Table 1.
Publicly available high-throughput drug combination screening datasets and large-scale cancer cell line genomics and transcriptomics datasets that can be used to develop drug synergy prediction models.
The datasets that were used in this work are highlighted in bold.
Fig 1.
A representation of the general architecture of the multimodal DL models developed in this study.
A model is defined as a combination of the learning algorithm itself and the data preparation steps required beforehand.
Fig 2.
Performance scores (Spearman correlation—left—and R2—right) of the different models tested in this study.
(A) Performance scores of models with different gene expression feature-encoding subnetworks. (B) Performance scores of models with different drug encoding subnetworks. (C) Performance scores of models trained with and without mutation (mut) and copy number variation (cnv) data in addition to gene expression and drug features. (D) Performance scores of non-DL models compared with one of the best DL models developed in this study.
Fig 3.
Top 20 most important features, ranked by mean absolute SHAP values.
Table 2.
Mean absolute SHAP values for features grouped by input type.
SHAP values for groups of features were calculated by summing the SHAP values of all of the individual features, and then the mean absolute SHAP value across all samples was calculated for the grouped features.
Fig 4.
Most important features (ranked by absolute SHAP values) for test set example 11835, shown as waterfall plots.
The plot starts at a base value of -22.754, which is the expected model output (determined based on a background dataset). Each row shows how each feature positively or negatively contributes to move the value from the expected output to the predicted value for this sample. (A) Top 20 features (from all features). (B) Top 20 drugA features. (C) Top 20 drugB features. (D) Top 20 gene expression features.