PMCID
string
Title
string
Sentences
string
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This communal effort could result in a dynamic resource capable of delivering progressively more accurate synergy estimates with minimal experimental overhead.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
While Combocat enables high-throughput combination screening in various contexts, some limitations are important to acknowledge.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
First, the accuracy of sparse mode inference is inherently dependent on the quality and diversity of the training data.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We tested three progressively stringent stratifications of our dense mode dataset (Supplementary Fig. 7a–c) and found essentially identical predictive performance (r≈0.97, Supplementary Fig. 7d–f), demonstrating no simple data leakage in predicting responses in Neuroblastoma.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
However, as new data are collaboratively added, preventing leakage through continuous hold-out testing will be essential to maintain the model’s robustness and generalizability.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Second, although dense and sparse modes’ experimental measurements largely agree, they will differ in some circumstances.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Sparse mode uses miniaturized volumes and a different plate format, which can occasionally introduce shifts in drug potency or dynamic range.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This can be related to technical factors like compound dispersion in smaller volumes, plate layout effects, or variations in cell confluence in smaller wells.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Finally, while Combocat supports Bliss and Loewe synergy metrics, each carries assumptions and limitations.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For example, Bliss can misestimate synergy for drugs targeting similar pathways, and Loewe cannot calculate synergy in the absence of dose-equivalence (Supplementary Fig. 8a, b)—making it undefined in 61.5% of sparse mode dose combinations tested (Supplementary Fig. 8c).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We compared the mean synergy scores in our combinations passing QC and filtering steps (n = 800) and observed a modest agreement between the Bliss and Loewe models (Spearman rho = −0.672, p < 2.2e-16, Supplementary Fig. 8d).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Importantly, additional reference models exist such as Highest Single Agent (HSA), Zero Interaction Potency (ZIP)(which models changes in potency of the dose-response curves), and others, which can be evaluated in future extensions of the platform.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We envision leveraging Combocat’s scalability to investigate more of the vast unexplored drug combination space, mapping out the synergy landscapes of thousands of drug combinations across many cell lines (Supplementary Fig. 5c).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
By integrating sparse mode workflows with continuously evolving machine learning models, rich synergy datasets can be generated, capable of capturing complex biological responses.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Such comprehensive synergy maps will empower researchers to identify conserved interactions, reveal lineage-specific vulnerabilities, and guide the rational selection of combination therapies.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This study complied with all relevant ethical regulations.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The CHP-134 cell line was purchased from Sigma (catalog #6122002) and maintained in RPMI 1640 medium (Corning, catalog #10-041-CM) supplemented with 10% FBS (Corning, catalog #35-015-CV) at 37 °C with 5% CO2.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Mycoplasma was routinely tested using the MycoAlert mycoplasma detection kit (Lonza, catalog # LT07-118), and the cells were verified negative for contamination.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For experiments with bacterial cultures, the Escherichia coli BW25113 strain (Horizon Discovery Ltd.) was cultured at 37 °C in Mueller-Hinton (MH) agar or broth (BBL).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Cultures in exponential phase (OD600 0.4–0.6) were diluted to OD600 of 0.0005 for plating in 384-well format.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Compounds were sourced from commercial vendors (see Supplementary Data 5).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Stocks were solubilized in DMSO and stored at −80 °C until 24 h before use, at which point they were thawed at room temperature and centrifuged at 1100 RPM for 5 min.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Across all assays, compounds were screened at 10 concentrations using 3-fold serial dilutions, typically with an upper limit of 25 µM. Most compounds were screened below and above their physiologically-relevant doses, and in a way positioning their suspected IC50s as the midpoint to attempt achieving a complete dose-response effect.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The choice of 135 compounds used in sparse mode comes from drugs targeting the top neuroblastoma-selective sensitizers from earlier CRISPR-anchor screens.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Specifically, drugs were chosen using the following approach:Our previous CRISPR knockout-anchor screens investigated 18 cell lines using 7 neuroblastoma standard-of-care drugs and 1 investigational compound (Doxorubicin, Etoposide, Retinoic Acid, Topotecan, Vincristine, Cisplatin, Phosphoramide Mustard, and JQAD1).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The CRISPR knockout library was designed to target 655 known druggable genes.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
These screens were designed to identify druggable gene knockouts that sensitize cancer cells to standard-of-care neuroblastoma drugs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To identify these high-potential drug-sensitizing gene knockouts from the resulting dataset, we applied a custom Bayesian hierarchical model (described in the “Methods” section of Lee and Wright et al.) to nominate genes with differential sensitization effects in the Neuroblastoma group, relative to the outgroup cell.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We then compiled the list of drugs that targeted the resulting neuroblastoma-selective sensitizing gene knockouts.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For genes targetable by >1 drug, the list was reduced manually, considering factors such as clinical approval, redundancy with other drugs, number of analogs, and compound availability.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Finally, we manually added high-priority investigational compounds to the list after consultation with multiple pediatric oncologists at St. Jude, yielding the final number of 135 unique drugs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The list of drugs is available in Supplementary Data 5.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Our previous CRISPR knockout-anchor screens investigated 18 cell lines using 7 neuroblastoma standard-of-care drugs and 1 investigational compound (Doxorubicin, Etoposide, Retinoic Acid, Topotecan, Vincristine, Cisplatin, Phosphoramide Mustard, and JQAD1).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The CRISPR knockout library was designed to target 655 known druggable genes.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
These screens were designed to identify druggable gene knockouts that sensitize cancer cells to standard-of-care neuroblastoma drugs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To identify these high-potential drug-sensitizing gene knockouts from the resulting dataset, we applied a custom Bayesian hierarchical model (described in the “Methods” section of Lee and Wright et al.) to nominate genes with differential sensitization effects in the Neuroblastoma group, relative to the outgroup cell.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We then compiled the list of drugs that targeted the resulting neuroblastoma-selective sensitizing gene knockouts.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For genes targetable by >1 drug, the list was reduced manually, considering factors such as clinical approval, redundancy with other drugs, number of analogs, and compound availability.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Finally, we manually added high-priority investigational compounds to the list after consultation with multiple pediatric oncologists at St. Jude, yielding the final number of 135 unique drugs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The list of drugs is available in Supplementary Data 5.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To enable comprehensive and reproducible measurement of pairwise dose combinations in a 10 × 10 matrix format, we designed a 384-well plate template that accommodates three replicate 10 × 10 combination matrices in a 384-well plate format, along with three replicates of the dose-response curves for each single-agent and twelve replicates of each positive and negative control (Fig. 1c).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
In each well, 200 nL of total compound was transferred via an Echo 655 acoustic liquid handler (Beckman Coulter): 100 nL of compound + 100 nL of DMSO for single-agent wells, or 100 nL of each compound for combination wells.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This was followed by 40 μL of cells, added using a Multidrop combi (Thermo Fisher).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Maintaining a 40 μL volume of cells minimized edge effects arising from evaporation.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Each dense mode plate tested exactly one drug combination, making the total plate count for a dense mode experiment scale linearly with the number of drug pairs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Sparse mode was designed to increase throughput and reduce resource usage by miniaturizing the assay into a 1536-well plate format (Supplementary Fig. 3a, b).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Three key features enable ultra-large combination screens in sparse mode:Miniaturized volumes: Only 20 nL of compound is transferred (10 nL of each drug for combinations or 10 nL of drug + 10 nL of DMSO for single-agent wells), followed by 4 μL of cells.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Edge wells are filled with culture medium to mitigate evaporation.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Doses are interleaved to minimize well-to-well signal contamination.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Measurement of the diagonal only: Each 10 × 10 matrix is represented by 10 dose pairs that combine each drug at a relative 1:1 ratio.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For example, the 6th-highest tested dose of Drug 1 with the 6th-highest tested dose of Drug 2.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This reduces the total measured dose pairs by 90%, compared to a fully-measured 10 × 10 matrix (Supplementary Fig. 3d).Separate single-agent and combination plates: Single-agent and combination plates are tested separately, with single-agents testing each drug’s 10-dose series, and combinations testing only the diagonal dose pairs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This allows single agents to be mapped onto combination matrices without being re-measured.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For example, to screen the combinations A_B and A_C, dense mode would measure A twice, whereas sparse mode would measure A once.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Within-plate controls facilitate normalization before integration into assembled matrices.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Miniaturized volumes: Only 20 nL of compound is transferred (10 nL of each drug for combinations or 10 nL of drug + 10 nL of DMSO for single-agent wells), followed by 4 μL of cells.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Edge wells are filled with culture medium to mitigate evaporation.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Doses are interleaved to minimize well-to-well signal contamination.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Measurement of the diagonal only: Each 10 × 10 matrix is represented by 10 dose pairs that combine each drug at a relative 1:1 ratio.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For example, the 6th-highest tested dose of Drug 1 with the 6th-highest tested dose of Drug 2.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This reduces the total measured dose pairs by 90%, compared to a fully-measured 10 × 10 matrix (Supplementary Fig. 3d).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Separate single-agent and combination plates: Single-agent and combination plates are tested separately, with single-agents testing each drug’s 10-dose series, and combinations testing only the diagonal dose pairs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This allows single agents to be mapped onto combination matrices without being re-measured.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For example, to screen the combinations A_B and A_C, dense mode would measure A twice, whereas sparse mode would measure A once.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Within-plate controls facilitate normalization before integration into assembled matrices.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The total number of plates required for a sparse mode screen can be calculated as:1[12pt] $$=c(r+)})$$=crn135+n2135where c is the total number of cell lines, r is the total number of single-agent plate replicates, and n is the number of unique drugs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The 135 value comes directly from the number of unique drugs able to fit within the usable (non-edge) wells of the 1536-well plates.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For both dense and sparse modes, compounds were transferred into empty 384- or 1536-well plates (Corning) using the Echo 655 acoustic liquid handler with Combocat protocol files (available at combocat.stjude.org).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
After compound transfer, cells were dispensed at seeding densities of 1000 cells/well in 384-well plates or 200 cells/well in 1536-well plates.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For E. coli cells (which were only screened in 384-well plates), an OD600 of 0.0005 was used.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
In 1536-well plates, cells were strained through a 70 µM cell strainer before being dispensed.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The final DMSO concentration after cell addition was 0.5% in all wells.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Plates were incubated at 37 °C for 72 h. Cell viability was assessed after the 3-day treatment using CellTiter-Glo (or BacTiter-Glo for bacterial cells) by adding either 25 µL (to 384-well plates) or 2 µL (to 1536-well plates) of prepared reagent and reading luminescence on an EnVision plate reader (PerkinElmer).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For the absorbance-based readout of cell viability, 8 µL of MTS reagent (Abcam) at 0.5× concentration was added to the cells at the endpoint time.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The MTS reagent was incubated with cells for 3 h, and plates were read at 490 nanometers on a Cytation 5 plate reader (BioTek).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The quality of all screened assay plates was assessed using the Z′ metric, which measures quality as a function of signal window and data variation, given by:2[12pt] $$^ }=1-_}+_}}_}-_}|}$$′=1−3σNeg+3σPosμNeg−μPoswhere σNeg/Pos and μNeg/Pos represent the standard deviation and means of the controls, respectively.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
A Z′ value of ≥0.5 is generally regarded as an excellent separation of controls.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The Combocat ensemble machine learning model was developed as a collection of 90 individual models, each built to predict one of the 90 non-measured indices of a sparse matrix.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To construct these models, an input dataset was first assembled containing all fully-measured data collected via dense mode, totaling over 800 combination matrices and their single-agent response values (Supplementary Data 1).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The input dataset was first filtered to remove any matrices with a mean cell death value < 10%.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Next, this dataset was split into 75% for training and 25% for the test set, respectively.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This resulted in a training set of 552 matrices (198,720 measurements) and a test set of 184 matrices (66,240 measurements).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Note that while the combination matrix itself has dimensions of 10 × 10, including the single-agent data effectively extends this into an 11 × 11 grid, where the 1st column and 11th row represent the single-agent responses (Supplementary Fig. 9a).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This format was used for associating a model with its respective index.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Each model was trained on 30 features: response values from 10 single-agent indices each from Drug 1 and Drug 2, and 10 indices corresponding to the matrix diagonal, which represent each drug’s concentration combined at a relative 1:1 ratio (Supplementary Fig. 9b).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
These 30 features correspond to the 30 indices measured experimentally in the sparse mode workflow (Supplementary Fig. 3d) and were used across the input data to generate the models predicting the 90 total indices of a sparse matrix (Supplementary Fig. 9c).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The training of each model involved fitting the tree-based regression model XGBoost.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Hyperparameter tuning was enabled to optimize model performance, which was accomplished using a space-filling grid design to efficiently test 40 combinations of the hyperparameters (detailed in Supplementary Data 6).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
A 10-fold cross-validation was used to select the optimal set of hyperparameters based on the lowest R value.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Once identified, each model was trained using its respective optimal hyperparameters on the entire training set.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The performance of each model was evaluated across all folds, gauged by R and RMSE.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For each model, the variable importance was measured across each of the 30 features (30 measured response values from sparse matrices) to evaluate their predictive impact (Supplementary Data 7).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Next, the trained models were validated on their test sets to ensure their predictive capability by comparing the observed vs. predicted cell death response values.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The comparison of observed vs. predicted response values across the 90 models collectively was used to summarize the predictive performance of the ensemble model.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Finally, the ensemble model was serialized in a format that allows for rapid deployment for future sparse matrix predictions.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
All model generation was performed using the tidymodels framework (tidymodels.org).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To ensure data integrity and mitigate the impact of spurious measurements, we implemented a rigorous quality control (QC) procedure focusing on three key metrics: single-agent standard deviation, residuals from single-agent dose-response curves, and monotonicity of the cell death response.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For each single-agent drug, we first calculated the standard deviation of the percentage cell death (% cell death) values across all replicates at each dose level, resulting in ten standard deviation values per agent corresponding to the ten doses tested.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Any dose with a percentage cell death standard deviation exceeding a predefined threshold (Tdefault) of 29 was flagged for disqualification.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
All combination data points involving that particular dose were also flagged for disqualification (Supplementary Fig. 2e).