Mardiyyah/TAPT_data_V2_split · Datasets at Hugging Face

PMCID stringclasses 24 values	Title stringclasses 24 values	Sentences stringlengths 2 40.7k
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	After fine-tuning on the training set of 1000 molecules (Table 3, entry 1), our model could retrieve 14% of the 6051 test molecules.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	When scaling down to a smaller training set of 50 molecules (the model gets trained on less than 1% of the data!),
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	it can still reproduce 2.5% of the test set, and performs 21.6 times better than the unbiased model (Table 3, entry 2).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	We additionally examined whether the model benefits from transfer learning.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Pretraining on the large data set is thus crucial to achieve good performance against Staphylococcus aureus.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	EOR: Enrichment over random.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Fine-tuning learning rate = 10.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	No pretraining.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	8 generate-test cycles.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Different training strategies on the Staphylococcus aureus data set with 1000 training and 6051 test examples.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Fine-tuning the pretrained model performs better than training from scratch (lower test loss [cross entropy] is better).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	The experiments we conducted so far are applicable if one already knows several actives.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	However, in drug discovery, one often does not have such a set to start with.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Then, the best molecules are selected, and based on the gained knowledge new molecules are designed, which closes the cycle.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	We thus do not use a set of known actives to start the structure generation procedure (see Figure 10).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Scheme of our de novo design cycle.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Molecules are generated by the chemical language model and then scored with the target prediction model (TPM).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	The inactives are filtered out, and the RNN is retrained.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Here, the TPM is a machine learning model, but it could also be a robot conducting synthesis and biological assays, or a docking program.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	We started with 100,000 sampled molecules from the unbiased chemical language model.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Then, using our target prediction model, we extracted the molecules classified as actives.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	After that, the RNN was fine-tuned for 5 epochs on the actives, sampling ≈10,000 molecules after each epoch.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	This EOR is higher than if the model is retrained directly on a set of 50 actives (entry 2).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Additionally, we obtained 60,988 unique molecules that the target prediction model classified as active.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	This suggests that, in combination with a target prediction or scoring model, our model can at least simulate the complete de novo design cycle.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	The Levenshtein distance of, e.g., benzene c1ccccc1 and pyridine c1ccncc1 would be 1.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	This is supported also by the distribution of the nearest neighbor fingerprint similarities of training and rediscovered molecules (ECFP4, Tanimoto, Figure 12).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Many rediscovered molecules are in the medium similarity regime.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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).
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Violin plot of the nearest-neighbor ECFP4-Tanimoto similarity distribution of the 50 training molecules against the rediscovered molecules in Table 3, entry 2.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Because we perform transfer learning, during fine-tuning, the model does not “forget” what it has learned.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	The model can generate large sets of novel molecules with similar physicochemical properties to the training molecules.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	This can be used to generate libraries for virtual screening.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	We see three main advantages of our method.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	First, it is conceptually orthogonal to established molecule generation approaches, as it learns a generative model for molecular structures.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Furthermore, it merges structure generation and optimization in one model.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	A weakness of our model is interpretability.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	In addition, different approaches for target prediction, for example, docking, could be evaluated.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Deep learning is not a panacea, and we join Gawehn et al. in expressing “some healthy skepticism” regarding its application in drug discovery.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	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.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	The complexity of the problem certainly warrants the investigation of novel approaches.
PMC5785775	Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.	Eventually, success in the wet lab will determine if the new wave of neural networks will prevail.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Despite their crucial role in health and disease, our knowledge of immune cells within human tissues remains limited.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Here, we surveyed the immune compartment of 16 tissues from 12 adult donors by single-cell RNA sequencing and VDJ sequencing generating a dataset of 360,000 cells.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	To systematically resolve immune cell heterogeneity across tissues, we developed CellTypist, a machine learning tool for rapid and precise cell type annotation.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Using this approach, combined with detailed curation, we determined the tissue distribution of finely phenotyped immune cell types, revealing hitherto unappreciated tissue-specific features and clonal architecture of T and B cells.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Our multi-tissue approach lays the foundation for identifying highly resolved immune cell types by leveraging a common reference dataset, tissue-integrated expression analysis and antigen receptor sequencing.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	The immune system is a dynamic and integrated network made up of many different cell types distributed across the body that act together to maintain tissue homeostasis and mediate protective immunity.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	In recent years, a growing appreciation of immune ontogeny and diversity across tissues has emerged.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	For example, we have gained insights into how macrophages derived in embryogenesis contribute to the unique adaptation of tissue-resident myeloid cells, such as Langerhans cells in the skin, microglia in the brain and Kupffer cells in the liver (1–3).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Other populations, such as innate lymphoid cells (ILCs), including natural killer (NK) cells and non-conventional (NKT, MAIT and γδ) T cells, have circulating counterparts but are highly enriched at barrier/mucosal sites where they mediate tissue defense and repair (4).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	In addition, following resolution of an immune response, antigen-specific, long-lived tissue-resident memory T cells (TRMs) persist in diverse sites, where they provide optimal protection against secondary infections (reviewed in (5–7)).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Studies in mice have revealed the central role of tissue immune responses in protective immunity, anti-tumor immunity, and tissue repair; however, human studies have largely focussed on peripheral blood as the most accessible site.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Given that circulating immune cells comprise only a subset of the entire immune cell landscape, understanding human immunity requires a comprehensive assessment of the properties and features of immune cells within and across tissues.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Recent progress in the analysis of tissue immune cells have implemented organ-focused approaches (8–12), while use of tissues obtained from organ donors allows for analysis of immune cells across multiple sites across an individual (13–19).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	We previously reported single-cell RNA sequencing (scRNA-seq) analysis of T cells in three tissues from two donors (20), identifying tissue-specific signatures.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	However, despite the effort to assemble murine (21) and human (22, 23) multi-tissue atlases, large-scale cross-tissue scRNA-seq studies that investigate tissue-specific features of innate and adaptive immune compartments have not been reported.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Furthermore, a fundamental challenge of increasingly large single cell transcriptomics data sets is cell type annotation, including identifying rare cell subsets and distinguishing novel discoveries from previously described cell populations.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Currently available automated annotation workflows leverage organ-focussed studies and lack a comprehensive catalogue of all cell types found across tissues.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Therefore, a single unified approach is required in order to provide an in-depth dissection of immune cell type and immune state heterogeneity across tissues.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	To address these needs, we comprehensively profiled immune cell populations isolated from a wide range of donor-matched tissues from 12 deceased individuals, generating nearly 360,000 single cell transcriptomes.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	To systematically annotate multi-tissue immune cells we developed CellTypist, a machine learning framework for cell type prediction initially trained on data from studies across 20 human tissues (see Supplementary Text) and then updated and extended by integrating immune cells from our dataset.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	To systematically assess immune cell type heterogeneity across human tissues, we performed scRNA-seq on 16 different tissues from 12 deceased organ donors (Fig. 1, A and B, and table S1).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	The tissues studied included primary (bone marrow) and secondary (spleen, lung-draining and mesenteric lymph nodes) lymphoid organs, mucosal tissues (gut and lung), as well as blood and liver.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	When available, we also included additional donor-matched samples from tissues such as thymus, skeletal muscle and omentum.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Immune cells were isolated from tissues as detailed in the Methods.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	After stringent quality control, we obtained a total of 357,211 high quality cells, of which 329,762 belonged to the immune compartment (fig. S1, A and B).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Robust cell type annotation remains a major challenge in single-cell transcriptomics.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	To computationally predict cellular heterogeneity in our dataset, we developed CellTypist (24), a cell type database, that in its current form is focused on immune cells, that provides a directly interpretable pipeline for the automatic annotation of scRNA-seq data (Fig. 1C).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	One of the unique and valuable aspects of CellTypist is that its training set encompasses a wide range of immune cell types across tissues.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	This is of critical importance given that immune compartments are shared across tissues, warranting accurate and automated cell annotation in an organ-agnostic manner.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	In brief, to develop CellTypist we integrated cells from 20 different tissues from 19 reference datasets (fig. S2) where we had deeply curated and harmonised cell types at two knowledge-driven levels of hierarchies (figs. S3 to S8).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	This was followed by a machine learning framework to train these cells using logistic regression with stochastic gradient descent learning (see methods).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Performance of the derived models, as measured by the precision, recall and global F1-score, reached ~0.9 for cell type classification at both the high- and low-hierarchy levels (Fig. 1C and fig. S9, A and B).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Notably, representation of a given cell type in the training data was a major determinant of its prediction accuracy (fig. S9C), implying higher model performance can be achieved by incorporating additional datasets.
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	Moreover, CellTypist prediction was robust to differences between training and query datasets including gene expression sparseness (fig. S10) and batch effects (fig. S11).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	First we applied CellTypist’s high-hierarchy (i.e. low-resolution, 32 cell types) classifier to our cross-tissue dataset (Fig. 1D), and found 15 major cell populations (fig. S1C).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	At this level of resolution, clear compositional patterns emerged in lymphoid versus non-lymphoid tissues, and within the lymphoid tissues between spleen versus lymph nodes, and appeared not to be driven by differences in dissociation protocols (fig. S12).
PMC7612735	Cross-tissue immune cell analysis reveals tissue-specific features in humans.	As the training datasets of CellTypist contained hematopoietic tissues with definitive annotations for progenitor populations, the classifier was able to resolve several progenitors including hematopoietic stem cells and multipotent progenitors (HSC/MPP), promyelocytes, early megakaryocytes, pre-B and pro-B cells.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

After fine-tuning on the training set of 1000 molecules (Table 3, entry 1), our model could retrieve 14% of the 6051 test molecules.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

When scaling down to a smaller training set of 50 molecules (the model gets trained on less than 1% of the data!),

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

it can still reproduce 2.5% of the test set, and performs 21.6 times better than the unbiased model (Table 3, entry 2).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

We additionally examined whether the model benefits from transfer learning.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Pretraining on the large data set is thus crucial to achieve good performance against Staphylococcus aureus.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

EOR: Enrichment over random.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Fine-tuning learning rate = 10.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

No pretraining.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

8 generate-test cycles.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Different training strategies on the Staphylococcus aureus data set with 1000 training and 6051 test examples.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Fine-tuning the pretrained model performs better than training from scratch (lower test loss [cross entropy] is better).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

The experiments we conducted so far are applicable if one already knows several actives.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

However, in drug discovery, one often does not have such a set to start with.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Then, the best molecules are selected, and based on the gained knowledge new molecules are designed, which closes the cycle.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

We thus do not use a set of known actives to start the structure generation procedure (see Figure 10).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Scheme of our de novo design cycle.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Molecules are generated by the chemical language model and then scored with the target prediction model (TPM).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

The inactives are filtered out, and the RNN is retrained.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Here, the TPM is a machine learning model, but it could also be a robot conducting synthesis and biological assays, or a docking program.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

We started with 100,000 sampled molecules from the unbiased chemical language model.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Then, using our target prediction model, we extracted the molecules classified as actives.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

After that, the RNN was fine-tuned for 5 epochs on the actives, sampling ≈10,000 molecules after each epoch.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

This EOR is higher than if the model is retrained directly on a set of 50 actives (entry 2).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Additionally, we obtained 60,988 unique molecules that the target prediction model classified as active.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

This suggests that, in combination with a target prediction or scoring model, our model can at least simulate the complete de novo design cycle.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

The Levenshtein distance of, e.g., benzene c1ccccc1 and pyridine c1ccncc1 would be 1.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

This is supported also by the distribution of the nearest neighbor fingerprint similarities of training and rediscovered molecules (ECFP4, Tanimoto, Figure 12).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Many rediscovered molecules are in the medium similarity regime.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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).

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Violin plot of the nearest-neighbor ECFP4-Tanimoto similarity distribution of the 50 training molecules against the rediscovered molecules in Table 3, entry 2.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Because we perform transfer learning, during fine-tuning, the model does not “forget” what it has learned.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

The model can generate large sets of novel molecules with similar physicochemical properties to the training molecules.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

This can be used to generate libraries for virtual screening.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

We see three main advantages of our method.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

First, it is conceptually orthogonal to established molecule generation approaches, as it learns a generative model for molecular structures.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Furthermore, it merges structure generation and optimization in one model.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

A weakness of our model is interpretability.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

In addition, different approaches for target prediction, for example, docking, could be evaluated.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Deep learning is not a panacea, and we join Gawehn et al. in expressing “some healthy skepticism” regarding its application in drug discovery.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

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.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

The complexity of the problem certainly warrants the investigation of novel approaches.

PMC5785775

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks.

Eventually, success in the wet lab will determine if the new wave of neural networks will prevail.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Despite their crucial role in health and disease, our knowledge of immune cells within human tissues remains limited.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Here, we surveyed the immune compartment of 16 tissues from 12 adult donors by single-cell RNA sequencing and VDJ sequencing generating a dataset of 360,000 cells.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

To systematically resolve immune cell heterogeneity across tissues, we developed CellTypist, a machine learning tool for rapid and precise cell type annotation.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Using this approach, combined with detailed curation, we determined the tissue distribution of finely phenotyped immune cell types, revealing hitherto unappreciated tissue-specific features and clonal architecture of T and B cells.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Our multi-tissue approach lays the foundation for identifying highly resolved immune cell types by leveraging a common reference dataset, tissue-integrated expression analysis and antigen receptor sequencing.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

The immune system is a dynamic and integrated network made up of many different cell types distributed across the body that act together to maintain tissue homeostasis and mediate protective immunity.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

In recent years, a growing appreciation of immune ontogeny and diversity across tissues has emerged.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

For example, we have gained insights into how macrophages derived in embryogenesis contribute to the unique adaptation of tissue-resident myeloid cells, such as Langerhans cells in the skin, microglia in the brain and Kupffer cells in the liver (1–3).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Other populations, such as innate lymphoid cells (ILCs), including natural killer (NK) cells and non-conventional (NKT, MAIT and γδ) T cells, have circulating counterparts but are highly enriched at barrier/mucosal sites where they mediate tissue defense and repair (4).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

In addition, following resolution of an immune response, antigen-specific, long-lived tissue-resident memory T cells (TRMs) persist in diverse sites, where they provide optimal protection against secondary infections (reviewed in (5–7)).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Studies in mice have revealed the central role of tissue immune responses in protective immunity, anti-tumor immunity, and tissue repair; however, human studies have largely focussed on peripheral blood as the most accessible site.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Given that circulating immune cells comprise only a subset of the entire immune cell landscape, understanding human immunity requires a comprehensive assessment of the properties and features of immune cells within and across tissues.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Recent progress in the analysis of tissue immune cells have implemented organ-focused approaches (8–12), while use of tissues obtained from organ donors allows for analysis of immune cells across multiple sites across an individual (13–19).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

We previously reported single-cell RNA sequencing (scRNA-seq) analysis of T cells in three tissues from two donors (20), identifying tissue-specific signatures.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

However, despite the effort to assemble murine (21) and human (22, 23) multi-tissue atlases, large-scale cross-tissue scRNA-seq studies that investigate tissue-specific features of innate and adaptive immune compartments have not been reported.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Furthermore, a fundamental challenge of increasingly large single cell transcriptomics data sets is cell type annotation, including identifying rare cell subsets and distinguishing novel discoveries from previously described cell populations.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Currently available automated annotation workflows leverage organ-focussed studies and lack a comprehensive catalogue of all cell types found across tissues.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Therefore, a single unified approach is required in order to provide an in-depth dissection of immune cell type and immune state heterogeneity across tissues.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

To address these needs, we comprehensively profiled immune cell populations isolated from a wide range of donor-matched tissues from 12 deceased individuals, generating nearly 360,000 single cell transcriptomes.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

To systematically annotate multi-tissue immune cells we developed CellTypist, a machine learning framework for cell type prediction initially trained on data from studies across 20 human tissues (see Supplementary Text) and then updated and extended by integrating immune cells from our dataset.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

To systematically assess immune cell type heterogeneity across human tissues, we performed scRNA-seq on 16 different tissues from 12 deceased organ donors (Fig. 1, A and B, and table S1).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

The tissues studied included primary (bone marrow) and secondary (spleen, lung-draining and mesenteric lymph nodes) lymphoid organs, mucosal tissues (gut and lung), as well as blood and liver.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

When available, we also included additional donor-matched samples from tissues such as thymus, skeletal muscle and omentum.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Immune cells were isolated from tissues as detailed in the Methods.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

After stringent quality control, we obtained a total of 357,211 high quality cells, of which 329,762 belonged to the immune compartment (fig. S1, A and B).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Robust cell type annotation remains a major challenge in single-cell transcriptomics.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

To computationally predict cellular heterogeneity in our dataset, we developed CellTypist (24), a cell type database, that in its current form is focused on immune cells, that provides a directly interpretable pipeline for the automatic annotation of scRNA-seq data (Fig. 1C).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

One of the unique and valuable aspects of CellTypist is that its training set encompasses a wide range of immune cell types across tissues.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

This is of critical importance given that immune compartments are shared across tissues, warranting accurate and automated cell annotation in an organ-agnostic manner.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

In brief, to develop CellTypist we integrated cells from 20 different tissues from 19 reference datasets (fig. S2) where we had deeply curated and harmonised cell types at two knowledge-driven levels of hierarchies (figs. S3 to S8).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

This was followed by a machine learning framework to train these cells using logistic regression with stochastic gradient descent learning (see methods).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Performance of the derived models, as measured by the precision, recall and global F1-score, reached ~0.9 for cell type classification at both the high- and low-hierarchy levels (Fig. 1C and fig. S9, A and B).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Notably, representation of a given cell type in the training data was a major determinant of its prediction accuracy (fig. S9C), implying higher model performance can be achieved by incorporating additional datasets.

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

Moreover, CellTypist prediction was robust to differences between training and query datasets including gene expression sparseness (fig. S10) and batch effects (fig. S11).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

First we applied CellTypist’s high-hierarchy (i.e. low-resolution, 32 cell types) classifier to our cross-tissue dataset (Fig. 1D), and found 15 major cell populations (fig. S1C).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

At this level of resolution, clear compositional patterns emerged in lymphoid versus non-lymphoid tissues, and within the lymphoid tissues between spleen versus lymph nodes, and appeared not to be driven by differences in dissociation protocols (fig. S12).

PMC7612735

Cross-tissue immune cell analysis reveals tissue-specific features in humans.

As the training datasets of CellTypist contained hematopoietic tissues with definitive annotations for progenitor populations, the classifier was able to resolve several progenitors including hematopoietic stem cells and multipotent progenitors (HSC/MPP), promyelocytes, early megakaryocytes, pre-B and pro-B cells.