PMCID
string
Title
string
Sentences
string
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Single-agent plates measure each drug across 10 doses, while combination plates only measure drug pairs at a relative 1:1 ratio—corresponding to the diagonal of a 10 × 10 matrix.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Single-agent response values are mapped onto dose-response curves and integrated with combination data to assemble the sparse matrix (Supplementary Fig. 3c, d), which effectively cuts the required number of measurements by 90%.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
By dramatically decreasing the experimental resource consumption per combination, sparse mode increases throughput and efficiency.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For example, the workflow supports up to 135 single agents, which yields 9045 combinations and would require just 73 plates when using six replicates of single-agent plates (Supplementary Fig. 3e, f).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Consequently, sparse mode’s miniaturized design and efficient plate usage make ultra-large-scale combination screens available with minimal resources.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To enhance the utility of our sparse screening workflow, we developed an ensemble machine learning model capable of predicting the non-measured response values within a sparse matrix, effectively filling in the gaps to reconstruct the completed matrix.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We leveraged the dense mode combination data to train the model, which provided fully-measured 10 × 10 matrices alongside their corresponding single-agent responses.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
From this, we curated a training set comprising 552 matrices (198,720 measurements) and a test set of 184 matrices (66,240 measurements) (Fig. 3a).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Each non-measured index within the sparse matrix is predicted by a dedicated regression model within the ensemble, allowing all 90 models to be finely tuned for their respective targets (Fig. 3b, see “Methods” section).Fig.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
3Machine learning supplements sparse screening data.a Composition of training and test sets used to develop the ensemble model.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The training set includes 552 fully-measured 10 × 10 combination matrices (light orange, 75% of total data), while the test set has 184 matrices (dark orange, 25% of total data).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
b Overview of the ensemble machine learning model architecture.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Ninety individual models are each designed to predict the response of a single non-measured index of a sparse matrix.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Fully-measured 10 × 10 matrices from the training set are downsampled to reflect the 30 measured values collected in sparse mode (treated as features) and the index a given model aims to predict (treated as the outcome variable).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Each model undergoes hyperparameter tuning and is fit with the XGBoost regression model (see “Methods” section).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
c Model performance measured by R across the 10 respective folds for each of the 90 models.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Box plot center line represents the median (0.947), bounds are 25th and 75th percentiles (0.934 and 0.959), and whiskers extend to the most extreme data points within 1.5× the interquartile range from the box edges.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The minimum is 0.860, the maximum is 0.983.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
d Variable importance (VI) contributed by the 30 measured values toward the model predicting index (represented by a yellow circle) across the test set.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
e Comparison of observed versus predicted cell death response values for the model predicting index across the test set.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
f–i Like (d, e) showing the VI and predictive performance for models and , respectively.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
j Violin plot summarizing the Euclidean distances between the target index of a model and its corresponding feature index with the maximum importance value, across all 90 models.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The dashed line represents the median Euclidean distance (d = 2) across all samples, with model highlighted as an example (red dot).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
k Example of the model predicting index , illustrating the Euclidean distance (d = 2) from its target index to the feature index with the maximum importance value.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
l Overall predictive performance of the model across all samples in the test set (n = 184).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
m Overview of the sparse mode workflow, integrating the experimental generation of sparse matrices with the machine learning-based imputation of non-measured response values.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
a Composition of training and test sets used to develop the ensemble model.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The training set includes 552 fully-measured 10 × 10 combination matrices (light orange, 75% of total data), while the test set has 184 matrices (dark orange, 25% of total data).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
b Overview of the ensemble machine learning model architecture.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Ninety individual models are each designed to predict the response of a single non-measured index of a sparse matrix.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Fully-measured 10 × 10 matrices from the training set are downsampled to reflect the 30 measured values collected in sparse mode (treated as features) and the index a given model aims to predict (treated as the outcome variable).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Each model undergoes hyperparameter tuning and is fit with the XGBoost regression model (see “Methods” section).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
c Model performance measured by R across the 10 respective folds for each of the 90 models.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Box plot center line represents the median (0.947), bounds are 25th and 75th percentiles (0.934 and 0.959), and whiskers extend to the most extreme data points within 1.5× the interquartile range from the box edges.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The minimum is 0.860, the maximum is 0.983.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
d Variable importance (VI) contributed by the 30 measured values toward the model predicting index (represented by a yellow circle) across the test set.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
e Comparison of observed versus predicted cell death response values for the model predicting index across the test set.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
f–i Like (d, e) showing the VI and predictive performance for models and , respectively.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
j Violin plot summarizing the Euclidean distances between the target index of a model and its corresponding feature index with the maximum importance value, across all 90 models.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The dashed line represents the median Euclidean distance (d = 2) across all samples, with model highlighted as an example (red dot).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
k Example of the model predicting index , illustrating the Euclidean distance (d = 2) from its target index to the feature index with the maximum importance value.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
l Overall predictive performance of the model across all samples in the test set (n = 184).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
m Overview of the sparse mode workflow, integrating the experimental generation of sparse matrices with the machine learning-based imputation of non-measured response values.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Our model demonstrated a high predictive accuracy, with a median R of 0.945 across 10-fold cross-validation (Fig. 3c).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Analyzing variable importance (VI) revealed that predictions were most strongly influenced by measured values closest to the target index.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
For instance, the model predicting the response at row 1, column 2 (Model), primarily relied on data from its two nearest measured indices (Fig. 3d).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Despite relying heavily on only two of the 30 measured values, this model achieved strong predictive performance (r = 0.98) between the observed and predicted values in the test set (Fig. 3e).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Similar patterns were observed throughout the ensemble, with strong performance across models predicting values near the diagonal (Fig. 3f, g) and those in the lower-right quadrant of the matrix (Fig. 3h, i).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
A broader analysis of all models confirmed a proximity-based dependency, with most predictions relying on measured indices within a Euclidean distance of 1 or 2 from the target index (Fig. 3j, k).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
These findings underscore the importance of leveraging local features for predicting sparse matrix values and highlight the advantage of training dedicated models for each of the 90 non-measured indices, enabling highly targeted predictions (Supplementary Data 2).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To evaluate the ensemble’s overall performance, we compared observed and predicted cell death responses across all non-measured indices in the test set.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This yielded a strong correlation (r = 0.98, Supplementary Data 3) (Fig. 3l), demonstrating the ensemble’s reliability in capturing key combination response effects.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
By predicting non-measured values, the ensemble model enriches the sparse screening workflow, providing detailed insights that enhance the interpretability of synergy results while preserving resource efficiency (Fig. 3m).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To demonstrate the scalability and practical utility of Combocat, we applied the sparse mode workflow to screen 9045 drug combinations in the neuroblastoma cell line CHP-134 (Supplementary Data 4), representing the largest dense combination screen reported in a single cell line.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We used 135 small molecules, encompassing approved chemotherapeutics, investigational compounds, and neuroblastoma-selective agents nominated from our earlier CRISPR screens. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Supplementary Data 5).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Synergy was quantified for each pair by their mean Blissadj. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
score—the adjusted Bliss synergy averaged across the matrix diagonal (10 measured dose pairs) (Fig. 4a).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Most combinations centered near zero (Fig. 4b), indicating a predominance of additive or non-synergistic interactions and aligning with prior findings that strong synergy is a rare phenomenon.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Fig. 4Ultrahigh-throughput screen of 9045 combinations in CHP-134 cells.a Heatmap of the mean adjusted Bliss synergy scores (Blissadj.)
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
across experimentally observed 10 dose pairs for each of the 9045 tested combinations.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
b–e Histograms of the mean Blissadj. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
, Moran’s I, QC flag count (observed), and mean % cell death, respectively, across the combination data.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
c–e highlight filters applied to exclude (gray) or include (green) combinations in the final hit list.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
f Filtered combinations (n = 594), ranked by their mean Blissadj. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
scores.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The top 15 hits are labeled above.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
g–n Comparison of the top 2 hits from sparse mode (upper matrices) re-screened using dense mode (lower matrices).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The cell death (g, h, k, l) and synergy (i, j, m, n) matrices are compared between the two modes.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
a Heatmap of the mean adjusted Bliss synergy scores (Blissadj.)
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
across experimentally observed 10 dose pairs for each of the 9045 tested combinations.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
b–e Histograms of the mean Blissadj. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
, Moran’s I, QC flag count (observed), and mean % cell death, respectively, across the combination data.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
c–e highlight filters applied to exclude (gray) or include (green) combinations in the final hit list.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
f Filtered combinations (n = 594), ranked by their mean Blissadj. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
scores.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The top 15 hits are labeled above.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
g–n Comparison of the top 2 hits from sparse mode (upper matrices) re-screened using dense mode (lower matrices).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The cell death (g, h, k, l) and synergy (i, j, m, n) matrices are compared between the two modes.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To prioritize drug combinations for further exploration, we used three filtering criteria that captured spatially coherent synergy patterns (Moran’s I), data reliability (QC flag counts), and biological relevance (mean cell death response, Fig. 4c–e, see “Methods” section).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
After filtering, 594 top-scoring pairs remained, ranked by mean Blissadj. (
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
(Fig. 4f).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
To validate these sparse mode predictions, we re-screened a subset of 40 combinations in dense mode: the top 30 from sparse mode (expected to display the highest synergy) plus 10 randomly selected combinations that had been excluded.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Measuring these combinations in a fully sampled 10 × 10 format enabled a direct comparison to sparse mode results.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Most top-ranked drug pairs retained strong synergy patterns, confirming that sparse mode can effectively prioritize synergistic combinations (Supplementary Fig. 4a).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
In contrast, the 10 random pairs exhibited weaker synergy in dense mode, matching their lower sparse mode scores.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
We additionally compared single-agent dose-response curves for the 47 unique drugs in the validation set and observed close alignment (rho = 0.858) of IC50 values and fitted dose-response curves (Supplementary Fig. 4b, c), reinforcing the consistency of single-agent measurements obtained in sparse mode.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Among the validated hits, the ATM inhibitor AZD1390 in combination with the PARP inhibitors Olaparib or Rucaparib stood out for their pronounced synergy (Fig. 4g, i, k, m).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This aligns mechanistically with targeting complementary DNA damage repair pathways: PARP inhibition leads to accumulation of single-strand breaks, while ATM inhibition impairs the double-strand break repair response.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Re-testing these combinations in dense mode confirmed high agreement (Fig. 4h, j, l, n), meaning strong, reliable synergies could be identified from the starting set of over 9000 screened combinations.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Collectively, these data establish Combocat’s sparse mode as an efficient and scalable means of uncovering synergy even in ultra-large-scale screens.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Combocat presents a step toward more comprehensive, flexible, and scalable drug combination screening.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
By uniting acoustic liquid handling and machine learning-assisted inference, we have demonstrated the feasibility of screening thousands of combinations while retaining the resolution and interpretability crucial to discovering strong synergistic interactions.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The principles underpinning Combocat are broadly applicable, with a design that supports compatibility with various assay readouts that align with specified volumes and plate formats—such as luminescence, absorbance (Supplementary Fig. 6), fluorescence, mass spectrometry, or high-content imaging.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Furthermore, the open-source acoustic liquid handler protocols ensure adaptability to other drug dispensing platforms with comparable capabilities, establishing Combocat as a scalable foundation for diverse experimental applications.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
The functionality and workflow of Combocat are designed to be straightforward, including its minimal and intuitive analytical pipeline (Supplementary Fig. 5a).
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
This streamlined architecture, along with open-source and detailed documentation, ensures that researchers can easily implement and customize Combocat experiments according to their needs.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
Our goal is to pursue continual refinement and expansion of the platform.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
One of the most compelling aspects of Combocat lies in its potential for community-driven advancement.
PMC12705714
An open-source screening platform accelerates discovery of drug combinations
As more research groups generate dense combination data, these datasets can be anonymized and contributed back to the community for re-training and improving the ensemble machine learning model (Supplementary Fig. 5b).