004). For western blotting the following antibodies were used: rabbit anti-ppMEK1/2 and mouse anti-myc 9B11 (Cell Signaling Technology); mouse anti-NRAS (C-20), rabbit anti-ERK2 (C-14), rabbit anti-ARAF (C-20), mouse anti-BRAF (F-7) (Santa Cruz Biotechnology); mouse anti-Tubulin, and mouse anti-ppERK1/2 (Sigma); mouse anti-CRAF (for western blotting) (BD Transduction Laboratories). For immunoprecipitation, the following antibodies were used: rabbit anti-myc (Abcam); rabbit anti-CRAF (C-20;Santa Cruz Biotechnology); mouse anti-BRAF (F-7) (Ab from Santa Cruz Biotechnology). Calf intestinal phosphatase (CIP) was from New England Biolabs (NEB). PD184352, sorafenib and PLX4720 were synthesized in-house; 885-A was synthesized by Evotec AG (Abingdon, UK). All drugs were prepared in DMSO. Synthetic routes are available on request. Human cell lines were cultured in DMEM (A375, WM852, HCT116, SW620, and PMWK) or RPMI (D04, MM485, MM415, and WM1791c) supplemented with 10% fetal bovine serum. For protein depletion, 3 × 10 D04 cells were transfected with 5nM CRAF (5′-AAGCACGCTTAGATTGGAATA-3′) or NRAS (5′-CATGGCACTGTACTCTTCTCG-3′) specific, or scrambled control (5′-AAACCGTCGATTTCACCCGGG-3′) siRNA using INTERFERin as recommended by the manufacturer (Polyplus Transfection SA). For transient expression studies, D04 cells were transfected using the Amaxa Nucleofector System as recommended by the manufacturer (Lonza). COS-7 cells were propagated, transfected, and extracted as described (Wan et al., 2004). For generation of stable lines, D04 cells were transfected with pMCEF-FLAG-CRAF or pMCEF-FLAG-CRAF using Effectene as recommended by the manufacturer (Invitrogen) and selected in G418 (1 mg/ml). Cell lysates were prepared with NP40 buffer as described (Wan et al., 2004). For immunoprecipitation, lysates were incubated with 2 μg BRAF F-7, 5 μg CRAF C-20 or 2 μg rabbit anti-myc antibodies, captured on Protein G sepharose 4B beads (Sigma) and analyzed by western blotting using standard protocols. Specific bands were detected using fluorescent-labeled secondary antibodies (Invitrogen; Li-COR Biosciences) and analyzed using an Odyssey Infrared Scanner (Li-COR Biosciences). For CIP treatment, immunoprecipitates were washed twice with NP40 lysis buffer, once in CIP buffer (50 mM Tris-Cl [pH 7.5], 150 mM NaCl, 10 mM MgCl2, and 1 mM EDTA), and incubated with CIP with or without 0.2 mM Na3VO4 and 7 mM EDTA. The immunoprecipitates were washed in CIP buffer and western blotted. Coupled RAF kinase assays were performed with immunoprecipitated CRAF or BRAF as described (Wan et al., 2004). Membrane fractionation was as described (Garnett et al., 2005). Experiments were performed under Home Office license authority in accordance with United Kingdom Coordinating Committee on Cancer Research Guidelines (Workman et al., 1988) and with local Ethics Committee approval. To activate CreERT2, mice were treated with four doses (10mg each) of topically applied tamoxifen as described (Dhomen et al., 2009). Genotyping was performed by PCR. Braf and Braf was analyzed as described for Braf and Braf respectively and Tyr::CreERT2 was analyzed as described (Dhomen et al., 2009). Kras was analyzed using primers 5′-CGCAGACTGTAGAGCAGCG-3′ and 5′-CCATGGCTTGAGTAAGTCTGC-3′. For expression analysis, RNA was prepared (QIAGEN RNEasy, QIAGEN) and first-strand cDNA synthesis was performed with 500ng total RNA and random hexanucleotides (Random Primers, Invitrogen). Specific genes were amplified under linear conditions for analysis as described (Dhomen et al., 2009). For Kras cDNA sequencing, a 238 bp fragment of Kras cDNA was PCR amplified using primers 5′-GGCGGCAGCGCTGTGGCGGCG-3′ and 5′-CGTAGGGTCATACTCATCCAC-3′ and sequenced using automated dideoxy sequencing. For immunohistochemistry (IHC), tissues were fixed and analyzed as described (Dhomen et al., 2009). Positive (a well characterized sample of mouse melanoma) and negative (omission of the primary antibody and substitution with preimmune serum) controls were included in each slide run. Immunohistochemical staining was analyzed by two of the authors on a multi-headed microscope. Tumor cell lines were established by mechanically dissociating tumors in DMEM/20%FCS/Primocin (0.1mg/ml - InvivoGen) and clonal lines were selected by limiting dilution. Extended Experimental ProceduresReagentsFor western blotting the following antibodies were used: rabbit anti-ppMEK1/2 and mouse anti-myc 9B11 (Cell Signaling Technology); mouse anti-NRAS (C-20), rabbit anti-ERK2 (C-14), rabbit anti-ARAF (C-20), mouse anti-BRAF (F-7) (Santa Cruz Biotechnology); mouse anti-Tubulin, and mouse anti-ppERK1/2 (Sigma); mouse anti-CRAF (for western blotting) (BD Transduction Laboratories). For immunoprecipitation, the following antibodies were used: rabbit anti-myc (Abcam); rabbit anti-CRAF (C-20; Santa Cruz Biotechnology); mouse anti-BRAF (F-7) (Santa Cruz Biotechnology). Calf intestinal phosphatase (CIP) was from New England Biolabs (NEB). PD184352, sorafenib and PLX4720 were synthesized in-house; 885-A was synthesized by Evotec AG (Abingdon, UK). All drugs were prepared in DMSO. Synthetic routes are available on request.Expression ConstructsThe expression vectors for wild-type human CRAF and wild-type human BRAF, pEFm/CRAF and pEFm/BRAF respectively have been described (Marais et al., 1995). Briefly, the vector backbone is pUC19 and the elongation factor 1α (EF1α) promoter is used to drive exogenous protein expression. The vector includes the first intron from human EF1α to assist mRNA processing during expression. The β-globin 5′ and 3′ untranslated regions (UTRs) are used to provide a strong translation start site (5′ UTR), and also to provide mRNA stability and a poly adenylation signal (3′ UTR). The vector introduces an amino-terminal myc-epitope tag (EQKLISEEDL) onto the exogenously expressed protein. The BRAF coding region includes the alternatively spliced exons 1 and 2 but not exons 8b or 10a and various modifications were introduced to provide additional restriction sites (without changing the amino acid sequence) and alterations to the 3′-UTR to allow easier manipulation of this construct. Standard PCR-directed mutagenesis approaches were used to generate the various mutations used in the study and all mutations were verified by automated dideoxy sequencing. The expression vector pMCEF/FLAG/CRAF uses the same expression cassette, but the backbone also possesses a neo resistance cassette to facilitate selection in the presence of G418. In addition, a version of this vector was used that incorporates a FLAG (DYKDDDKGS), rather than a myc-epitope tag.Cell CultureHuman cell lines were cultured in DMEM (A375, WM852, HCT116, SW620, PMWK, SKMel24, SKMel28, D25) or RPMI (D04, MM485, MM415, WM1791c) supplemented with 10% fetal bovine serum. African green monkey kidney COS-7 cells were cultured in DMEM supplemented with 10% fetal bovine serum. All cell lines were incubated at 37°C and 10% CO2. For inhibitor treatment, the drugs were dissolved in DMSO and added to the medium for 2-5 hr. When two compounds were used, the first was added 30 min prior to the second. For cell growth assays, cells were seeded in 96-well plates and treated with drug in quadruplicate in a 10-point titration assay for 5 days. The amount of growth (% DMSO controls) was determined using sulphorhodamine B reagent (Monks et al., 1991) as follows. 1,000-10,000 cells (depending on cell type) were plated into 96-well plates in 100 μL medium. After 24 hr, compounds prepared in DMSO (10mM stocks) were serially diluted in culture medium at 2× the final required concentrations and 100μL were added to the cells to nine final concentrations of 0.005 μM-100 μM. After a further 5 days, the cells are fixed in trichloroacteic acid (10%), and stained with sulforhodamine-B (0.1%). After rinsing, the bound stain was disolved using 100 μL 10 mM Tris (pH 8.0) and the absorbance at 540 nm determined. The data were analyzed by nonlinear regression to a four-parameter logistic equation (Graphpad Prism, Graphpad Software Inc., San Diego, CA, USA) and the GI50 value determined.siRNA Transfections3 × 10 D04 cells per 35 mm diameter well were seeded in 2ml growth medium the day before transfection. The cells were either mock-transfected or transfected with 6nM CRAF-specific (5′-AAGCACGCTTAGATTGGAATA-3′) or NRAS-specific (5′-CATGGCACTGTACTCTTCTCG-3′) siRNA using INTERFERin as recommended by the manufacturer (Polyplus Transfection SA). Briefly, 0.6 μl of 20 μM siRNA and 6 μl of INTERFERin were combined in a total of 200 μl serum free medium in RNase-free tubes. The mix was vortexed for 10 s and incubated for 5-10 min before adding the complexes dropwise to the cells. The cells were serum-starved the day after transfecting and extracts were prepared 48 hr after transfection.DNA TransfectionsFor transient protein expression in D04 cells, Lonza Nucleofector Technology (Lonza, Cologne AG) was used. 2 μg of DNA was mixed with 1x10 cells resuspended in 100 μl of Nucleofection Solution V in an Amaxa-certified cuvette and transfected using program T030. The cells were re-plated into 35mm diameter tissue culture wells and incubated for 48 hr before preparation of cell extracts.For generation of stable lines, D04 cells were transfected using Effectene (Invitrogen) and selected in G418. 3-4x10 cells were plated in 35 mm diameter wells and incubated overnight. 0.4 μg of DNA diluted into 100 μl of DNA condensation buffer (EC) and 3.2 μl enhancer were mixed vigorously and incubated for 2-5 min. 10 μl Effectene reagent was added and the mixture was incubated for another 5-10 min. The cells were washed with 2ml PBS and 1.6ml fresh serum containing medium was added. The DNA complexes were diluted with 600 μl of culture medium and the mixture added to the cells drop-wise. After six hours, the medium was replaced with 2ml of fresh growth medium. After 48 hr, the cells were replated into several 10cm dishes in a 10-fold dilution series and incubated in G418 (1mg/ml) for selection. The medium was refreshed weekly and after 2-3 weeks, single colonies were selected and expanded.For transient expression in COS-7 cells, Lipofectamine (Invitrogen) was used. 2x10 cells were plated into 35mm diameter wells and incubated overnight. 75 to 200 ng of expression plasmid (depending on construct) was mixed with empty vector to a total of 700 ng DNA in 16μl PBS. Typically, 3 μl of Lipofectamine in 13 μl of PBS was added to the DNA on the surface of a bacterial plate and incubated (Lipofectamine is inactivated by binding to polypropylene) for 15 min at room temperature. The cells were washed twice with 1ml serum-free DMEM, and then overlaid with 800 μl of serum free DMEM. 200 μl of serum free DMEM was added to the DNA:Lipofectamine mix, and the total volume was added to the cells. After six hours, the complexes were removed and replaced with 2ml of normal culture medium. Cell extracts were prepared two days following transfection.Preparation of Cell LysatesCulture medium was aspirated from cells and cells were placed on ice and washed three times in ice-cold PBS. Depending on the assays, the cells were scraped into 50-200 μl Nonidet P40 (NP40) extraction buffer (Table S3) and incubated on ice for five minutes. The cells were sheared by passing through a pipette tip several times and the samples were centrifuged at 20,000 × g for 5 min at 4°C and the soluble fraction was harvested.RAF CoimmunoprecipitationsImmunoprecipitations were performed in 300 μl cell lysates from one 35mm diameter well for endogenous protein or from 2-3 wells for transfected protein. Endogenous BRAF or CRAF were immunoprecipitated with 2μg BRAF F-7 or 5μg CRAF C-20 respectively and myc-tagged BRAF and CRAF with 2 μg rabbit anti-myc antibody. The antibody-protein complex was captured using 20 μl of a 1:1 Protein G sepharose 4B beads (Sigma-Aldrich) mixture in NP40 lysis buffer (Table S3) and immunoprecipitates (IPs) were mixed for 2 hr at 4°C on a rotation wheel. Thereafter, the IPs were washed three times with 300 μl of NP40 lysis buffer (Table S3) before analysis on standard sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE). Specific bands were detected using fluorescent-labeled secondary antibodies (Invitrogen; Li-COR Biosciences) and analyzed using an Odyssey Infrared Scanner (Li-COR Biosciences). For CIP treatment, immunoprecipitates were washed twice with NP40 lysis buffer (Table S3) and once in CIP buffer (Table S3). Thereafter immunoprecipitates were incubated with 30 μl CIP buffer containing 5 units of CIP in the presence or absence of 0.2 mM Na3VO4 phosphatase inhibitor and 7mM EDTA. Controls were incubated in CIP buffer without CIP. The reactions were performed at 30°C for 30 min before analysis on SDS-PAGE.Cell Fractionation ExperimentsD04 cells were plated on two 10cm dishes per treatment and grown to confluency. After treatment, cells were washed three times with cold PBS, washed once with 20mM HEPES pH 7.4 and then lysed by scraping in 20mM HEPES pH 7.4 supplemented with protease inhibitors (1ml per 2 plates). Cells were disrupted by passing them through a 9G syringe ten times, followed by another ten times through a 19G syringe (Terumo Medical). Lysates were centrifuged at 900 × g for five minutes to pellet the nuclear proteins. The supernatant was transferred to fresh 1.5ml ultracentrifuge tubes (Beckman Coulter) and 200 μl removed as a total lysate control. The remainder was centrifuged at 100,000 × g for 30 min at 4°C to separate the cytosolic fraction from the membrane fraction. The supernatant containing the cytosolic fraction was transferred to a fresh 1.5ml tube and the pelleted membrane fraction washed once in 20 mM HEPES pH 7.4 before resuspension in 200 μl 20mM HEPES pH 7.4/1% Triton X-100. For analysis on SDS-PAGE, the concentration of protein was determined by Bradford protein assay (Bio-Rad Laboratories) using purified bovine serum albumin (BSA) as a standard as described by the manufacturer. Equal amounts of protein were loaded for the cytosolic fraction and total cell lysate. Three times as much protein was loaded for the membrane fraction.RAF Kinase AssaysThe in vitro kinase activity of endogenous RAF proteins or myc-tagged RAF proteins transiently expressed in COS-7 cells was measured using a coupled kinase cascade assay with GST-MEK, GST-ERK and myelin basic protein (MBP) (Sigma-Aldrich) as sequential substrates. ERK activation was quantified by measuring the incorporation of [P]-orthophosphate (PerkinElmer) into MBP. For measurement of endogenous BRAF kinase activity, D04 or A375 cells were seeded in 10cm dishes and harvested in 300 μl of NP40 buffer (Table S3) as described above. Protein concentrations were determined and equal amounts of protein were immunoprecipitated as described above.For measurement of mutant BRAF kinase activities, COS-7 cells were transiently transfected with myc-tagged BRAF and cells in one 35 mm diameter well were harvested in 200μl of NP40 buffer (Table S3). The relative concentrations of exogenously expressed RAF in these cell lysates were determined by quantitative western blotting using the myc antibody (Cell Signaling Technology) specified above. Bands were quantified using the Odyssey infrared imaging system (LI-COR Biosciences). Equivalent amounts of RAF were immunoprecipitated using rabbit myc antibody (Abcam) as specified above.Endogenous and transiently transfected RAF proteins were immunoprecipitated in a total of 300 μl NP40 buffer (Table S3) for 2 hr at 4°C and immunoprecipitates were washed sequentially three times with chilled wash buffer (Table S3) containing decreasing concentrations of KCl (1M KCl, 0.1M KCl and no KCl). The first-step reaction was initiated by addition of 20 μl MKK buffer (containing GST-MEK and GST-ERK, Table S3) to the beads and incubated at 30°C for 10 min in the case of myc-tagged BRAF or 30 min for endogenous BRAF and CRAF. Reactions were terminated by the addition of 20 μl KILL buffer (Table S3), which contains EDTA to chelate Mg ions and inhibit kinase activity. The reaction supernatants were collected from the beads and transferred into fresh tubes for the second step reaction. 5 μl of supernatant was incubated with 25 μl MBP buffer containing [γ-P]ATP (PerkinElmer) for ten minutes at 30°C in triplicate to measure ERK activity. The second reaction was terminated by spotting 20 μl of reaction mix onto a 1cm piece of P81 paper (VWR International), which was then dropped into 400ml 25mM orthophosphate solution. The papers were washed three times in 400 ml 25 mM orthophosphate solution to remove the unincorporated ATP and the [P]-orthophosphate incorporated into MBP was determined using Cerenkov counting. For transfected samples, the background counts were determined using lysates of cells transfected with the empty vector. For endogenous protein, samples in which no RAF was immunoprecipitated were used. Background values were removed and to ensure linearity, assays were used at below 50% saturation. To determine BRAF and BRAF sensitivity to 885-A, immunoprecipitated BRAF was preincubated with drug in KCl-free buffer for 10 min at room temperature prior to the first-step reaction.To measure the activity of purified BRAF, a 96-well DELFIA-based assay system was used. Full-length rabbit MEK1 protein was expressed with a GST tag at the N-terminus and a C-terminal histidine tag in Escherichia coli JM109 bacteria and purified by nickel-agarose affinity chromatography. Full length BRAF protein was generated by infection of SF9 insect cells with a recombinant baculovirus expressing full-length human BRAF with an N-terminal histidine tag and purified as above. For the kinase assays, all incubations were at room temperature with shaking. 4 μg GST-MEK1, 100-200ng purified BRAF and 1 μl inhibitor at the required concentrations (0.001 to 100 μM final concentration) were added to the wells of glutathione-coated plates and preincubated for 10 min. ATP in DELFIA assay buffer (20 μL; Table S3), to give a final concentration of 100 μM, was added to each well, and the plates were incubated for 45 min. The plates were washed 3X with 200 μl 0.1% tween20/water. Primary antibody (rabbit anti-phospho MEK1/2 diluted 1/2000, Cell Signaling Technologies) and Eu-labeled anti-rabbit secondary antibody (diluted 1/1000, Perkin-Elmer) were preincubated for 30 min and 100 μl was added to the plates and incubated for a further hour. The plates were washed as before, and 100 μl DELFIA enhancement solution (Perkin-Elmer, Turku, Finland) was added. The plates were sealed and incubated for 30 min and europium counts measured on Spectramax M5 plate reader (Molecular Devices, Wokingham, UK). IC50 values were determined using GraphPad Prism (Graphpad Software Inc., San Diego, CA, USA).Transgenic MiceExperiments were performed under Home Office license authority with local Ethics Committee approval. To activate CreERT2, four doses of tamoxifen (Sigma; 10mg each in 100% ethanol) were applied topically to the shaven skin on the backs of the mice every other day for 7 days. Genotyping was performed by PCR using DNA prepared with DNeasy kits (QIAGEN). Braf and Braf were analyzed using primers: A) 5′-GCCCAGGCTCTTTATGAGAA-3; B) 5′-AGTCAATCATCCACAGAGACCT-3′; and C) 5′-GCTTGGCTGGACGTAAACTC-3′. A+B detects the wild-type BRAF allele (466 bp product) and Braf, the Cre-recombinase recombined allele (518pb product). A+C detects the targeted allele Braf (140 bp). Tyr::CreERT2 was analyzed using primers 5′-GAAGCAACTCATCGATTG-3′ and 5′-TGAAGGGTCTGGTAGGATCA-3′. Kras was analyzed using primers 5′-CGCAGACTGTAGAGCAGCG-3′ and 5′-CCATGGCTTGAGTAAGTCTGC-3′.mRNA expression analysis was performed by RT-PCR. RNA was prepared using the RNEasy kit (QIAGEN). First-strand cDNA synthesis was performed with 500ng total RNA and random hexanucleotides (Random Primers, Invitrogen). Tyrosinase (Tyr), was detected using primers 5′-TGGTTCCTTTCATACCGCTC-3′ and 5′-CAGATACGACTGGCTTGTTCC-3′; Dct with 5′-GCAAGATTGCCTGTCTCTCC-3′ and 5′-AGTCCAGTGTTCCGTCTGCT-3′; Pax3 with 5′-CCAGGATGATGCGGCCCGGCCCGGG-3′ and 5′-AGGATGCGGCTGATAGAACTCACTG-3′; and silver/gp100 (Si) with 5′-GGAGAGGTGGCCAGGTATC-3′ and 5′-CAGTAATGGTGAAGGTTGAAC-3′. The control Gapdh was detected with 5′- GATGGCCCCTCGGAAAGCT-3′ 5′-CCAGTGAGCTTCCCGTTCAGC-3′. To sequence Kras cDNA, a 238 bp product from Kras cDNA was PCR amplified using primers 5′-GGCGGCAGCGCTGTGGCGGCG-3′ and 5′-CGTAGGGTCATACTCATCCAC-3′ and directly sequenced using these primers and automated dideoxy sequencing.For immunohistochemistry (IHC), tumors were fixed in 10% buffered formalin and embedded in paraffin. Sections (3-10μm) were stained with hematoxylin and eosin using standard protocols. For S100 and Ki67 staining, antigen retrieval was performed in citrate buffer (pH 6.0, 30 min) and revealed using a rabbit polyclonal antibody (Dako, 1/1000), the Rabbit Envision Peroxidase kit and the AEC substrate chromogen (Dako) for S100, and a rat monoclonal antibody (Dako,1/25), the rat Vectastain ABC kit (Vector Labs, USA) and DAB as chromagen for Ki67. Positive (a well characterized sample of mouse melanoma) and negative (omission of the primary antibody and substitution with preimmune serum) controls were included in each slide run. Immunohistochemical staining was analyzed by two of the authors on a multi-headed microscope. Tumor cell lines were established by mechanically dissociating tumors in DMEM/20%FCS/Primocin (0.1mg/ml - InvivoGen) and clonal lines were selected by limiting dilution. For western blotting the following antibodies were used: rabbit anti-ppMEK1/2 and mouse anti-myc 9B11 (Cell Signaling Technology); mouse anti-NRAS (C-20), rabbit anti-ERK2 (C-14), rabbit anti-ARAF (C-20), mouse anti-BRAF (F-7) (Santa Cruz Biotechnology); mouse anti-Tubulin, and mouse anti-ppERK1/2 (Sigma); mouse anti-CRAF (for western blotting) (BD Transduction Laboratories). For immunoprecipitation, the following antibodies were used: rabbit anti-myc (Abcam); rabbit anti-CRAF (C-20; Santa Cruz Biotechnology); mouse anti-BRAF (F-7) (Santa Cruz Biotechnology). Calf intestinal phosphatase (CIP) was from New England Biolabs (NEB). PD184352, sorafenib and PLX4720 were synthesized in-house; 885-A was synthesized by Evotec AG (Abingdon, UK). All drugs were prepared in DMSO. Synthetic routes are available on request. The expression vectors for wild-type human CRAF and wild-type human BRAF, pEFm/CRAF and pEFm/BRAF respectively have been described (Marais et al., 1995). Briefly, the vector backbone is pUC19 and the elongation factor 1α (EF1α) promoter is used to drive exogenous protein expression. The vector includes the first intron from human EF1α to assist mRNA processing during expression. The β-globin 5′ and 3′ untranslated regions (UTRs) are used to provide a strong translation start site (5′ UTR), and also to provide mRNA stability and a poly adenylation signal (3′ UTR). The vector introduces an amino-terminal myc-epitope tag (EQKLISEEDL) onto the exogenously expressed protein. The BRAF coding region includes the alternatively spliced exons 1 and 2 but not exons 8b or 10a and various modifications were introduced to provide additional restriction sites (without changing the amino acid sequence) and alterations to the 3′-UTR to allow easier manipulation of this construct. Standard PCR-directed mutagenesis approaches were used to generate the various mutations used in the study and all mutations were verified by automated dideoxy sequencing. The expression vector pMCEF/FLAG/CRAF uses the same expression cassette, but the backbone also possesses a neo resistance cassette to facilitate selection in the presence of G418. In addition, a version of this vector was used that incorporates a FLAG (DYKDDDKGS), rather than a myc-epitope tag. Human cell lines were cultured in DMEM (A375, WM852, HCT116, SW620, PMWK, SKMel24, SKMel28, D25) or RPMI (D04, MM485, MM415, WM1791c) supplemented with 10% fetal bovine serum. African green monkey kidney COS-7 cells were cultured in DMEM supplemented with 10% fetal bovine serum. All cell lines were incubated at 37°C and 10% CO2. For inhibitor treatment, the drugs were dissolved in DMSO and added to the medium for 2-5 hr. When two compounds were used, the first was added 30 min prior to the second. For cell growth assays, cells were seeded in 96-well plates and treated with drug in quadruplicate in a 10-point titration assay for 5 days. The amount of growth (% DMSO controls) was determined using sulphorhodamine B reagent (Monks et al., 1991) as follows. 1,000-10,000 cells (depending on cell type) were plated into 96-well plates in 100 μL medium. After 24 hr, compounds prepared in DMSO (10mM stocks) were serially diluted in culture medium at 2× the final required concentrations and 100μL were added to the cells to nine final concentrations of 0.005 μM-100 μM. After a further 5 days, the cells are fixed in trichloroacteic acid (10%), and stained with sulforhodamine-B (0.1%). After rinsing, the bound stain was disolved using 100 μL 10 mM Tris (pH 8.0) and the absorbance at 540 nm determined. The data were analyzed by nonlinear regression to a four-parameter logistic equation (Graphpad Prism, Graphpad Software Inc., San Diego, CA, USA) and the GI50 value determined. 3 × 10 D04 cells per 35 mm diameter well were seeded in 2ml growth medium the day before transfection. The cells were either mock-transfected or transfected with 6nM CRAF-specific (5′-AAGCACGCTTAGATTGGAATA-3′) or NRAS-specific (5′-CATGGCACTGTACTCTTCTCG-3′) siRNA using INTERFERin as recommended by the manufacturer (Polyplus Transfection SA). Briefly, 0.6 μl of 20 μM siRNA and 6 μl of INTERFERin were combined in a total of 200 μl serum free medium in RNase-free tubes. The mix was vortexed for 10 s and incubated for 5-10 min before adding the complexes dropwise to the cells. The cells were serum-starved the day after transfecting and extracts were prepared 48 hr after transfection. For transient protein expression in D04 cells, Lonza Nucleofector Technology (Lonza, Cologne AG) was used. 2 μg of DNA was mixed with 1x10 cells resuspended in 100 μl of Nucleofection Solution V in an Amaxa-certified cuvette and transfected using program T030. The cells were re-plated into 35mm diameter tissue culture wells and incubated for 48 hr before preparation of cell extracts. For generation of stable lines, D04 cells were transfected using Effectene (Invitrogen) and selected in G418. 3-4x10 cells were plated in 35 mm diameter wells and incubated overnight. 0.4 μg of DNA diluted into 100 μl of DNA condensation buffer (EC) and 3.2 μl enhancer were mixed vigorously and incubated for 2-5 min. 10 μl Effectene reagent was added and the mixture was incubated for another 5-10 min. The cells were washed with 2ml PBS and 1.6ml fresh serum containing medium was added. The DNA complexes were diluted with 600 μl of culture medium and the mixture added to the cells drop-wise. After six hours, the medium was replaced with 2ml of fresh growth medium. After 48 hr, the cells were replated into several 10cm dishes in a 10-fold dilution series and incubated in G418 (1mg/ml) for selection. The medium was refreshed weekly and after 2-3 weeks, single colonies were selected and expanded. For transient expression in COS-7 cells, Lipofectamine (Invitrogen) was used. 2x10 cells were plated into 35mm diameter wells and incubated overnight. 75 to 200 ng of expression plasmid (depending on construct) was mixed with empty vector to a total of 700 ng DNA in 16μl PBS. Typically, 3 μl of Lipofectamine in 13 μl of PBS was added to the DNA on the surface of a bacterial plate and incubated (Lipofectamine is inactivated by binding to polypropylene) for 15 min at room temperature. The cells were washed twice with 1ml serum-free DMEM, and then overlaid with 800 μl of serum free DMEM. 200 μl of serum free DMEM was added to the DNA:Lipofectamine mix, and the total volume was added to the cells. After six hours, the complexes were removed and replaced with 2ml of normal culture medium. Cell extracts were prepared two days following transfection. Culture medium was aspirated from cells and cells were placed on ice and washed three times in ice-cold PBS. Depending on the assays, the cells were scraped into 50-200 μl Nonidet P40 (NP40) extraction buffer (Table S3) and incubated on ice for five minutes. The cells were sheared by passing through a pipette tip several times and the samples were centrifuged at 20,000 × g for 5 min at 4°C and the soluble fraction was harvested. Immunoprecipitations were performed in 300 μl cell lysates from one 35mm diameter well for endogenous protein or from 2-3 wells for transfected protein. Endogenous BRAF or CRAF were immunoprecipitated with 2μg BRAF F-7 or 5μg CRAF C-20 respectively and myc-tagged BRAF and CRAF with 2 μg rabbit anti-myc antibody. The antibody-protein complex was captured using 20 μl of a 1:1 Protein G sepharose 4B beads (Sigma-Aldrich) mixture in NP40 lysis buffer (Table S3) and immunoprecipitates (IPs) were mixed for 2 hr at 4°C on a rotation wheel. Thereafter, the IPs were washed three times with 300 μl of NP40 lysis buffer (Table S3) before analysis on standard sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE). Specific bands were detected using fluorescent-labeled secondary antibodies (Invitrogen; Li-COR Biosciences) and analyzed using an Odyssey Infrared Scanner (Li-COR Biosciences). For CIP treatment, immunoprecipitates were washed twice with NP40 lysis buffer (Table S3) and once in CIP buffer (Table S3). Thereafter immunoprecipitates were incubated with 30 μl CIP buffer containing 5 units of CIP in the presence or absence of 0.2 mM Na3VO4 phosphatase inhibitor and 7mM EDTA. Controls were incubated in CIP buffer without CIP. The reactions were performed at 30°C for 30 min before analysis on SDS-PAGE. D04 cells were plated on two 10cm dishes per treatment and grown to confluency. After treatment, cells were washed three times with cold PBS, washed once with 20mM HEPES pH 7.4 and then lysed by scraping in 20mM HEPES pH 7.4 supplemented with protease inhibitors (1ml per 2 plates). Cells were disrupted by passing them through a 9G syringe ten times, followed by another ten times through a 19G syringe (Terumo Medical). Lysates were centrifuged at 900 × g for five minutes to pellet the nuclear proteins. The supernatant was transferred to fresh 1.5ml ultracentrifuge tubes (Beckman Coulter) and 200 μl removed as a total lysate control. The remainder was centrifuged at 100,000 × g for 30 min at 4°C to separate the cytosolic fraction from the membrane fraction. The supernatant containing the cytosolic fraction was transferred to a fresh 1.5ml tube and the pelleted membrane fraction washed once in 20 mM HEPES pH 7.4 before resuspension in 200 μl 20mM HEPES pH 7.4/1% Triton X-100. For analysis on SDS-PAGE, the concentration of protein was determined by Bradford protein assay (Bio-Rad Laboratories) using purified bovine serum albumin (BSA) as a standard as described by the manufacturer. Equal amounts of protein were loaded for the cytosolic fraction and total cell lysate. Three times as much protein was loaded for the membrane fraction. The in vitro kinase activity of endogenous RAF proteins or myc-tagged RAF proteins transiently expressed in COS-7 cells was measured using a coupled kinase cascade assay with GST-MEK, GST-ERK and myelin basic protein (MBP) (Sigma-Aldrich) as sequential substrates. ERK activation was quantified by measuring the incorporation of [P]-orthophosphate (PerkinElmer) into MBP. For measurement of endogenous BRAF kinase activity, D04 or A375 cells were seeded in 10cm dishes and harvested in 300 μl of NP40 buffer (Table S3) as described above. Protein concentrations were determined and equal amounts of protein were immunoprecipitated as described above. For measurement of mutant BRAF kinase activities, COS-7 cells were transiently transfected with myc-tagged BRAF and cells in one 35 mm diameter well were harvested in 200μl of NP40 buffer (Table S3). The relative concentrations of exogenously expressed RAF in these cell lysates were determined by quantitative western blotting using the myc antibody (Cell Signaling Technology) specified above. Bands were quantified using the Odyssey infrared imaging system (LI-COR Biosciences). Equivalent amounts of RAF were immunoprecipitated using rabbit myc antibody (Abcam) as specified above. Endogenous and transiently transfected RAF proteins were immunoprecipitated in a total of 300 μl NP40 buffer (Table S3) for 2 hr at 4°C and immunoprecipitates were washed sequentially three times with chilled wash buffer (Table S3) containing decreasing concentrations of KCl (1M KCl, 0.1M KCl and no KCl). The first-step reaction was initiated by addition of 20 μl MKK buffer (containing GST-MEK and GST-ERK, Table S3) to the beads and incubated at 30°C for 10 min in the case of myc-tagged BRAF or 30 min for endogenous BRAF and CRAF. Reactions were terminated by the addition of 20 μl KILL buffer (Table S3), which contains EDTA to chelate Mg ions and inhibit kinase activity. The reaction supernatants were collected from the beads and transferred into fresh tubes for the second step reaction. 5 μl of supernatant was incubated with 25 μl MBP buffer containing [γ-P]ATP (PerkinElmer) for ten minutes at 30°C in triplicate to measure ERK activity. The second reaction was terminated by spotting 20 μl of reaction mix onto a 1cm piece of P81 paper (VWR International), which was then dropped into 400ml 25mM orthophosphate solution. The papers were washed three times in 400 ml 25 mM orthophosphate solution to remove the unincorporated ATP and the [P]-orthophosphate incorporated into MBP was determined using Cerenkov counting. For transfected samples, the background counts were determined using lysates of cells transfected with the empty vector. For endogenous protein, samples in which no RAF was immunoprecipitated were used. Background values were removed and to ensure linearity, assays were used at below 50% saturation. To determine BRAF and BRAF sensitivity to 885-A, immunoprecipitated BRAF was preincubated with drug in KCl-free buffer for 10 min at room temperature prior to the first-step reaction. To measure the activity of purified BRAF, a 96-well DELFIA-based assay system was used. Full-length rabbit MEK1 protein was expressed with a GST tag at the N-terminus and a C-terminal histidine tag in Escherichia coli JM109 bacteria and purified by nickel-agarose affinity chromatography. Full length BRAF protein was generated by infection of SF9 insect cells with a recombinant baculovirus expressing full-length human BRAF with an N-terminal histidine tag and purified as above. For the kinase assays, all incubations were at room temperature with shaking. 4 μg GST-MEK1, 100-200ng purified BRAF and 1 μl inhibitor at the required concentrations (0.001 to 100 μM final concentration) were added to the wells of glutathione-coated plates and preincubated for 10 min. ATP in DELFIA assay buffer (20 μL; Table S3), to give a final concentration of 100 μM, was added to each well, and the plates were incubated for 45 min. The plates were washed 3X with 200 μl 0.1% tween20/water. Primary antibody (rabbit anti-phospho MEK1/2 diluted 1/2000, Cell Signaling Technologies) and Eu-labeled anti-rabbit secondary antibody (diluted 1/1000, Perkin-Elmer) were preincubated for 30 min and 100 μl was added to the plates and incubated for a further hour. The plates were washed as before, and 100 μl DELFIA enhancement solution (Perkin-Elmer, Turku, Finland) was added. The plates were sealed and incubated for 30 min and europium counts measured on Spectramax M5 plate reader (Molecular Devices, Wokingham, UK). IC50 values were determined using GraphPad Prism (Graphpad Software Inc., San Diego, CA, USA). Experiments were performed under Home Office license authority with local Ethics Committee approval. To activate CreERT2, four doses of tamoxifen (Sigma; 10mg each in 100% ethanol) were applied topically to the shaven skin on the backs of the mice every other day for 7 days. Genotyping was performed by PCR using DNA prepared with DNeasy kits (QIAGEN). Braf and Braf were analyzed using primers: A) 5′-GCCCAGGCTCTTTATGAGAA-3; B) 5′-AGTCAATCATCCACAGAGACCT-3′; and C) 5′-GCTTGGCTGGACGTAAACTC-3′. A+B detects the wild-type BRAF allele (466 bp product) and Braf, the Cre-recombinase recombined allele (518pb product). A+C detects the targeted allele Braf (140 bp). Tyr::CreERT2 was analyzed using primers 5′-GAAGCAACTCATCGATTG-3′ and 5′-TGAAGGGTCTGGTAGGATCA-3′. Kras was analyzed using primers 5′-CGCAGACTGTAGAGCAGCG-3′ and 5′-CCATGGCTTGAGTAAGTCTGC-3′. mRNA expression analysis was performed by RT-PCR. RNA was prepared using the RNEasy kit (QIAGEN). First-strand cDNA synthesis was performed with 500ng total RNA and random hexanucleotides (Random Primers, Invitrogen). Tyrosinase (Tyr), was detected using primers 5′-TGGTTCCTTTCATACCGCTC-3′ and 5′-CAGATACGACTGGCTTGTTCC-3′; Dct with 5′-GCAAGATTGCCTGTCTCTCC-3′ and 5′-AGTCCAGTGTTCCGTCTGCT-3′; Pax3 with 5′-CCAGGATGATGCGGCCCGGCCCGGG-3′ and 5′-AGGATGCGGCTGATAGAACTCACTG-3′; and silver/gp100 (Si) with 5′-GGAGAGGTGGCCAGGTATC-3′ and 5′-CAGTAATGGTGAAGGTTGAAC-3′. The control Gapdh was detected with 5′- GATGGCCCCTCGGAAAGCT-3′ 5′-CCAGTGAGCTTCCCGTTCAGC-3′. To sequence Kras cDNA, a 238 bp product from Kras cDNA was PCR amplified using primers 5′-GGCGGCAGCGCTGTGGCGGCG-3′ and 5′-CGTAGGGTCATACTCATCCAC-3′ and directly sequenced using these primers and automated dideoxy sequencing. For immunohistochemistry (IHC), tumors were fixed in 10% buffered formalin and embedded in paraffin. Sections (3-10μm) were stained with hematoxylin and eosin using standard protocols. For S100 and Ki67 staining, antigen retrieval was performed in citrate buffer (pH 6.0, 30 min) and revealed using a rabbit polyclonal antibody (Dako, 1/1000), the Rabbit Envision Peroxidase kit and the AEC substrate chromogen (Dako) for S100, and a rat monoclonal antibody (Dako,1/25), the rat Vectastain ABC kit (Vector Labs, USA) and DAB as chromagen for Ki67. Positive (a well characterized sample of mouse melanoma) and negative (omission of the primary antibody and substitution with preimmune serum) controls were included in each slide run. Immunohistochemical staining was analyzed by two of the authors on a multi-headed microscope. Tumor cell lines were established by mechanically dissociating tumors in DMEM/20%FCS/Primocin (0.1mg/ml - InvivoGen) and clonal lines were selected by limiting dilution.
|
PMC5785775
|
Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks
|
In de novo drug design, computational strategies are used to generate novel molecules with good affinity to the desired biological target. In this work, we show that recurrent neural networks can be trained as generative models for molecular structures, similar to statistical language models in natural language processing. We demonstrate that the properties of the generated molecules correlate very well with the properties of the molecules used to train the model. In order to enrich libraries with molecules active toward a given biological target, we propose to fine-tune the model with small sets of molecules, which are known to be active against that target. Against Staphylococcus aureus, the model reproduced 14% of 6051 hold-out test molecules that medicinal chemists designed, whereas against Plasmodium falciparum (Malaria), it reproduced 28% of 1240 test molecules. When coupled with a scoring function, our model can perform the complete de novo drug design cycle to generate large sets of novel molecules for drug discovery.Chemistry is the language of nature. Chemists speak it fluently and have made their discipline one of the true contributors to human well-being, which has “change[d] the way you live and die”. This is particularly true for medicinal chemistry. However, creating novel drugs is an extraordinarily hard and complex problem. One of the many challenges in drug design is the sheer size of the search space for novel molecules. It has been estimated that 10 drug-like molecules could possibly be synthetically accessible. Chemists have to select and examine molecules from this large space to find molecules that are active toward a biological target. Active means for example that a molecule binds to a biomolecule, which causes an effect in the living organism, or inhibits replication of bacteria. Modern high-throughput screening techniques allow testing of molecules on the order of 10 in the lab. However, larger experiments will get prohibitively expensive. Given this practical limitation of in vitro experiments, it is desirable to have computational tools to narrow down the enormous search space. Virtual screening is a commonly used strategy to search for promising molecules among millions of existing or billions of virtual molecules. Searching can be carried out using similarity-based metrics, which provides a quantifiable numerical indicator of closeness between molecules. In contrast, in de novo drug design, one aims to directly create novel molecules that are active toward the desired biological target. Here, like in any molecular design task, the computer has to(i)create molecules,(ii)score and filter them, and(iii)search for better molecules, building on the knowledge gained in the previous steps. create molecules, score and filter them, and search for better molecules, building on the knowledge gained in the previous steps. Task i, the generation of novel molecules, is usually solved with one of two different protocols. One strategy is to build molecules from predefined groups of atoms or fragments. Unfortunately, these approaches often lead to molecules that are very hard to synthesize. Therefore, another established approach is to conduct virtual chemical reactions based on expert coded rules, with the hope that these reactions could then also be applied in practice to make the molecules in the laboratory. These systems give reasonable drug-like molecules and are considered as “the solution” to the structure generation problem. We generally share this view. However, we have recently shown that the predicted reactions from these rule-based expert systems can sometimes fail. Also, focusing on a small set of robust reactions can unnecessarily restrict the possibly accessible chemical space. Task ii, scoring molecules and filtering out undesired structures, can be solved with substructure filters for undesirable reactive groups in conjunction with established approaches such as docking or machine learning (ML) approaches. The ML approaches are split into two branches: Target prediction classifies molecules into active and inactive, and quantitative structure–activity relationships (QSAR) seek to quantitatively predict a real-valued measure for the effectiveness of a substance (as a regression problem). As molecular descriptors, signature fingerprints, extended-connectivity (ECFP), and atom pair (APFP) fingerprints and their fuzzy variants are the de facto standard today. Convolutional networks on graphs are a more recent addition to the field of molecular descriptors. Jastrzebski et al. proposed to use convolutional neural networks to learn descriptors directly from SMILES strings. Random forests, support vector machines, and neural networks are currently the most widely used machine learning models for target prediction. This leads to task iii, the search for molecules with the right binding affinity combined with optimal molecular properties. In earlier work, this was performed (among others) with classical global optimization techniques, for example genetic algorithms or ant-colony optimization. Furthermore, de novo design is related to inverse QSAR. While in de novo design a regular QSAR mapping X → y from molecular descriptor space X to properties y is used as the scoring function for the global optimizer, in inverse QSAR one aims to find an explicit inverse mapping y → X, and then maps back from optimal points in descriptor space X to valid molecules. However, this is not well-defined, because molecules are inherently discrete (the space is not continuously populated), and the mapping from a target property value y to possible structures X is one-to-many, as usually several different structures with very similar properties can be found. Several protocols have been developed to address this, for example enumerating all structures within the constraints of hyper-rectangles in the descriptor space. Gómez-Bombarelli et al. proposed to learn continuous representations of molecules with variational autoencoders, based on the model by Bowman et al., and to perform Bayesian optimization in this vector space to optimize molecular properties. While promising, this approach was not applied to create active drug molecules and often produced syntactically invalid molecules and highly strained or reactive structures, for example cyclobutadienes. In this work, we suggest a complementary, completely data-driven de novo drug design approach. It relies only on a generative model for molecular structures, based on a recurrent neural network, that is trained on large sets of molecules. Generative models learn a probability distribution over the training examples; sampling from this distribution generates new examples similar to the training data. Intuitively, a generative model for molecules trained on drug molecules would “know” how valid and reasonable drug-like molecules look and could be used to generate more drug-like molecules. However, for molecules, these models have been studied rarely, and rigorously only with traditional models such as Gaussian mixture models (GMM). Recently, recurrent neural networks (RNNs) have emerged as powerful generative models in very different domains, such as natural language processing, speech, images, video, formal languages, computer code generation, and music scores. In this work, we highlight the analogy of language and chemistry, and show that RNNs can also generate reasonable molecules. Furthermore, we demonstrate that RNNs can also transfer their learned knowledge from large molecule sets to directly produce novel molecules that are biologically active by retraining the models on small sets of already known actives. We test our models by reproducing hold-out test sets of known biologically active molecules. To connect chemistry with language, it is important to understand how molecules are represented. Usually, they are modeled by molecular graphs, also called Lewis structures in chemistry. In molecular graphs, atoms are labeled nodes. The edges are the bonds between atoms, which are labeled with the bond order (e.g., single, double, or triple). One could therefore envision having a model that reads and outputs graphs. Several common chemistry formats store molecules in such a manner. However, in models for natural language processing, the input and output of the model are usually sequences of single letters, strings or words. We therefore employ the SMILES format, which encodes molecular graphs compactly as human-readable strings. SMILES is a formal grammar which describes molecules with an alphabet of characters, for example c and C for aromatic and aliphatic carbon atoms, O for oxygen, and −, =, and # for single, double, and triple bonds (see Figure 1). To indicate rings, a number is introduced at the two atoms where the ring is closed. For example, benzene in aromatic SMILES notation would be c1ccccc1. Side chains are denoted by round brackets. To generate valid SMILES, the generative model would have to learn the SMILES grammar, which includes keeping track of rings and brackets to eventually close them. In morphine, a complex natural product, the number of steps between the first 1 and the second 1, indicating a ring, is 32. Having established a link between molecules and (formal) language, we can now discuss language models. Examples of molecules and their SMILES representation. To correctly create smiles, the model has to learn long-term dependencies, for example, to close rings (indicated by numbers) and brackets. Given a sequence of words (w1, ..., wi), language models predict the distribution of the (i+1)th word wi+1. For example, if a language model received the sequence “Chemistry is”, it would assign different probabilities to possible next words: “fascinating”, “important”, or “challenging” would receive high probabilities, while “runs” or “potato” would receive very low probabilities. Language models can both capture the grammatical correctness (“runs” in this sentence is wrong) and the meaning (“potato” does not make sense). Language models are implemented, for example, in message autocorrection in many modern smartphones. Interestingly, language models do not have to use words. They can also be based on characters or letters. In that case, when receiving the sequence of characters chemistr, it would assign a high probability to y, but a low probability to q. To model molecules instead of language, we simply swap words or letters with atoms, or, more practically, characters in the SMILES alphabet, which form a (formal) language. For example, if the model receives the sequence c1ccccc, there is a high probability that the next symbol would be a “1”, which closes the ring, and yields benzene. More formally, to a sequence S of symbols si at steps ti ∈ T, the language model assigns a probability of1where the parameters θ are learned from the training set. In this work, we use a recurrent neural network (RNN) to estimate the probabilities of eq 1. In contrast to regular feedforward neural networks, RNNs maintain state, which is needed to keep track of the symbols seen earlier in the sequence. In abstract terms, an RNN takes a sequence of input vectors x1:n = (x1, ..., xn) and an initial state vector h0, and returns a sequence of state vectors h1:n = (h1, ..., hn) and a sequence of output vectors y1:n = (y1, ..., yn). The RNN consists of a recursively defined function R, which takes a state vector hi and input vector xi+1 and returns a new state vector hi+1. Another function O maps a state vector hi to an output vector yi.234 The state vector hi stores a representation of the information about all symbols seen in the sequence so far. As an alternative to the recursive definition, the recurrent network can also be unrolled for finite sequences (see Figure 2). An unrolled RNN can be seen as a very deep neural network, in which the parameters θ are shared among the layers, and the hidden state ht is passed as an additional input to the next layer. Training the unrolled RNN to fit the parameters θ can then simply be done by using backpropagation to compute the gradients with respect to the loss function, which is categorical cross-entropy in this work. (a) Recursively defined RNN. (b) The same RNN, unrolled. The parameters θ (the weight matrices of the neural network) are shared over all time steps. As the specific RNN function, in this work, we use the long short-term memory (LSTM), which was introduced by Hochreiter and Schmidhuber. It has been used successfully in many natural language processing tasks, for example in Google’s neural machine translation system. For excellent in-depth discussions of the LSTM, we refer to the articles by Goldberg, Graves, Olah, and Greff et al. To encode the SMILES symbols as input vectors xt, we employ the “one-hot” representation. This means that if there are K symbols, and k is the symbol to be input at step t, then we can construct an input vector xt with length K, whose entries are all zero except the kth entry, which is one. If we assume a very restricted set of symbols , input c would correspond to xt = (1, 0, 0), 1 to xt = (0, 1, 0), and to xt = (0, 0, 1). The probability distribution Pθ(st+1|st, ..., s1) of the next symbol given the already seen sequence is thus a multinomial distribution, which is estimated using the output vector yt of the recurrent neural network at time step t by5where yt corresponds to the kth element of vector yt. Sampling from this distribution would then allow generating novel molecules: After sampling a SMILES symbol st+1 for the next time step t + 1, we can construct a new input vector xt+1, which is fed into the model, and via yt+1 and eq 5 yields Pθ(st+2|st+1, ..., s1). Sampling from the latter generates st+2, which serves again also as the model’s input for the next step (see Figure 3). This symbol-by-symbol sampling procedure is repeated until the desired number of characters have been generated. Symbol generation and sampling process. We start with a random seed symbol s1, here c, which gets converted into a one-hot vector x1 and input into the model. The model then updates its internal state h0 to h1 and outputs y1, which is the probability distribution over the next symbols. Here, sampling yields s2 = 1. Converting s2 to x2 and feeding it to the model leads to updated hidden state h2 and output y2, from which we can sample again. This iterative symbol-by-symbol procedure can be continued as long as desired. In this example, we stop it after observing an EOL () symbol, and obtain the SMILES for benzene. The hidden state hi allows the model to keep track of opened brackets and rings, to ensure that they will be closed again later. To indicate that a molecule is “completed”, each molecule in our training data finishes with an “end of line” (EOL) symbol, in our case the single character (which means that the training data is just a simple SMILES file). Thus, when the system outputs an EOL, a generated molecule is finished. However, we simply continue sampling, thus generating a regular SMILES file that contains one molecule per line. In this work, we used a network with three stacked LSTM layers, using the Keras library. The model was trained with back-propagation through time, using the ADAM optimizer at standard settings. To mitigate the problem of exploding gradients during training, a gradient norm clipping of 5 is applied. For many machine learning tasks, only small data sets are available, which might lead to overfitting with powerful models such as neural networks. In this situation, transfer learning can help. Here, a model is first trained on a large data set for a different task. Then, the model is retrained on the smaller data set, which is also called fine-tuning. The aim of transfer learning is to learn general features on the bigger data set, which also might be useful for the second task in the smaller data regime. To generate focused molecule libraries, we first train on a large, general set of molecules, then perform fine-tuning on a smaller set of specific molecules, and after that start the sampling procedure. To verify whether the generated molecules are active on the desired targets, standard target prediction was employed. Machine learning based target prediction aims to learn a classifier c: M → to decide whether a molecule m ∈ molecular descriptor space M is active or not against a target. The molecules are split into actives and inactives using a threshold on a measure for the substance effectiveness. pIC50 = −log10(IC50) is one of the most widely used metrics for this purpose. IC50 is the half maximal inhibitory concentration, that is, the concentration of drug that is required to inhibit 50% of a biological target’s function in vitro. To predict whether the generated molecules are active toward the biological target of interest, target prediction models (TPMs) were trained for all the tested targets (5-HT2A, Plasmodium falciparum and Staphylococcus aureus). We evaluated random forest, logistic regression, (deep) neural networks, and gradient boosting trees (GBT) as models with ECFP4 (extended connectivity fingerprint with a diameter of 4) as the molecular descriptor. We found that GBTs slightly outperformed all other models and used these as our virtual assay in all studies (AUC[5-HT2A] = 0.877, AUC[Staph. aur.] = 0.916). ECFP4 fingerprints were generated with CDK version 1.5.13. scikit-learn, XGBoost, and Keras were used as the machine learning libraries. For 5-HT2A and Plasmodium, molecules are considered as active for the TPM if their IC50 reported in ChEMBL is <100 nM, which translates to a pIC50 > 7, whereas for Staphylococcus, we used pMIC > 3. The chemical language model was trained on a SMILES file containing 1.4 million molecules from the ChEMBL database, which contains molecules and measured biological activity data. The SMILES strings of the molecules were canonicalized (which means finding a unique representation that is the same for isomorphic molecular graphs) before training with the CDK chemoinformatics library, yielding a SMILES file that contained one molecule per line. It has to be noted that ChEMBL contains many peptides, natural products with complex scaffolds, Michael acceptors, benzoquinones, hydroxylamines, hydrazines, etc., which is reflected in the generated structures (see below). This corresponds to 72 million individual characters, with a vocabulary size of 51 unique characters. 51 characters is only a subset of all SMILES symbols, since the molecules in ChEMBL do not contain many of the heavy elements. As we have to set the number of symbols as a hyperparameter during model construction, and the model can only learn the distribution over the symbols present in the training data, this implies that only molecules with these 51 SMILES symbols seen during training can be generated during sampling. The 5-HT2A, the Plasmodium falciparum, and the Staphylococcus aureus data sets were also obtained from ChEMBL. As these molecules were intended to be used in the rediscovery studies, they were removed from the training data before fitting the chemical language model. To evaluate the models for a test set T, and a set of molecules GN generated from the model by sampling, we report the ratio of reproduced molecules , and enrichment over random (EOR), which is defined as6where n = |GN ∩ T| is the number of reproduced molecules from T by sampling a set GN of |GN| = N molecules from the fine-tuned generative model, and m = |RM ∩ T| is the number of reproduced molecules from T by sampling a set RM of |RM| = M molecules from the generic, unbiased generative model trained only on the large data set. Intuitively, EOR indicates how much better the fine-tuned models work when compared to the general model. In this work, we address two points: First, we want to generate large sets of diverse molecules for virtual screening campaigns. Second, we want to generate smaller, focused libraries enriched with possibly active molecules for a specific target. For the first task, we can train a model on a large, general set of molecules to learn the SMILES grammar. Sampling from this model would generate sets of diverse, but unfocused molecules. To address the second task, and to obtain novel active drug molecules for a target of interest, we perform transfer learning: We select a small set of known actives for that target and we refit our pretrained chemical language model with this small data set. After each epoch, we sample from the model to generate novel actives. Furthermore, we investigate if the model actually benefits from transfer learning, by comparing it to a model trained from scratch on the small sets without pretraining. We employed a recurrent neural network with three stacked LSTM layers, each with 1024 dimensions, and each one followed by a dropout layer, with a dropout ratio of 0.2, to regularize the neural network. The model was trained until convergence, using a batch size of 128. The RNN was unrolled for 64 steps. It had 21.3 × 10 parameters. During training, we sampled a few molecules from the model every 1000 minibatches to inspect progress. Within a few 1000 steps, the model starts to output valid molecules (see Table 1). To generate novel molecules, 50,000,000 SMILES symbols were sampled from the model symbol-by-symbol. This corresponded to 976,327 lines, from which 97.7% were valid molecules after parsing with the CDK toolkit. Removing all molecules already seen during training yielded 864,880 structures. After filtering out duplicates, we obtained 847,955 novel molecules. A few generated molecules were randomly selected and depicted in Figure 4. The Supporting Information contains more structures. The created structures are not just formally valid but also mostly chemically reasonable. A few randomly selected, generated molecules. Ad = Adamantyl. In order to check if the de novo compounds could be considered as valid starting points for a drug discovery program, we applied the internal AstraZeneca filters. At AstraZeneca, this flagging system is used to determine if a compound is suitable to be part of the high-throughput screening collection (if flagged as “core” or “backup”) or should be restricted for particular use (flagged as “undesirable” since it contains one or several unwanted substructures, e.g., undesired reactive functional groups). The filters were applied to the generated set of 848 k molecules, and they flagged most of them, 640 k (75%), as either core or backup. Since the same ratio (75%) of core and backup compounds has been observed for the ChEMBL collection, we therefore conclude that the algorithm generates preponderantly valid screening molecules and faithfully reproduces the distribution of the training data. To determine whether the properties of the generated molecules match the properties of the training data from ChEMBL, we followed the procedure of Kolb: We computed several molecular properties, namely, molecular weight, BertzCT, the number of H-donors, H-acceptors, and rotatable bonds, logP, and total polar surface area for randomly selected subsets from both sets with the RDKit library version 2016.03.1. Then, we performed dimensionality reduction to 2D with t-SNE (t-distributed stochastic neighbor embedding, a technique analogous to PCA), which is shown in Figure 5. Both sets overlap almost completely, which indicates that the generated molecules very well recreate the properties of the training molecules. t-SNE projection of 7 physicochemical descriptors of random molecules from ChEMBL (blue) and molecules generated with the neural network trained on ChEMBL (green), to two unitless dimensions. The distributions of both sets overlap significantly. Furthermore, we analyzed the Bemis–Murcko scaffolds of the training molecules and the sampled molecules. Bemis–Murcko scaffolds contain the ring systems of a molecule and the moieties that link these ring systems, while removing any side chains. They represent the scaffold, or “core” of a molecule, which series of drug molecules often have in common. The number of common scaffolds in both sets divided by the union of all scaffolds in both sets (Jaccard index) is 0.12, which indicates that the language model does not just modify side chain substituents but also introduces modifications at the molecular core. To generate novel ligands for the 5-HT2A receptor, we first selected all molecules with pIC50 > 7 which were tested on 5-HT2A from ChEMBL (732 molecules), and then fine-tuned our pretrained chemical language model on this set. After each epoch, we sampled 100,000 chars, canonicalized the molecules, and removed any sampled molecules that were already contained in the training set. Following this, we evaluated the generated molecules of each round of retraining with our 5-HT2A target prediction model (TPM). In Figure 6, the ratio of molecules predicted to be active by the TPM after each round of fine-tuning is shown. Before fine-tuning (corresponding to epoch 0), the model generates almost exclusively inactive molecules. Already after 4 epochs of fine-tuning the model produced a set in which 50% of the molecules are predicted to be active. Epochs of fine-tuning vs ratio of actives. In order to assess the novelty of the de novo molecules generated with the fine-tuned model, a nearest neighbor similarity/diversity analysis has been conducted using a commonly used 2D fingerprint (ECFP4) based similarity method (Tanimoto index).Figure 7 shows the distribution of the nearest neighbor Tanimoto index generated by comparing all the novel molecules and the training molecules before and after n epochs of fine-tuning. For each bin, the white bars indicate the molecules generated from the unbiased, general model, while the darker bars indicate the molecules after several epochs of fine-tuning. Within the bins corresponding to lower similarity, the number of molecules decreases, while the bins of higher similarity get populated with increasing numbers of molecules. The plot thus shows that the model starts to output more and more similar molecules to the target-specific training set. Notably, after a few rounds of training not only are highly similar molecules produced but also molecules covering the whole range of similarity, indicating that our method could deliver not only close analogues but also new chemotypes or scaffold ideas to a drug discovery project. To have the best of both worlds, that is, diverse and focused molecules, we therefore suggest to sample after each epoch of retraining and not just after the final epoch. Nearest-neighbor Tanimoto similarity distribution of the generated molecules for 5-HT2A after n epochs of fine-tuning against the known actives. The generated molecules are distributed over the whole similarity range. Generated molecules with a medium similarity can be interesting for scaffold-hopping. Plasmodium falciparum is a parasite that causes the most dangerous form of malaria. To probe our model on this important target, we used a more challenging validation strategy. We wanted to investigate whether the model could also propose the same molecules that medicinal chemists chose to evaluate in published studies. To test this, first, the known actives against Plasmodium falciparum with a pIC50 > 8 were selected from ChEMBL (Table 2). Then, this set was split randomly into a training (1239 molecules) and a test set (1240 molecules). The chemical language model was then fine-tuned on the training set. 7500 molecules were sampled after each of the 20 epochs of refitting. EOR: Enrichment over random. This yielded 128,256 unique molecules. Interestingly, we found that our model was able to “redesign” 28% of the unseen molecules of the test set. In comparison to molecules sampled from the unspecific, untuned model, an enrichment over random (EOR) of 66.9 is obtained. With a smaller training set of 100 molecules, the model can still reproduce 7% of the test set, with an EOR of 19.0. To test the reliance on pIC50 we chose to use another cutoff of pIC50 > 9, and took 100 molecules in the training set and 1022 in the test set. 11% of the test set could be recreated, with an EOR of 35.7. To visually explore how the model populates chemical space, Figure 8 shows a t-SNE plot of the ECFP4 fingerprints of the test molecules and 2000 generated molecules that were predicted to be active by the target prediction model for Plasmodium falciparum. It indicates that the model has generated many similar molecules around the test examples. t-SNE plot of the pIC50 > 9 test set (blue) and the de novo molecules predicted to be active (green). The language model populates chemical space around the test molecules. To evaluate a different target, we furthermore conducted a series of experiments to reproduce known active molecules against Staphylococcus aureus. Here, we used actives with a pMIC > 3. MIC is the mean inhibitory concentration, the lowest concentration of a compound that prevents visible growth of a microorganism. As above, the actives were split into a training and a test set. However, here, the availability of the data allows larger test sets to be used. After fine-tuning on the training set of 1000 molecules (Table 3, entry 1), our model could retrieve 14% of the 6051 test molecules. When scaling down to a smaller training set of 50 molecules (the model gets trained on less than 1% of the data!), it can still reproduce 2.5% of the test set, and performs 21.6 times better than the unbiased model (Table 3, entry 2). Using a lower learning rate (0.0001, entry 3) for fine-tuning, which is often done in transfer learning, does not work as well as the standard learning rate (0.001, entry 2). We additionally examined whether the model benefits from transfer learning. When trained from scratch, the model performs much worse than the pretrained and subsequently fine-tuned model (see Figure 9 and Table 3, entry 4). Pretraining on the large data set is thus crucial to achieve good performance against Staphylococcus aureus. EOR: Enrichment over random. Fine-tuning learning rate = 10. No pretraining. 8 generate-test cycles. Different training strategies on the Staphylococcus aureus data set with 1000 training and 6051 test examples. Fine-tuning the pretrained model performs better than training from scratch (lower test loss [cross entropy] is better). The experiments we conducted so far are applicable if one already knows several actives. However, in drug discovery, one often does not have such a set to start with. Therefore, high throughput screenings are conducted to identify a few hits, which serve as a starting point for the typical cyclical drug discovery process: Molecules get designed, synthesized, and then tested in assays. Then, the best molecules are selected, and based on the gained knowledge new molecules are designed, which closes the cycle. Therefore, as a final challenge for our model, we simulated this cycle by iterating molecule generation (“synthesis”), selection of the best molecules with the machine learning based target prediction (“virtual assay”), and retraining the language model with the best molecules (“design”) with Staphylococcus aureus as the target. We thus do not use a set of known actives to start the structure generation procedure (see Figure 10). Scheme of our de novo design cycle. Molecules are generated by the chemical language model and then scored with the target prediction model (TPM). The inactives are filtered out, and the RNN is retrained. Here, the TPM is a machine learning model, but it could also be a robot conducting synthesis and biological assays, or a docking program. We started with 100,000 sampled molecules from the unbiased chemical language model. Then, using our target prediction model, we extracted the molecules classified as actives. After that, the RNN was fine-tuned for 5 epochs on the actives, sampling ≈10,000 molecules after each epoch. The resulting molecules were filtered with the target prediction model, and the new actives appended to the actives from the previous round, closing the loop. Already after 8 iterations, the model reproduced 416 of the 7001 test molecules from the previous task, which is 6% (Table 3, entry 5), and exhibits an EOR of 59.6. This EOR is higher than if the model is retrained directly on a set of 50 actives (entry 2). Additionally, we obtained 60,988 unique molecules that the target prediction model classified as active. This suggests that, in combination with a target prediction or scoring model, our model can at least simulate the complete de novo design cycle. Our results indicate that the general model trained on a large molecule set has learned the SMILES rules and can output valid, drug-like molecules, which resemble the training data. However, sampling from this model does not help much if we want to generate actives for a specific target: We would have to generate very large sets to find actives for that target among the diverse range of molecules the model creates, which is indicated by the high EOR scores in our experiments. When fine-tuned to a set of actives, the probability distribution over the molecules captured by our model is shifted toward molecules active toward our target. To study this, we compare the Levenshtein (string edit) distance of the generated SMILES to their nearest neighbors in the training set in Figure 11. The Levenshtein distance of, e.g., benzene c1ccccc1 and pyridine c1ccncc1 would be 1. Figure 11 shows that while the model often seems to have made small replacements in the underlying SMILES, in many cases it also made more complex modifications or even generated completely different SMILES. This is supported also by the distribution of the nearest neighbor fingerprint similarities of training and rediscovered molecules (ECFP4, Tanimoto, Figure 12). Many rediscovered molecules are in the medium similarity regime. Histogram of Levenshtein (string edit) distances of the SMILES of the reproduced molecules to their nearest neighbor in the training set (Staphylococcus aureus, model retrained on 50 actives). While in many cases the model makes changes of a few symbols in the SMILES, resembling the typical modifications applied when exploring series of compounds, the distribution of the distances indicates that the RNN also performs more complex changes by introducing larger moieties or generating molecules that are structurally different, but isofunctional to the training set. Violin plot of the nearest-neighbor ECFP4-Tanimoto similarity distribution of the 50 training molecules against the rediscovered molecules in Table 3, entry 2. The distribution suggests that the model has learned to make typical small functional group replacements, but can also reproduce molecules which are not too similar to the training data. Because we perform transfer learning, during fine-tuning, the model does not “forget” what it has learned. A plausible explanation why the model works is therefore that it can transfer the modifications that are regularly applied when series of molecules are studied, to the molecules it has seen during fine-tuning. In this work, we have shown that recurrent neural networks based on the long short-term memory (LSTM) can be applied to learn a statistical chemical language model. The model can generate large sets of novel molecules with similar physicochemical properties to the training molecules. This can be used to generate libraries for virtual screening. Furthermore, we demonstrated that the model performs transfer learning when fine-tuned to smaller sets of molecules active toward a specific biological target, which enables the creation of novel molecules with the desired activity. By iterating cycles of structure generation with the language model, scoring with a target prediction model (TPM) and retraining of the model with increasingly larger sets of highly scored molecules, we showed that we do not even need a set of known active molecules to start our procedure with, as the TPM could also be a docking program, or a robot conducting synthesis and biological testing. We see three main advantages of our method. First, it is conceptually orthogonal to established molecule generation approaches, as it learns a generative model for molecular structures. Second, our method is very simple to set up, to train, and to use; it can be adapted to different data sets without any modifications to the model architecture; and it does not depend on hand-encoded expert knowledge. Furthermore, it merges structure generation and optimization in one model. A weakness of our model is interpretability. In contrast, existing de novo design methods settled on virtual reactions to generate molecules, which has advantages as it minimizes the chance of obtaining “overfit”, weird molecules, and increases the chances to find synthesizable compounds. To extend our work, it is just a small step to cast molecule generation as a reinforcement learning problem, where the pretrained LSTM generator could be seen as a policy, which can be encouraged to create better molecules with a reward signal obtained from a target prediction model. In addition, different approaches for target prediction, for example, docking, could be evaluated. Deep learning is not a panacea, and we join Gawehn et al. in expressing “some healthy skepticism” regarding its application in drug discovery. Generating molecules that are almost right is not enough, because in chemistry, a miss is as good as a mile, and drug discovery is a “needle in the haystack” problem—in which also the needle looks like hay. Nevertheless, given that we have shown in this work that our model can rediscover those needles, and other recent developments, we believe that deep neural networks can be complementary to established approaches in drug discovery. The complexity of the problem certainly warrants the investigation of novel approaches. Eventually, success in the wet lab will determine if the new wave of neural networks will prevail.
|
PMC3629873
|
Preliminary evaluation of the CellFinder literature curation pipeline for gene expression in kidney cells and anatomical parts
|
Biomedical literature curation is the process of automatically and/or manually deriving knowledge from scientific publications and recording it into specialized databases for structured delivery to users. It is a slow, error-prone, complex, costly and, yet, highly important task. Previous experiences have proven that text mining can assist in its many phases, especially, in triage of relevant documents and extraction of named entities and biological events. Here, we present the curation pipeline of the CellFinder database, a repository of cell research, which includes data derived from literature curation and microarrays to identify cell types, cell lines, organs and so forth, and especially patterns in gene expression. The curation pipeline is based on freely available tools in all text mining steps, as well as the manual validation of extracted data. Preliminary results are presented for a data set of 2376 full texts from which >4500 gene expression events in cell or anatomical part have been extracted. Validation of half of this data resulted in a precision of ∼50% of the extracted data, which indicates that we are on the right track with our pipeline for the proposed task. However, evaluation of the methods shows that there is still room for improvement in the named-entity recognition and that a larger and more robust corpus is needed to achieve a better performance for event extraction. Database URL: http://www.cellfinder.org/Biomedical literature curation is the process of automatically and/or manually compiling biological data from scientific publications and making it available in a structured and comprehensive way. Databases that integrate information derived in some way from scientific publications include, for instance, model organism databases (1), protein–protein interactions (2) and gene–chemical–disease relationships (3). Typical literature curation workflows include the following steps (4): triage (selection of relevant publications), biological entities identification (e.g. genes/proteins, diseases, etc.), extraction of relationships (e.g. protein–protein interactions, gene expression, etc.), association of biological processes with experimental evidence, data validation and recoding into the database. Therefore, literature curation requires a careful reading of publications by domain experts, which is known to be a time-consuming task. Additionally, the increasing growth of available publications prevents a comprehensive manual curation of intended facts and previous studies show that it is not feasible (5). Recent advances in text mining methods have facilitated its application in most of the literature curation stages. Challenges have contributed to the improvement and availability of a variety of methods for named-entity prediction (6), and more specifically for gene/protein prediction and normalization (7, 8). Also binary relationships (9) and event extraction (10) have been improved, and its current performance allows its use on large scale projects (11). Finally, integrated ready-to-use workbenches have also been available, such as @Note (12), Argo (13), MyMiner (14) and Textpresso (15), although the performance and scalability to larger projects is still dubious for some of them. A comparison between some of them is found in this survey on annotation tools for the biomedical domain (16). Previous reports (17, 18) and experiments (19) have confirmed the feasibility of text mining to assist literature curation and recent surveys (4, 20) show that, indeed, it is already part of many biological databases workflows. For instance, text mining support is being explored for the triage stage in FlyBase (21), for curation of regulatory annotation in (22) and also in the AgBase (23), Biomolecular Interaction Network Database (BIND) (24), Immune Epitope Database (IEDB) (25) and The Comparative Toxicogenomics Database (CTD) (26) databases. Additionally, many solutions have been proposed for the CTD database during a recent collaborative task (27). Further, Textpresso has been widely used to prioritize document and for Gene Ontology (GO) terms (28) annotation in WormBase and The Arabidopsis Information Resource (TAIR) (29). Named-entity recognition has also been included in the curation workflow of Mouse Genome Informatics (MGI) (30) for gene/protein extraction, and in Xenbase (31) for gene and anatomy terms, for instance. Finally, few databases have tried automatic relationships extraction methods: protein phosphorylation information has been extracted using rule-based pattern templates (32), recreation of events has been carried out for the Human Protein Interaction Database (HHPID) database (33) and revalidation of relationships for the PharmGKB database (34). We present the first description of the curation pipeline for the CellFinder database (http://www.cellfinder.org/), a repository of cell research, which aims to integrate data derived from many sources, such as literature curation and microarray data. It is based on a novel ontology [Cell: Expression, Localization, Development, Anatomy (CELDA) (http://cellfinder.org/about/ontology)], which allows standardization and integration to other available ontologies on the cell and anatomy domains. Hence, the CellFinder platform provides a framework for comprehensive descriptions of human tissues, cells and commonly used model organisms on molecular and functional levels, in vivo and in vitro. The CellFinder pipeline for literature curation integrates state-of-art freely available tools for the document triage, recognition of a variety of entity types and extraction of biological processes. Curation is carried out for full text documents available at the PubMed Central Open Access (PMC OA) subset (http://www.ncbi.nlm.nih.gov/pmc/tools/openftlist/), and manual intervention from curators is currently only necessary for querying new documents for curation and validation of the derived biological processes. In both cases, web-based tools are being used, which allow their integration into the CellFinder web site. We are not aware of prior usage of available systems for the automatic extraction of biological events. For instance, Xenbase manually annotates gene expression events (31), whereas others databases use proprietary systems (34) or tools, which do not allow re-use for other domains (33). Our literature curation pipeline has been evaluated using a dataset on the kidney cell research. The kidney consists of >26 cell types, which arise and organize into several anatomical structures during a conserved developmental process (35). Kidney disease culminates from a common sclerotic pathway involving epithelial-mesenchymal transition, extracellular matrix remodeling and vascular changes (36). Multiple renal and non-renal (e.g. inflammatory) cell types are involved in these processes, with dynamic gene expression patterns and functions (37). Therefore, to identify relevant research describing cells and their interactions in normal and diseased kidney, we decided to include species-independent experimental and clinical data of renal disease and of kidney development in CellFinder. For the kidney cell use case, information is compiled about characterization of gene expression profiles in cells and other anatomical locations, such as tissues and organs. Hence, named-entity extraction is performed for genes, proteins, cell lines, cell types, tissues and organs. Gene expression events are then extracted between a gene/protein and a certain cell or anatomical part. The sentence below illustrates one such example (PMID 18989465): On the other hand, the podoplanin expression occurs in the differentiating odontoblasts and the expression is sustained in differentiated odontoblasts, indicating that odontoblasts have the strong ability to express podoplanin. On the other hand, the podoplanin expression occurs in the differentiating odontoblasts and the expression is sustained in differentiated odontoblasts, indicating that odontoblasts have the strong ability to express podoplanin. We are aware of only two previous publications, which report extraction of gene expression in anatomical locations from biomedical texts. OpenDMAP (38) uses Protégé and UIMA-based components, and it has been evaluated for three applications: protein transport, protein interactions and cell type-specific gene expression. OpenDMAP extract genes/proteins and cells using A Biomedical Named Entity Recognizer (ABNER) (39) and a short list of trigger words. Relationships between the triple gene-cell-trigger are identified based on manual pattern templates. It reports precision of 64% and recall of 16% from an evaluation of 324 NCBI’s GeneRIFs, which consists of short descriptions of gene functions. A more comprehensive study on the expression of genes in anatomical location was carried out in (40) with the Gene Expression Text Miner system. The work included extending 150 abstracts from the BioNLP corpus (41) with annotations for anatomical parts and cell lines, as well as relationships to the existing gene expression events. Genes/proteins were extracted using GNAT (42), anatomical part and cell line recognition was performed by Linnaeus (43) using 13 anatomical ontologies and one for cell lines. A list of expression triggers was manually built, and association between the entities is also rule-based. Evaluation on the extended 150 abstracts resulted in a precision of almost 60% and a recall of 24%. The next section will describe the CellFinder curation pipeline and the methods that are used in each stage. Results for the experiments performed for most of the steps are shown in the section ‘Results’ followed by discussion on the more important aspects of the pipeline in the section ‘Discussion and future work’. The curation pipeline for the CellFinder database includes the following steps (cf. Figure 1): triage of potential relevant documents, retrieval of full text, linguistic pre-processing, named-entity recognition, post-processing, relationship extraction, manual validation of the results and integration of gene expression events into the database. This section describes details on the methods used in each phase. Figure 1.Overview of the literature curation pipeline for the CellFinder database. It includes the following steps: triage of potential relevant documents, retrieval of full text, preprocessing (sentence splitting, tokenization and parsing), named-entity recognition (genes, proteins, cell lines, cell types, organs, tissues, expression triggers), gene expression events extraction, manual validation of the results and integration into the database. Automatic procedures are shown in red, whereas the manual ones are shown in blue. Overview of the literature curation pipeline for the CellFinder database. It includes the following steps: triage of potential relevant documents, retrieval of full text, preprocessing (sentence splitting, tokenization and parsing), named-entity recognition (genes, proteins, cell lines, cell types, organs, tissues, expression triggers), gene expression events extraction, manual validation of the results and integration into the database. Automatic procedures are shown in red, whereas the manual ones are shown in blue. Document triage is usually the first step in any literature curation workflow and consists of retrieving potential relevant publications for manual curation or for further processing by a text mining pipeline. In the CellFinder project, we aim to curate only full texts documents, which are available for text mining purposes, i.e. the ones included in the PMC OA subset. Although it is a much smaller collection than the whole Medline, this subset currently contains >200 000 documents. In our pipeline, document triage was performed by querying MedlineRanker (44), a machine learning based text categorization system. We have performed eight queries to MedlineRanker as follows: ‘kidney tubular epithelial EMT’, ‘kidney vascular endothelial interstitium’, ‘kidney glomerular basement membrane’, ‘kidney mesangial space podocyte’, ‘kidney development differentiation pronephros’, ‘kidney extra cellular matrix, ‘kidney regeneration mesenchymal precursor’ and ‘corticomedullary junction’. The search terms were aimed to identify cells, genes and structures that relate to cells contained in nephrons and tubules, such as epithelial cells, endothelial cells and podocytes, as well as cell changes associated with mesenchymal–epithelial transition (EMT) and fibrosis, changes in extracellular matrix and relevant proteins and in cells during kidney development, such as mesenchymal precursor cells. Each query retrieved a list of 10 000 (MedlineRanker’s cut-off) potential PMC relevant documents, including many repeated documents found across lists. After a post-processing step, which included verification on whether documents were part of the PMC OA subset and exclusion of repeated entries, a list of 2376 documents was derived. Documents were retrieved from PMC and were processed through our text mining pipeline. Full texts documents were first split by sentences using the OpenNLP toolkit (http://opennlp.apache.org/) and then parsed by the Brown Laboratory for Linguistic Information Processing (BLLIP) parser (https://github.com/dmcc/bllip-parserV) (45) (also known as McClosky-Charniak parser). Part-of-speech tags, tokenization and full parsing were derived from the BLLIP parser output. Dependency trees were built using the Stanford parser (http://nlp.stanford.edu/software/lex-parser.shtml). Part-of-speech, tokenization and parsing information are only necessary for the gene expression extraction (cf. ‘Event Extraction’ below). Named-entity recognition has been performed for five entity types: genes/proteins, cell lines, cell types, anatomical parts and gene expression triggers. Extraction is based on available state-of-art systems and dictionary or ontology-based approaches, without any adaption nor retraining. Methods are similar to the ones investigated in previous experiments performed with the CellFinder corpus (46). To enable data integration into the CellFinder database, all extracted mentions must be normalized to any of the ontologies or terminologies currently supported by our database: Cell Ontology (CL) (47), Cell Line Ontology (CLO) (48), EHDAA2 (49), Experimental Factor Ontology (EFO) (50), Foundational Model of Anatomy (FMA) (51), GO (52), Adult Mouse Anatomy (MA) (53) and Uberon (54). We identify genes using GNAT (42), a system for extraction and normalization of gene and protein mentions. GNAT assigns confidence scores (up to 1.0) to the gene/protein candidates. Based on previous experiments (46), we have decided for a threshold score of 0.25 for filtering out potentially wrong gene/protein predictions. GNAT provides identifiers for all gene mentions with respect to the EntrezGene database (55). Cell lines are recognized based on the version 4.0 of Cellosaurus (ftp://ftp.nextprot.org/pub/current_release/controlled_vocabularies/ cellosaurus.txt), a manually curated vocabulary of cell lines provided by the Swiss Institute of Bioinformatics. Synonyms from Cellosaurus were automatically expanded according to space and hyphens, such as ‘BSF-1’, ‘BSF 1’ and ‘BSF1’, resulting in a list of >41 000 synonyms for 15 245 registered cell lines. Matching of the derived list of synonyms and the full texts is performed by Linnaeus (43). For the recognition of cell types and anatomical parts, we use Metamap (56), a system for Unified Medical Language System (UMLS) concept extraction. We configured Metamap to generate acronym variants and restricted results by the following semantic types: ‘Cell’ for cell types and ‘Anatomical Structure’, ‘Body Location or Region’, ‘Body Part, Organ or Organ Component’, ‘Body Space or Junction’, ‘Body Substance’, ‘Body System’, ‘Embryonic Structure’, ‘Fully Formed Anatomical Structure’ and ‘Tissue’ for anatomical parts. Metamap uses natural language processing techniques for breaking the text into phrases and further match them to UMLS concepts. From the potential matches returned by Metamap, we record not only the ones with highest score but also those that have the longest matching with the respective phrase. Cell types have also been extracted using an ontology-based approach in which synonyms from the CL are matched against the full texts. It consists on a list of 2786 cell types from 1491 terms and matching is again performed by Linnaeus (43). Finally, triggers are extracted based on a list of 509 expression triggers, which was built manually. Terms from the list are matched against the full text using Lingpipe (http://alias-i.com/lingpipe/). Metamap includes a step for acronym resolution, which returns a list of the pairs of abbreviations and long forms found as equivalent. However, Metamap sometimes recognizes the plural of some abbreviations but not the singular form or it does not return some abbreviations as a mention, but only the long forms. For instance, for cell types, Metamap recognizes ‘hESCs’ as an acronym for ‘human embryonic stem cells’, but not its singular form ‘hESC’. Further, although it lists the pair ‘hESCs’ and ‘human embryonic stem cells’ as being equivalent, only the long form is returned as a mention. Based on the list of pairs of abbreviations and long forms returned by Metamap, we try to match missed abbreviations and singular forms using Lingpipe. Metamap returns annotations with regard to Concept Unique Identifier (CUI) terms, the original UMLS identifiers. Whenever available, we map them to FMA and GO terms using mappings available at the UMLS database. CUI terms are also mapped to other ontologies and terminologies supported by UMLS, but not by CellFinder, such as the CRISP Thesaurus (http://www.nlm.nih.gov/research/umls/sourcereleasedocs/current/CSP/).To increase the recall of anatomical terms, we mapped UMLS CUI terms to CRISP terms [using mappings available at BioPortal (57)], and then further to other ontologies supported by CellFinder (e.g. CL, CLO, EHDAA2, MA, Uberon). Annotations returned by Metamap, which could not be automatically mapped to any supported ontology, are not removed, as identifiers could still be provided manually before integration of the data into the CellFinder database (not yet supported in the current curation workflow). Blacklists of manually curated mentions and identifiers are used for filtering out potential false predictions for all four entity types. This list was manually built based on the analysis of wrongly extracted annotations from the two corpora used for evaluation (cf. section ‘Results’). The list of mentions contains only one entry for cell line (‘FL’), 39 for anatomical parts (e.g. ‘organism’, ‘tissue’ and ‘analysis’), 31 entries for cell types (e.g. ‘cell’ and ‘stem cell’) and 79 entries for genes/proteins (e.g. ‘anti’, ‘repair’, ‘or in’). The list of identifiers include those which refer to broad concepts such as ‘cell’ (FMA:68646) or ‘tissue’ (FMA:9637). We filter out extracted mentions associated to any of the identifiers in this list. Results from sentence splitting, tokenization, part-of-speech tagging, parsing, dependency tags and named entities are integrated into the so-called ‘Interaction XML’ file format (https://github.com/jbjorne/TEES/wiki/TEES-Overview) (58) used by the Turku Event Extraction System (TEES) (59). TEES is an event extraction system, which uses multiclass Support Vector Machine on a rich graph-based feature set for trigger, edge and negation detection. Despite recent improvement of relation extraction methods (10), TEES seems to be the only available system suitable to be re-trained with novel corpora from any domain without the need of performing changes in its source code. We trained TEES in a gold-standard set of 20 full text annotated documents, 10 on human embryonic stem cell research (hereafter called CF-hESC), whose entities annotations have been previously published (46) and a new set of 10 full texts documents on kidney stem cell research (hereafter called CF-Kidney). Both corpora have been manually annotated with the five entity types (gene/proteins, cell lines, cell types, anatomical parts, expression triggers) and gene expression events (cf. example in Figure 2). These events are composed of a trigger, which is always linked to two arguments, a gene/protein (hereafter called ‘Gene’ argument) and a cell line, cell type or anatomical part (hereafter called ‘Cell’ argument). We split both corpora into three parts (training, development and test) and perform experiments using one corpus or a combination of both for training. Details on the corpora are shown in Table 1. Figure 2.Examples of gene expression events for the kidney stem cell corpus (PMID 17389645, PMCID PMC1885650). Each expression trigger (dark yellow) is always related with only one gene/protein (in blue) and only one cell (in yellow) or anatomical part (in red). However, the corpus was also annotated with entities, which do not take part in any event. Visualization of the corpus was provided by Brat annotation tool (60). Table 1.Statistics on the corporaFeaturesCF-hESCCF-KidneyTrainingDevelopmentTestTrainingDevelopmentTestDocuments622622Sentences13792595391578618383Sentences with entities9441633021344527314Sentences with events1472640240210122Entities41585831260483434431748 Genes/proteins126416335514401338782 Cell lines198721411181 Cell types155617952491725972 Anatomical parts92113717321161380617 Expression triggers2193267350458276Relationships944160390114414041320 Expression-Gene/protein47284195572702660 Expression-CellLine13636145 Expression-CellType4355612241139886 Expression-anatomy241837147299574Information is shown for the training, development and test data sets of the CF-hESC and CF-Kidney data sets. It includes number of documents, sentences, sentences with entities and sentences with events. Number of annotations is presented by entity type, and the number of events also shown according to the entities participating in the relationships. Examples of gene expression events for the kidney stem cell corpus (PMID 17389645, PMCID PMC1885650). Each expression trigger (dark yellow) is always related with only one gene/protein (in blue) and only one cell (in yellow) or anatomical part (in red). However, the corpus was also annotated with entities, which do not take part in any event. Visualization of the corpus was provided by Brat annotation tool (60). Statistics on the corpora Information is shown for the training, development and test data sets of the CF-hESC and CF-Kidney data sets. It includes number of documents, sentences, sentences with entities and sentences with events. Number of annotations is presented by entity type, and the number of events also shown according to the entities participating in the relationships. TEES receives the Interaction XML file as input and returns a new XML file, which includes predictions for the ‘Cell’ and ‘Gene’ relationships. The later are subsequently combined to compose complete gene expression events by checking the presence of both a ‘Gene’ and a ‘Cell’ relationship linked to the same trigger. TEES relationships are restricted to entities present in the same sentence; therefore, the same restriction is valid for all derived events. We applied TEES-trained models on the kidney cell data set of 2376 full texts. Results were manually validated using Bionotate (61), a collaborative open-source text annotation tool. Bionotate presents a snippet of text along with annotated entities, a question, and a list of possible answers. Curators were instructed to give one answer per snippet, and although Bionotate allows changing the span of the named entities, for this experiment, curators were asked only to answer the question. Bionotate selects snippets randomly among all those included in its repository. A snippet is no longer presented to the user when a certain number of agreements (equal answers) have been reached. For this experiment, one answer from any of our expert curators suffices. We have converted the output from TEES event extractor system to the XML format of the Bionotate. Snippets are composed of the sentence in which the event occurs and the two previous and subsequent sentences, for a better understanding of the context (cf. Figure 3). Additionally, a link to the respective PubMed entry is provided, in case those curators needed to check the abstract or full text of the publication before answering the question. The questions assessed whether there was a gene expression event taking place in the snippet, including its negation, whether the named entities were correctly recognized or if the publication was relevant for the kidney cell research. This resulted in the following possible answers: Yes, an event is taking place and all entities are correct. Yes, but the text says the gene expression is NOT taking place. No, no event is taking place although all entities are correct. No, this is not a gene expression trigger. No, this is not a gene. No, this is not a cell or anatomical part. No, both gene and cell or anatomical part are incorrect. No, the snippet (publication) does not seem to be relevant for CellFinder. Figure 3.Screen-shot of Bionotate configured for the validation of the gene expression events. Three named-entities are always pre-annotated: a trigger (in green), a gene (in blue) and a cell line, cell type or anatomical part (in red). The answers assess whether the biological event is taking place, its negation, the accuracy of the named-entity recognition and the relevancy of the publication from where the snippet was derived. Screen-shot of Bionotate configured for the validation of the gene expression events. Three named-entities are always pre-annotated: a trigger (in green), a gene (in blue) and a cell line, cell type or anatomical part (in red). The answers assess whether the biological event is taking place, its negation, the accuracy of the named-entity recognition and the relevancy of the publication from where the snippet was derived. In this section, we describe the evaluation performed for the methods used in the various stages of the text mining pipeline. We also present an overview of the data, which have been extracted by our curators with the help of the pipeline. The triage phase has not been directly evaluated, except for the answer number 8 during the manual validation of results (cf. ‘Manual validation’ in this section). Evaluation of the named-entity recognition and event extraction will be shown in terms of precision (P), recall (R) and f-score (F). Precision represents the ratio of the correct predictions of a particular system among all the returned ones. On the other hand, recall corresponds to the ratio of gold-standard annotations, which were actually returned by the system. Finally, the f-score is a harmonic average of both measures and shows the overall performance of a system. During the pre-processing step, sentence splitting in all 2376 full text documents resulted in a total of 581 350 sentences. Parsing and dependency tags conversion was successfully for 578 572 of them. The parsing information is only used by the TEES system (cf. ‘Event extraction’ in section ‘Methods and materials’), which means that although named-entity recognition was carried out in all sentences, only those correctly parsed ones were analyzed by TEES. Named-entity extraction was evaluated on the development and test gold-standard documents belonging to the human embryonic and kidney stem cell research (cf. Table 1), but only the development data sets were used for further improvements of methods, such as trigger list or blacklist construction and error analysis (cf. section ‘Discussion and future work’). Table 2 shows the evaluation of each entity type for both corpora. The ‘Exact’ evaluation assesses annotations, which matched regarding span and entity type, whereas ‘Overlap+Type’ allowed overlapping spans for annotations of the same type and ‘Overlap’ let annotations to have different types. The latter is particularly helpful regarding overlapping annotations between cell lines, cell types and anatomical parts, as any of these entity types corresponds to the same argument ‘Cell’ in the gene expression event (cf. Figure 2). Table 2.Evaluation of the automatic named-entity recognition on the CF- hESC and CF-Kidney corporaCorporaMatchEntity types (recall/F-score)GenesC. linesC. typesAnatomyExpressionCF-hESCDevelopmentEx.0.61/0.540.68/0.610.14/0.150.34/0.340.72/0.15OT0.75/0.650.94/0.850.62/0.660.48/0.450.91/0.19Ov.0.82/0.690.94/0.810.70/0.730.72/0.620.97/0.20TestEx.0.68/0.650.40/0.490.25/0.280.30/0.250.45/0.08OT0.76/0.720.58/0.650.58/0.650.43/0.350.54/0.09Ov.0.77/0.710.61/0.690.77/0.820.81/0.710.55/0.10CF-KidneyDevelopmentEx.0.34/0.451.00/0.330.17/0.260.69/0.750.68/0.43OT0.35/0.461.00/0.330.18/0.270.88/0.870.69/0.43Ov.0.46/0.561.00/0.340.77/0.800.90/0.890.76/0.47TestEx.0.69/0.761.00/0.330.89/0.860.67/0.740.80/0.42OT0.70/0.771.00/0.330.93/0.890.69/0.760.80/0.42Ov.0.70/0.771.00/0.330.94/0.910.72/0.770.81/0.42Results are shown for the development and test data sets in the format recall/F-score. Matching is evaluated regarding same span and entity type (Ex.), overlapping span and same type (OT) and overlapping span of any entity type (Ov.). Evaluation of the automatic named-entity recognition on the CF- hESC and CF-Kidney corpora Results are shown for the development and test data sets in the format recall/F-score. Matching is evaluated regarding same span and entity type (Ex.), overlapping span and same type (OT) and overlapping span of any entity type (Ov.). Recall is particularly low for genes/proteins in the development data set of the CF-Kidney corpus owing to a high number of annotations from a few genes/proteins, which have been missed by the system: ‘Gata3’ (155), ‘Ret’ (97) and ‘EpCAM’ (83). Some of these were found by GNAT but with a recall lower than the threshold we have considered. Cell lines are very rare in the CF-Kidney corpus, and the eight identical cell lines of the development data set and the only one of the test data set were correctly extracted (thus recall 1.0). Finally, recall is also particularly low for cell types in the development data set, even when allowing overlaps. Indeed, there is a great variety of cell types (>100), which could not be recognized, especially cell types, which in fact represent gene expressions events, such as ‘NCAM + NTRK2 + cells’ or ‘Gata3−/Ret− cells’. The ontology mapping post-processing step could automatically map a total of 171 (CF-hESC corpus) and 121 (CF-Kidney corpus) additional annotations to an identifier from any of the ontologies supported in CellFinder. They had been previously extracted by Metamap, but they were associated only to the UMLS CUI identifier. However, 1342 (33%) and 961 (16%) of the extracted annotations, respectively, remain assigned only to the UMLS CUI identifier, with respect to the total number of cell types and anatomical parts. The acronym resolution procedure has resulted in a slight increase in recall for cell types and anatomy, without loss of f-score (result not shown). For instance, recall for cell types in the CF-hESC corpus increased from 64 to 70% (result not shown) owing to the extraction of acronyms such as ‘MEF’ (mouse embryonic fibroblast) or ‘EB’ (embryoid body), which have not been previously returned by Metamap. Finally, blacklist filtering of terms also allowed a modest improvement of precision for both corpora (result not shown). For instance, precision for genes/proteins in the CF-hESC corpus increased from 43 to 50% (result not shown) owing to filtering out annotations such as ‘or in’ or ‘membrane’, which had been recognized by GNAT and genes or proteins. The named-entity extraction methods were run on the 2376 full texts and resulted in a total of >2 200 000 mentions for all five entity types. Details on the extracted annotations are presented in Table 3, such as the number of mentions for each entity type, distinct text spans and distinct identifiers. Table 3.Statistics on the extracted named entitiesAnnotationsGenesC. linesC. typesAnatomyExpressionDistinct mentions702 82981 074183 820565 860681 370Distinct spans34 2221825914214 874892Distinct ids34 35311 87511504300For each entity type, the number of annotations, distinct spans and identifiers is shown. Sometimes more than one identifier is assigned to a mention, therefore their high number. Trigger words (Expression) are not normalized to any ontology. Statistics on the extracted named entities For each entity type, the number of annotations, distinct spans and identifiers is shown. Sometimes more than one identifier is assigned to a mention, therefore their high number. Trigger words (Expression) are not normalized to any ontology. To extract gene expression events, we investigated training TEES on three models: CF-hESC corpus (6 full text documents), CF-Kidney corpus (6 full text documents) and a mix of both (12 full text documents) (hereafter called CF-Both). Input to TEES should include three data sets: training, development and test. During the training step, TEES automatically configures its parameters using the development data set and presents an evaluation of its own for the test set. Details on the performance of the relationship extraction is shown in Table 4 for the three training models, as well as for the complete events further performed by the authors. This is the performance of TEES without the influence of the named-entity recognition predictions of our text mining pipeline, as only gold-standard documents are used during the training step. Recall of the relationships range from 60 to 70% while precision is also good, from 60 to almost 90%. Both the recall and precision drop when considering the complete events, and recall is not always as high as the argument with the lower recall. This is due to the fact that TEES predicts the ‘Cell’ and ‘Gene’ relationships independently, and many of them are not associated to the same trigger. Table 4.Evaluation of TEES during trainingData setsRelationshipDevelopmentTestPRFPRFCF-hESCCell0.860.560.680.770.450.57Gene0.910.680.780.820.900.86Event0.600.350.440.380.530.44CF-KidneyCell0.710.500.590.620.680.65Gene0.600.820.690.730.750.74Event0.170.490.250.120.560.20CF-BothCell0.770.550.650.690.640.67Gene0.670.810.730.690.840.76Event0.550.480.510.500.560.53Evaluation is shown for the ‘Cell’ and ‘Gene’ relationships and for the development and test data sets, as described in Table 1. The complete events derived from a ‘Cell’ and a ‘Gene’ argument associated to the same trigger are also shown. For each training run, evaluation is carried out on the corresponding development and test data sets, i.e. two documents for each single corpus (CF-hESC and CF-Kidney) and four documents when training on the joined corpus (CF-Both). Predictions were performed over the gold-standard named-entity annotations. ‘P’ refers to ‘Precision’, ‘R’ to ‘Recall’ and ‘F’ to ‘F-score’. Evaluation of TEES during training Evaluation is shown for the ‘Cell’ and ‘Gene’ relationships and for the development and test data sets, as described in Table 1. The complete events derived from a ‘Cell’ and a ‘Gene’ argument associated to the same trigger are also shown. For each training run, evaluation is carried out on the corresponding development and test data sets, i.e. two documents for each single corpus (CF-hESC and CF-Kidney) and four documents when training on the joined corpus (CF-Both). Predictions were performed over the gold-standard named-entity annotations. ‘P’ refers to ‘Precision’, ‘R’ to ‘Recall’ and ‘F’ to ‘F-score’. In Table 5, we show the performance of TEES relationship extraction when using the predictions obtained in the named-entity recognition step, as well as gene expression events derived from the binary relationships. This is the final performance of our text mining pipeline for the extraction of gene expression events on cell and anatomical locations. Additionally, we include the performance for the prediction of the triplets gene-cell-trigger, which represent every possible combination of annotations from these three arguments in the same sentence. Therefore, it represents the higher possible recall for the event extraction provided the predicted named entities. Table 5.Evaluation of gene expression extractionData setsRelationship/EventDevelopmentTestPredictionsPRFPRFCF-hESCCell0.430.060.100.760.330.4614 551Gene0.350.220.270.760.790.77112 372Events0.500.080.140.270.050.084280Triplets0.060.510.100.050.350.09CF-KidneyCell0.440.020.050.520.570.55109 934Gene0.620.060.100.770.690.735520Event115Triplets0.020.190.040.020.280.05CF-BothCell1.00.010.020.700.640.6769 079Gene0.330.010.010.690.840.763792Event178Triplets0.020.220.040.030.300.05We have trained the TEES system on three data sets: CF-hESC, CF-Kidney and CF-Both. Results for the ‘Cell’ and ‘Gene’ relationships were provided by TEES during processing of the documents. Performance for complete events is evaluated allowing overlapping matches for entity spans, but with equality of entity types and argument types. The triplets correspond to every possible combination of the triggers, genes/proteins, cells or anatomical parts in the same sentence, i.e. the highest possible recall for any relationship extraction system provided the predictions for the entities. The ‘Pred.’ column presents the number of relationships or complete events, which have been extracted from the 2376 full texts on kidney research when using each of the training models. ‘P’ refers to ‘Precision’, ‘R’ to ‘Recall’ and ‘F’ to ‘F-score’. Evaluation of gene expression extraction We have trained the TEES system on three data sets: CF-hESC, CF-Kidney and CF-Both. Results for the ‘Cell’ and ‘Gene’ relationships were provided by TEES during processing of the documents. Performance for complete events is evaluated allowing overlapping matches for entity spans, but with equality of entity types and argument types. The triplets correspond to every possible combination of the triggers, genes/proteins, cells or anatomical parts in the same sentence, i.e. the highest possible recall for any relationship extraction system provided the predictions for the entities. The ‘Pred.’ column presents the number of relationships or complete events, which have been extracted from the 2376 full texts on kidney research when using each of the training models. ‘P’ refers to ‘Precision’, ‘R’ to ‘Recall’ and ‘F’ to ‘F-score’. Results are shown using the approximate span matching, i.e. for each argument, overlapping matches are allowed, but entities should have the same type as well as equality of the argument type (‘Cell’ or ‘Gene’). For the development data set and when using the CF-Kidney corpus for training TEES, whether alone or together with the CF-hESC corpus, no complete event was extracted. This is due to two reasons: (i) the low recall of genes/proteins and cell types for the CF-Kidney corpus (cf. Table 2, evaluation OT) and (ii) the inability of the CF-Kidney model to extract events from documents from other domains, i.e with different cell type nomenclature. Indeed, no gene expression events have been extracted from the two development documents of the CF-hESC corpus included in the development data set of the CF-Both corpus. This probably due to the high complexity and variability of the cell types in the CF-Kidney corpus, with examples such as ‘NCAM− cell’ or ‘EpCAM−NCAM−NTRK2+ cells’. We have run TEES using the three models (CF-hESC, CF-Kidney and CF- Both) on the 2376 documents and the named-entities previously extracted (cf. Table 3). We have obtained only 115 and 178 gene expression events for the CF-Kidney and CF-Both models, respectively, whereas the CF-hESC model retrieved 4280 events. The latter were derived from almost 127 000 binary relationships, i.e. the complete events correspond to only 14% of the original extracted relationships. The last column of Table 5 summarizes the number of relationships and derived events, which have been obtained using each training model. The gene expression events obtained with the three models were converted to the Bionotate XML format, and snippets were loaded into its repository. Curators have manually validated 2741 snippets, which contained events predicted by the three distinct models. Results are summarized in Table 6. The validated data, one file per snippet in the Bionotate’s XML format, is available for download at the CellFinder web site (http://cellfinder.org/about/annotation/). Table 6.Evaluation of the gene expression snippets in BionotateAnswersCF-hESCCF-KidneyCF-BothTotalNo. snippets%No. snippets%No. snippets%No. snippets%1. Yes120449.13429.563.3124445.42. Yes (negation)471.932.600501.83. No (but entities correct)2189.087.010.62278.34. No (trigger wrong)1948.02824.37843.830011.05. No (gene wrong)34614.1119.663.436313.26. No (cell/anatomy wrong)2078.52622.695.12428.87. No (gene/cell/anatomy wrong)552.243.510.6602.28. No (irrelevant document)1777.210.97743.22559.3Total24481001151001781002741100A total of 2741 snippets (gene expression events) were validated. These events were predicted by the three models used for training TEES event extraction system. Percentages for each answer are also shown. Evaluation of the gene expression snippets in Bionotate A total of 2741 snippets (gene expression events) were validated. These events were predicted by the three models used for training TEES event extraction system. Percentages for each answer are also shown. Validation for the events extracted using the CF-hESC model, the best performing one according to the evaluation and the number of predictions, can be summarized as follows. About 51% (answers 1 and 2) of the gene expression events have been extracted correctly, as well as the participating entities. This includes both positive and negative statements of gene expression in cell in anatomical parts. Exactly 17% (answers 3 and 4) of the snippets described processes not related to gene expression, although the gene, cell or anatomy were correctly recognized. Almost 25% (answers 5, 6 and 7) of the extracted events contained a wrong identified gene/protein, cell/anatomy or both of them, which means that precision was higher than the average for the named-entity recognition (cf. Table 2). Finally, 7.2% of the snippets turned out to belong to documents, which are irrelevant to the kidney cell domain, which gives a hint on the performance of the triage step. We have described our preliminary text mining pipeline for the extraction of five entity types and gene expression events. In this section, we discuss the most important results derived from this first experiment with our text mining curation pipeline. In the named-entity recognition step, we have considered only state-of-art and freely available tools, and we did not train specific systems with the gold-standard corpora discussed here. Results for entity extraction are in-line with previous published ones (46), although data sets are different and, therefore, results are not directly comparable. A high recall is preferable over a high precision, as events cannot be predicted if the participating entities have not been previously extracted. On the other hand, a high number of wrong predictions slow down the validation process, and therefore, a balance between precision and recall (given by the f-score) is also desirable. Provided the still low recall for some entities, and the consequent low recall of the event extraction, future work should still focus on the improvement of the named-entity prediction. Regarding genes/proteins extraction, most of the missing annotations could have been recognized by GNAT if we had used a lower threshold. Other tools could also be combined with GNAT, such as GeneTUKit (62) or BANNER (63). Additionally, use of domain-specific post-processing, such as ‘whitelists’ of genes/proteins, would certainly help, and future work will concentrate on these two approaches. Recall for genes/proteins increases considerably for both development data sets when allowing overlaps and an improvement is also perceived when type equality is relieved, which shows that some genes overlap with some cells names or anatomical parts, such as ‘C34’ (a gene) and ‘C34 cell’ (a cell type). Cell lines are not as common as cell types in our corpora, specially in the CF-Kidney corpus where this entity type is almost non-existent (cf Table 1). However, it plays an important role in the cell research, and scientific literature reports many gene expression events, which take place in cell cultures. Restricting our evaluation to the CF-hESC corpus, recall varies from 60 to >90% when allowing overlapping spans (cf. Table 2), but it is still not satisfactory, and dictionary-based methods might not be sufficient. Missing annotations for cell lines are mostly due to the absence of the synonym in any of the available thesaurus or ontologies, such as ‘SD56’, which is not included in Cellosaurus. Thus, future work will include training a machine learning system for cell line recognition, including annotation of additional gold-standard documents. Improvement of the event extraction starts with the improvement of the recall for the named entities. Performance of cell types and anatomical parts are rather variable. A good recall is usually obtained when releasing equality of types, and further experiments should consider unifying the cell types and anatomical parts in our corpora. If fact, previous studies on the CF-hESC corpus show that inter-annotator agreement for these entity types was low (46). Overlaps between cell types and anatomical parts should not be a problem for the gene expression event extraction, as both entity types takes part in the ‘Cell’ argument. Cell types were sometimes poorly recognized for the CF-Kidney data set, owing to the high variability of the nomenclature and the presence of gene expression in its contents, such as ‘NCAM+NTRK2+ cells’ or ‘Gata3−/Ret− cells’. Thus, improvements on cell type extraction should also focus on training machine learning algorithms. Mapping cell types with such a pattern to an identifier is also a challenge, as these terms are not included in any available ontology. The prior identification of the original cell type in which the gene is being expressed can help in the normalization of these cells, an information that is usually present in the text, although not always in the same sentence. Expression triggers are extracted based on a manually curated list, which assures a high recall. Low recall, such as the ones for the development data set of the CF-Kidney corpus are due to unusual trigger words, such as ‘-’ (negative expression), ‘dim’ and ‘bright’. We obtained the gene expression events using the TEES edge detection module, which extracted relationships between expression triggers and a gene/protein, cell or anatomy. TEES allows training the system with novel corpora, and during the training step, examples are generated for all combinations of entities provided in the training corpus. Therefore, a few relationships returned by TEES are related to the wrong entity type. For instance, it extracts some ‘Gene’ arguments associated to cells or anatomical parts and some ‘Cell’ arguments related to genes, although no such examples can be found in any of our gold-standard corpora. TEES extracts the relationships independently. Therefore, the recall of the binary relationships does not correspond to the recall of the complete gene expression event. Future work on event extraction will also include trying additional event extraction systems, such as (64, 65). Use of more annotated documents might also improve the event extraction. Further experiments can also be performed using available corpora, such as the set of annotated abstracts of the Gene Expression Text Miner corpus (40). Additionally, a careful analysis of the wrongly extracted events returned by TEES when using gold-standard annotations (cf. low precision for CF-Kidney corpus in Table 4) could reveal inconsistencies in the manual annotations in our corpora. To avoid huge differences between development and test results, a cross-validation could have been investigated. In summary, a cross-validation in a larger and more robust corpus could provide more stable results. Nevertheless, these preliminaries results on extraction of gene expression in cells and anatomical parts are certainly interesting for the many groups working on event extraction, as this is one of the first curation experiment to use a event extraction system, which had not been developed by the authors. Additionally, it is probably the first external evaluation of TEES on a new corpus, one of the very few event extraction systems available to the public. Finally, the use of corpora from two distinct cell research domains shows how large differences in results are dependent on the corpus and the corresponding learned model. Processing of the data set of 2376 full text documents for kidney cell research resulted in a high number of entities but apparently a low number of extracted events. However, recall is unknown, as well as the number of publications, which described expression of genes in cells and anatomical parts for the kidney cell research. The number of correct gene expression events is certainly low compared with the number of processed documents, but number of irrelevant publications in our collection is also unknown and could be higher than 6%, as reported by answer number 8 of the validation (cf. Table 6). Next event extraction tasks will involve recognition of additional relationships, such as identifying the cell type or tissue from which a certain cell line was derived. Future work will also include additional biological processes, such as cell differentiation. These relationships have already been annotated in the two gold-standard corpora discussed here and involve the same entities whose recognition is already included in our pipeline. Manual validation of 2741 snippets reported that half of them contained correctly recognized entities and gene expression events, which is in line with the precision of TEES shown in Table 5. Curators reported that most mistakes concentrated on incomplete extraction of genes/proteins and cell types, such as the recognition of ‘TGF’ instead of ‘TGF-beta’. Feedback from the validation will help to improve both recall and precision for the named-entity recognition by adding more terms to the blacklists (potential wrong predictions) and by creating ‘whitelists’ (potential missing annotations). Curators reported a positive first experience with Bionotate, although changes in visual interface, short-cuts and functional features have been suggested as future work. Next experiments will also focus on the validation of the identifiers, which were automatically assigned during the named-entity recognition, as well as allowing curators to change the span of the pre-annotated entities, a feature already supported by Bionotate. Validation of the normalized identifiers is an important step before final integration of the results into the CellFinder database. Version 2.0 of Bionotate (66) supports this functionality and will certainly be considered for integration in our pipeline. We presented here our preliminary results for the text mining pipeline for curation of gene expression events in cells in anatomical parts for the CellFinder database. Our pipeline relies only on open-source or freely available tools, and evaluation for each stage has been carried out based on gold-standard corpora. We are not aware of previous database curation pipelines where text mining methods have been used in all of the following stages: triage, named-entity recognition and event extraction. We performed named-entity extraction extraction for genes/proteins, cell lines, cell types, tissues, organs and gene expression triggers. Gene expression events were extracted using machine learning algorithms trained on manually annotated corpora from two domains, human embryonic stem cells and kidney cell research. Results for both the name-entity recognition and event extraction steps are promising, although improvements are still necessary to achieve a higher recall and precision. The text mining pipeline has been used to process 2376 full texts documents on kidney cell research and resulted in a total of >60 000 distinct entities and >4500 gene expression events. Half of the events have been manually validated by experts, and about half of them were classified as describing a gene expression taking place in a cell or anatomical part.
|
PMC11334037
|
Insulin-like growth factor 2 targets IGF1R signaling transduction to facilitate metastasis and imatinib resistance in gastrointestinal stromal tumors
|
Gastrointestinal stromal tumors (GISTs) are typical gastrointestinal tract neoplasms. Imatinib is the first-line therapy for GIST patients. Drug resistance limits the long-term effectiveness of imatinib. The regulatory effect of insulin-like growth factor 2 (IGF2) has been confirmed in various cancers and is related to resistance to chemotherapy and a worse prognosis. To further investigate the mechanism of IGF2 specific to GISTs. IGF2 was screened and analyzed using Gene Expression Omnibus (GEO: GSE225819) data. After IGF2 knockdown or overexpression by transfection, the phenotypes (proliferation, migration, invasion, apoptosis) of GIST cells were characterized by cell counting kit 8, Transwell, and flow cytometry assays. We used western blotting to evaluate pathway-associated and epithelial-mesenchymal transition (EMT)-associated proteins. We injected transfected cells into nude mice to establish a tumor xenograft model and observed the occurrence and metastasis of GIST. Data from the GEO indicated that IGF2 expression is high in GISTs, associated with liver metastasis, and closely related to drug resistance. GIST cells with high expression of IGF2 had increased proliferation and migration, invasiveness and EMT. Knockdown of IGF2 significantly inhibited those activities. In addition, OE-IGF2 promoted GIST metastasis in vivo in nude mice. IGF2 activated IGF1R signaling in GIST cells, and IGF2/IGF1R-mediated glycolysis was required for GIST with liver metastasis. GIST cells with IGF2 knockdown were sensitive to imatinib treatment when IGF2 overexpression significantly raised imatinib resistance. Moreover, 2-deoxy-D-glucose (a glycolysis inhibitor) treatment reversed IGF2 overexpression-mediated imatinib resistance in GISTs. IGF2 targeting of IGF1R signaling inhibited metastasis and decreased imatinib resistance by driving glycolysis in GISTs.Core Tip: Our study found that insulin-like growth factor 2 (IGF2) regulated metastasis and imatinib resistance in gastrointestinal stromal tumors (GISTs). IGF2 interacted with IGF1R to regulate glycolysis. Our results confirm that IGF2 targeting of IGF1R signaling inhibited metastasis and improved imatinib chemosensitivity by driving glycolysis in GISTs and indicated that IGF2 might be used to reverse imatinib resistance in GIST patients. Primary gastrointestinal stromal tumors (GISTs) account for 2% of gastrointestinal tumors. GISTs are encoded by the receptor tyrosine kinase gene KIT or PDGFRA. These mutations cause ligand-dependent activation and constitutive activation of signal transduction mediated by PDGFRA or KIT. The downstream molecular pathways of the KIT mutation include PI3K/AKT, JAK-STAT, Src family kinases, and Ras-ERK). Activation of molecular pathways follows KIT activation and leads to the occurrence of GISTs tumors by activation of cell proliferation and inhibition of apoptosis signals . Imatinib remains the primary treatment of GIST patients with advanced or metastatic tumors. Imatinib significantly improves the prognosis of patients in the advanced stages of the disease, but those undergoing imatinib treatment often encounter challenges associated with both primary and secondary drug resistance, which, unfortunately, restricts long-term efficacy. Insulin-like growth factor 2 (IGF2) is a genomic imprinting gene in growth on the chromosome 11 short arm. IGF2 overexpression is observed in a variety. of cancers and is related to chemotherapy resistance and a worse prognosis[12-14]. Studies of IGF1R have increased recently. Insulin-like growth factor (IGF) is comprised of the two ligands IGF1 and IGF2, their target tyrosine kinase receptors, IGF1 receptor (IGF1R) and the insulin receptor, as well as the IGF2 receptor (IGF2R) and IGF-binding proteins that regulate IGF ligand availability. IGF1R, is a tyrosine kinase receptor with binding affinity for both IGF1 and IGF2 ligands. Upon ligand binding, the activated tyrosine kinase domain initiates signaling cascades that specifically activate the GPTase Ras-Raf-ERK/MAPK and PI3K-AKT/mTOR pathways. These pathways, regulate the proliferation rate and apoptosis of cancer cells. The IGF pathway family gene expression (such as IGF1, IGF2, and IGF1R) has been reported to distinguish subsets of GISTs wild type for KIT and PDGFRA. Although data on IGF1R in GISTs have been reported[20-22], further research on the mechanisms of IGF2 and IGF1R in GISTs is needed. Sequencing data from the Gene Expression Omnibus (GEO) database (GSE225819 and GSE155880) were examined by bioinformatics. We found that IGF2 acted as a cancer-promoting factor and was involved in cell proliferation, apoptosis, liver metastasis, and epithelial-mesenchymal transition (EMT) in GISTs. Moreover, the role of IGF2 in GIST cells and the IGF2-IGF1R regulatory axis contributed to imatinib resistance of GISTs by regulating glycolysis and represents a target for GISTs therapy. Gene expression data based on RNA sequencing were obtained from the GEO. Two eligible datasets (GSE225819, GSE155880) were combined. The aligned reads were calculated by FeatureCounts (subread/2.0, http://subread.sourceforge.net/) and differentially expressed genes (DEGs) were analyzed by the R package DESeq2/3.1.0 (https://bioconductor.org/packages/release/bioc/html/DESeq2.html). A total of 2578 DEGs (1398 downregulated, and 1188 upregulated) were identified by screening GSE225819, including 20 normal samples and 20 GISTs samples with liver metastasis (|log2FC| > 1; P < 0.05) (Supplementary Table 1). Based on Deseq2, 1386 DEGs (939 downregulated, and 447 upregulated) were identified by screened GSE155880 including seven Imatinib-sensitive samples and seven imatinib-resistant GIST patients (|log2FC| > 1; P < 0.05) (Supplementary Table 2). RGM-1 normal human gastric mucosal cells, GIST882, and GIST-T1 cells were cultured in Iscove's modified Dulbecco's medium containing10% fetal bovine serum and 1% antibiotics, The culture temperature was 37 °C with 5% CO2. The imatinib concentration was increased from 1 nM to 100 nM over 10 mon and repeated to obtain imatinib-resistant GIST882 (GIST882-R) and GISTT1 (GISTT1R) cells. GIST882 and GIST-T1 cells were transfected with OE-IGF2, sh-IGF2 plasmids and sh-NC, OE-NC negative controls (RiboBio, Beijing, China) using Lipofectamine 3000 (Invitrogen, Waltham, MA, United States) and cultured for 2 d. Transfection efficiency was determined by western blotting. Imatinib mesylate was purchased from Selleckchem (Houston, TX, United States). GIST-T1 and GIST-882 cells were treated with serial dilutions of 1 μM imatinib in dimethyl sulfoxide for 4 h. We lysed transfected cells with RIPA buffer, the total protein was purified, and the protein concentration was determined with bicinchoninic kits (ThermoFisher Scientific, Waltham, MA, United States). The proteins were resolved by 10% SDS-PAGE and transferred to PVDF membranes for incubation with anti-IGF2 (1:1000, ab177467; Abcam, Cambridge, United Kingdom), anti-vimentin (1:1000, ab92547; Abcam), anti-N-cadherin (1:1000, ab76011; Abcam), anti-E-cadherin (1:1000, ab40772; Abcam), anti-Twist1 (1:1000, ab50887; Abcam), anti-IGF1R (1:1000, ab182408; Abcam), anti-p-IGF1R (1:1000, ab39398; Abcam), anti-PI3K (1:1000, ab302958; Abcam), anti-AKT (1:1000, MA5-14916; Invitrogen), anti-phospho-AKT (1:1000, PA5-95669; Invitrogen), and anti-β-actin (1:1000, ab8227; Abcam) primary antibodies overnight at 4 °C after blocking with skimmed milk (5%). After washing the primary antibodies away, the proteins were incubated with the anti-rabbit secondary antibody (1:5000; SA00001-2; SanYing Biotechnology Inc, Wuhan, China) for 1 h. The protein bands were visualized using an ECL chemiluminescence system, and the protein blots were quantified with Image J. The concentration of IGF2 was measured using ELISA kits (Abcam) according to the manufacturer′s instructions. The samples were prepared from cell culture supernatants and the IGF2 concentration was measured at 450 nm using a microplate reader. We determined GIST cell proliferation by cell counting kit-8 (CCK-8) assay. OE-IGF2- or sh-IGF2-transfected GIST882 and GIST-T1 cells were added to 96-well plates (1 × 10/well). After 1 d, we added CCK-8 reagent (10 μL, Catalog No. AD10; Dojindo Molecular Technologies, Kumamoto, Japan) to each well at room temperature. Absorbance was monitored at 0, 24, 48, 72, and 96 h and the half inhibitory concentration of imatinib was determined at 450 nm. After overnight incubation, the cells were treated with imatinib at 0, 20, 40, 60, and 80 μmol/L for 48 h. CompuSyn software was used to calculate the combination index using the Chou-Talalay method to determine the antagonistic influence. For the migration assay, GIST cells were seeded into 8 µm well Transwell chambers (Corning; Corning, NY, United States). The upper chamber was filled with 200 µL serum-free medium containing 2 × 10 cells and the lower chamber was filled with 500 μL complete medium (10% FBS). After 48 h, the cells were fixed with formaldehyde and stained with 0.2% crystal violet for 10 min. To assay cell invasion, 500 μL culture supernatant was collected from transfected cells and added to the upper Transwell chamber. GIST cells (2 × 10 cells) in about 200 μL serum-free medium were added to the lower chamber. The cells were cultured for 2 d at 37 °C with 5% CO2. After culturing, cells remaining in the lower chamber were removed with cotton swabs and those in the upper chamber were stained with 0.2% crystal violet for 5 min. We used an inverted microscope to count the cells that had migrated through the membrane and invaded the upper chamber. We bought 5-wk-old; male BALB/c nude mice from Vital River Laboratories (Beijing, China) and housed them for 1 wk to adapt to the environment. GIST-T1 cells (5 × 10) transfected with OE-IGF2/OE-NC, sh-IGF2/sh-NC were injected into the inguinal skin and the mice were monitored for growth of the tumor for 7 d before being randomized to four groups and treated with imatinib 50 mg/kg daily. After 4 wk, we killed the mice with an overdose of pentobarbital. All animal experiments were approved by the Animal Ethics Committee of Beijing Viewsolid Biotechnology Co. LTD (Protocol No. VS2126A00170) and all methods followed the ARRIVE guidelines. We fixed the liver tissue of mice in neutral formalin (10%), embedded it in paraffin, cut the tissue into 4 µm sections, and stained it with hematoxylin and eosin (HE). The sections were observed with a microscope. Cells were incubated in commercial seahorse XF assay medium plus pyruvate (1 mmol/L), glucose (10 mmol/L) and glutamine (2 mmol/L) 37 °C for 1 h in a CO2-free incubator. The rate of extracellular acidification was measured before and after addition of oligomycin, glucose, and 2-deoxy-D-glucose (2-DG). FCCP, a mitochondrial uncoupling agent; oligomycin, an ATP synthase inhibitor; 2-DG, a glycolysis inhibitor; rotenone; and antimycin A were added and metabolic energy consumption was assayed with a Seahorse XF96 Analyzer (Agilent, Santa Clara, CA, United States). The concentration of lactate in transfected cells was determined by ELISA with lactate assay kits (MAK064; Sigma-Aldrich, St Louis, MO, United States) according to the manufacturer’s protocol. The optical density of each well was determined at 570 nm (Plate Reader AF2000; Eppendorf, Waltham, MA, United States). GIST cell apoptosis was assayed by flow cytometry (LSRII; BD Biosciences, Franklin Lakes, NJ, United States). using annexin V-FITC apoptosis detection kits. The apoptosis rate was determined by analysis of Q2 and Q3 quadrant cells. We used GraphPad Prism 7.0 for data analysis. Data were reported as mean ± standard deviation of three independent experiments. Single-group comparisons were done with Student’s t-tests. Multiple group differences were compared by analysis of variance. P < 0.05 indicated significance. Based on the limma R package, a total of 2578 (DEGs 1398 downregulated and 1188 upregulated) were screened out from GEO: GSE225819 data, including 20 normal samples and 20 GIST samples with liver metastasis (|log2FC| > 1; P < 0.05), suggesting that these DEGs may be involved in liver metastasis in GIST patients (Figure 1A). The top 10 upregulated genes were PENK, IGF2, GPR20, CTSL, SCRG1, PNMAL1, NKX3-2, ANO1, PLAT, and BCHE. The top 10 downregulated genes were ATP4B, GKN1, MT1G, GKN2, ATP4A, SPINK1, TSPAN8, TFF1, KCNE2, and REG1A (Supplementary Table 1). Based on the Deseq2, 1386 DEGs (939 downregulated and 447 upregulated) were screened out in GSE155880, including seven Imatinib-sensitive samples and seven imatinib-resistant GIST patients (|log2FC| > 1; P < 0.05, Figure 1B). The intersection of the two analyses indicated that only IGF2 was involved in the drug resistance regulation and GIST metastasis in these DEGs (Supplementary Table 2). Moreover, we evaluated IGF2 expression in the GIST cell line. By western blotting, expression levels of IGF2 in GIST882, GIST882-R, GIST-T1, and GIST-T1-R were higher than those in normal RGM-1. Furthermore, IGF2 was significantly over expressed in GIST882-R/GIST-T1-R compared with other cell lines GIST882/GIST-T1 (P < 0.01, P < 0.001; Figure 1C). In addition, the expression levels of IGF2 in culture supernatants were measured using ELISA and compared (Figure 1D). We found that the ELISA and western blot results (P < 0.05, P < 0.001) were similar. IGF2 expression was high in drug-resistant GIST cell lines, suggesting that IGF2 overexpression may be closely related to drug resistance. High expression of insulin-like growth factor 2 in gastrointestinal stromal tumors with liver metastasis and closely related to drug resistance. A: Differentially expressed genes in gastrointestinal stromal tumors (GIST) with liver metastasis tissues and normal gastric tissues (|log2FC| > 1; P < 0.05); B: Differentially expressed genes in imatinib sensitive and in seven Imatinib-resistant GIST patients (|log2FC| > 1; P < 0.05); C: Western blot assay of insulin-like growth factor 2 (IGF2) protein expression in GIST cell lines (GIST882, GIST882-R, GIST-T1, GIST-T1-R); D: ELISA of IGF2 expression in GIST cell lines (GIST882, GIST882-R, GIST-T1, GIST-T1-R). Data are mean ± standard deviation. P < 0.05; P < 0.01; P < 0.001. We transfected GIST882 and GIST-T1 cells with an IGF2 overexpressing plasmid (OE-IGF2) or a shRNA to knock down IGF2 (sh-IGF2). Western blotting detected the efficiency of cell transfection (Figure 2A). IGF2 was highly expressed in OE-IGF2-transfected cells compared with OE-NC cells, while IGF2 expression was low in sh-IGF2-transfected cells (P < 0.001). ELISA also found that IGF2 expression high in OE-IGF2 group compared with OE-NC-GIST882 and GIST-T1 cells and IGF2 was low expressed in sh-IGF2-transfected cells (Figure 2B, P < 0.05, P < 0.01, P < 0.001). The CCK-8 results showed that cell viability was significantly increased after exogenous expression of IGF2, sh-IGF2 transfection inhibited GIST882 and GIST-T1 cell viability (Figure 2C, P < 0.001). Likewise, the Transwell assays found more migrating and invading OE-IGF2-GIST882 and GIST-T1 cells compared with their respective control cells (Figure 2D and E, P < 0.001). We also found that sh-IGF2 transfection inhibited cell viability, migration and invasion. In addition, western blotting detect EMT-related proteins (E-cadherin, vimentin, Twist1, and N-cadherin) expression in cells. Silencing IGF2 increased E-cadherin expression, and inhibited vimentin, Twist1, and N-cadherin expression, but IGF2 overexpression had the opposite experimental findings (Figure 2F, P < 0.001). To further verify the functional role of IGF2 on the growth of GISTs, we performed nude mouse tumorigenesis experiments. OE-IGF2 transfected-GIST-T1 cell lines were injected into the spleen. We found that OE-IGF2 promoted the GIST-T1 cell metastasis in vivo, showing a significant decline in the number of liver metastatic nodules (Figure 2G and H, P < 0.01). Insulin-like growth factor 2 promotes malignant characteristics and metastasis of gastrointestinal stromal tumors. A: Western blot measured the transfection efficiency of OE-insulin-like growth factor 2 (IGF2) or sh-IGF2 in gastrointestinal stromal tumors (GIST) 882 and GIST-T1 cells; B: ELISA of IGF2 expression in OE-IGF2 or sh-IGF2 transfected GIST882 and GIST-T1 cells; C: Cell counting kit-8 assay assessed cell viability in GIST882 and GIST-T1 cells; D: Transwell assay evaluated the migration of OE-IGF2- or sh-IGF2-transfected cells (scar bar = 50 μm); E: Transwell assays of the invasiveness of OE-IGF2 or sh-IGF2 transfected cells. (scar bar = 50 μm); F: Detection of proteins involved in epithelial-mesenchymal transition (vimentin, N-cadherin, E-cadherin, Twist1) in OE-IGF2 or sh-IGF2 transfected cells; G: Liver tissue from tumor xenografts in nude mice injected withOE-IGF2 transfected GIST-T1 cells; H: Liver metastasis determined by hematoxylin-eosin staining. Data are mean ± standard deviation. P < 0.05; P < 0.01; P < 0.001. IGF1R mRNA expression was increased in GIST-T1 and GIST882 cells transfected with OE-IGF2, and IGF1R mRNA expression was decreased after sh-IGF2 transfection (Figure 3A, P < 0.001). PI3K-Akt signaling is the IGF2-IGF1R signal principal downstream target. Expression of IGF2-IGF1R pathway-associated proteins (IGF1R, p-IGF1R, PI3K, AKT, p-AKT) in GIST-T1 cells was measured by western blotting. IGF2 overexpression increased the expression of IGF1R, p-IGF1R, PI3K, AKT, and p-AKT in GIST-T1 cells. The opposite result was noted after IGF2 knockdown (Figure 3B, P < 0.01, P < 0.001). Although sh-IGF2 reduced IGF1R, p-IGF1R, PI3K, AKT, and p-AKT expression in GIST-T1 cells, it was partially restored by overexpression of IGF2R (Figure 3C, P < 0.01, P < 0.001). Insulin-like growth factor 2 activated the IGF1R signaling in gastrointestinal stromal tumors cells. A: Quantitative reverse transcriptase PCR assay of IGF1R mRNA expression in gastrointestinal stromal tumors (GIST) 882 and GIST-T1 cells after OE-insulin-like growth factor 2 (IGF2) or sh-IGF2 transfection; B: Detection of protein levels (IGF1R, p-IGF1R, PI3K, AKT, and p-AKT) involved in the PI3K/AKT in OE-IGF2 or sh-IGF2 transfected-GIST-T1 cells by western blot assay; C: Detection of protein levels (IGF1R, p-IGF1R, PI3K, AKT, and p-AKT) involved in the PI3K/AKT in GIST-T1 cells after sh-IGF2 and OE-IGF2R transfection by western blot assay. P < 0.01; P < 0.001. We analyzed glucose consumption and lactate production in GIST cells. Sh-IGF2 inhibited glucose consumption (Figure 4A), and lactate production in GIST882 and GIST-T1 cells (Figure 4B), but IGF2 overexpression had the opposite experimental findings (P < 0.001). To examine the role of the Warburg effect in liver metastasis of GISTs, we treated OE-NC-GIST882 and OE-IGF2-GIST882 cells with 2-deoxyglucose (2-DG, a glycolysis inhibitor) for 24 hat 0, 4, 8, and 16 mmol/L. 2-DG significantly inhibited glycolysis (Figure 4C, P < 0.05, P < 0.01, P < 0.001) and Transwell assays found that 2-DG treatment inhibited the promoting effect of OE-IGF2 on GIST882 and GIST-T1 cell invasion and migration (Figure 4D and E, P < 0.001). Similarly, OE-IGF2 increased vimentin, Twist1, and N-cadherin expression and inhibited E-cadherin expression in cells, but the expression was partially restored by 2-DG treatment (Figure 4F, P < 0.001). Insulin-like growth factor 2/IGF1R-mediatedglycolysisis required for gastrointestinal stromal tumors with liver metastasis. A: Extracellular acidification rate was measured; B: Lactate production in gastrointestinal stromal tumors (GIST) 882 and GIST-T1 cells transfected with sh-insulin-like growth factor 2 (IGF2) or OE-IGF2 were measured; C: Lactate production in OE-IGF2-GIST882 and GIST-T1 cells cotreated with 2-deoxy-D-glucose (2-DG) (0, 4, 8, and 16 mmol/L); D: Transwell assay of the migration ability of the OE-IGF2-cells cotreated with 2-DG (scar bar = 50 μm); E: Transwell assay of the invasiveness of OE-IGF2-cells cotreated with 2-DG (scar bar = 50 μm); F: Assay of proteins involved in epithelial-mesenchymal transition (vimentin, N-cadherin, E-cadherin, Twist1) in OE-IGF2-cells cotreated with 2-DG. Data are mean ± standard deviation. P < 0.05; P < 0.01; P < 0.001. Figure 1 shows that IGF2 was involved in regulating drug resistance. Next, we will further verify. To test whether IGF2 also regulated drug resistance in GISTs in vivo, we established a xenograft model by inoculating sh-NC or sh-IGF2-GIST-T1 cells into nude mice. In the sh-IGF2-GIST-T1 mouse xenograft model, tumor volume and growth were inhibited by sh-IGF2, and imatinib had the same influence on tumor growth and volume. Combined treatment with imatinib and sh-IGF2 was more effective for reducing tumor progression than single treatment (Figure 5A-C, P < 0.001). The western blot results revealed that expression of IGF1R, p-IGF1R, AKT, PI3K, and p-AKT in tumor tissue was suppressed in both sh-IGF2-transfected cells and after imatinib treatment. Moreover, combined imatinib and sh-IGF2 were more effective than single therapy (Figure 5D, P < 0.001). The above data suggest that IGF2/IGF1R regulate imatinib resistance. Insulin-like growth factor 2/IGF1R regulates imatinib resistance of gastrointestinal stromal tumors by regulating glycolysis. A: Tumor growth in xenografted nude mice; B: Tumor volumes in sh-insulin-like growth factor 2 (IGF2)-gastrointestinal stromal tumors (GIST)-T1 mouse xenograft models treated with imatinib; C: After 35 d, the mice were killed and the tumors were weighed; D: Assay of IGF1R, p-IGF1R, PI3K, AKT, and p-AKT in tumor tissue by western blotting; E: Assay of drug sensitivity in OE-IGF2-GIST882 and GIST-T1 cells treated with 2-deoxy-D-glucose (2-DG); F: Flow cytometry assay of apoptosis of OE-IGF2-GIST882 and GIST-T1 cells treated with 2-DG. Data are mean ± standard deviation. P < 0.05; P < 0.01; P < 0.001. In addition, previous data shows that IGF2 regulates glycolysis in GIST cells. IGF2 regulates cell sensitivity to imatinib through its influence on glycolysis. We used 2-DG to inhibit glycolysis in GIST cells. OE-IGF2 increased drug sensitivity in GIST882 and GIST-T1 cells, but after treatment with 2-DG, transfection with OE-IGF2 no longer changed drug sensitivity in GIST cells (Figure 5E, P < 0.001). Flow cytometric analysis showed that sh-IGF2 suppressed imatinib-induced apoptosis and OE-IGF2 reduced imatinib-induced apoptosis in GIST cells. Treatment with 2-DG and transfection with OE-IGF2 no longer influenced imatinib-induced apoptosis in GIST cells (Figure 5F, P < 0.001). Therefore, the results show that IGF2 regulated imatinib sensitivity in GIST cells by affecting glycolysis. GISTs is the most frequent malignant gastrointestinal sarcoma and causes significant patient harm. Recently, anticancer drug resistance has become a significant challenge to the treatment of GISTs. Treatment with tyrosine kinase inhibitors (TKIs) has led to substantial improvement of survival, both for patients with localized GISTs and those with advanced disease. As the first-line TKI, imatinib offers treatment for advanced and metastatic GISTs, adjuvant therapy in high-risk GISTs and neoadjuvant treatment to downsize large tumors prior to resection. We explored the mechanism of IGF2 in imatinib resistance in GISTs and whether IGF2 enhanced metastasis and imatinib resistance by driving glycolysis by targeting IGF1R signaling transduction. IGF2, identified as an imprinted gene, exhibits a significant impact on cancer progression when its imprinting is lost or relaxed, leading to heightened autocrine IGF2 levels and increased secretion in malignant cells. Numerous studies have revealed the upregulation of IGF2 in various cancers such as hepatocellular carcinoma, correlating with resistance to chemotherapy and a poorer prognosis[12-14]. Our investigation, which focused on DEGs associated with liver metastasis and drug resistance in GISTs, we observed elevated levels of IGF2 in GISTs cases linked to liver metastasis and drug resistance. Our comprehensive analysis included assessment of cell proliferation, viability, migration, and invasiveness. The findings strongly suggest that overexpression of IGF2 induce the proliferation, metastasis, and EMT of GIST cells. IGF1R, is a tyrosine kinase receptor that can be triggered by IGF2 and has a pivotal role in regulating mammalian development, metabolism, and growth. IGF1R is known to be upregulated in various human solid tumors. Its involvement in cell promoting cell proliferation and inhibiting programmed cell death is facilitated by activation of its tyrosine kinase and the subsequent engagement of the Ras/Raf/MEK and PI3K/AKT/mTOR signaling pathways. The IGF2-IGF1R signaling axis assumes critical significance in governing cell proliferation, differentiation, EMT, migration, drug resistance, and maintaining stemness in malignancies. This investigation further demonstrated the activation of IGF1R signaling by IGF2 in GIST cells. It highlights the role of IGF2 as a pivotal chromatin factor that controls the expression level of IGF1R and modulates downstream signaling by the PI3K/AKT pathway. IGF2 also upregulated the expression of glycolytic and mitochondrial respiration markers. IGF2 overexpression has also been shown to cause metabolic reprogramming in breast cancer. As expected, we also that IGF2 mediated the glycolysis in GISTs by targeting IGF1R signaling. Increased expression of IGF2 is a common occurrence in various cancers and has been associated with increased resistance to chemotherapy, leading to a poorer prognosis. Regarding GISTs, the standard first-line therapeutic approach involves the use of imatinib. Imatinib, a potent TKI, is the primary treatment for GISTs, and significantly contributes to the progression-free survival of GIST patients. Our investigation revealed a noteworthy correlation of increased IGF2 expression with the induction of GISTs resistance to imatinib concurrently with a reduction of imatinib-induced apoptosis in GIST cells. These findings underscore IGF2 as a potential regulator of GISTs imatinib resistance, and a promising target for interventions aimed at reversing such resistance. Intriguingly, our study further showed that IGF2 regulates cellular sensitivity to imatinib by modulating glycolysis. The study had some limitations of this study. First, except for GIST cells, the role of IGF2 on GIST patient samples needs verification. Even though we found that IGF-2 overexpression increased the resistance of GIST cells to imatinib in cell culture, the clinical effect needs to be verified. Secondly, our results allows speculation that IGF2 was involved in the resistance to chemotherapy and a worse GISTs prognosis. However, the molecular mechanism of IGF2 specific to GISTs requires further investigation. We will consider these issues in future studies. In addition, studies have found that hypoglycemia in patients with non-islet cell tumor-induced GISTs may be aggravated by imatinib. A recent case study reported that a GISTs that produced big-IGF2 also caused severe hypoglycemia. We also hope to investigate that in future experiments. This study investigated IGF2 regulation of metastasis and imatinib resistance in GISTs. IGF2 interacted with IGF1R to regulate glycolysis. Our results found that IGF2 targeting of IGF1R signaling improved metastasis and imatinib chemosensitivity via driving glycolysis in GISTs and support potential use of IGF2 to reverse imatinib resistance in GISTs patients.
|
PMC1462997
| "Derivation of normal macrophages from human embryonic stem (hES) cells for applications in HIV gene(...TRUNCATED)
| "Many novel studies and therapies are possible with the use of human embryonic stem cells (hES cells(...TRUNCATED)
|
PMC12217045
|
M1 Macrophage‐Derived Extracellular Particles Induce Cell Death in MDA‐MB‐231 Cells
| "Triple‐negative breast cancer (TNBC), a leading cause of female mortality worldwide, presents a t(...TRUNCATED)
|
PMC10287567
|
An integrated cell atlas of the lung in health and disease
| "Single-cell technologies have transformed our understanding of human tissues. Yet, studies typicall(...TRUNCATED)
|
PMC1315352
|
Development of antibodies to human embryonic stem cell antigens
| "Using antibodies to specific protein antigens is the method of choice to assign and identify cell l(...TRUNCATED)
|
README.md exists but content is empty.
- Downloads last month
- 9