Drug combination therapy has become a promising therapeutic strategy for cancer treatment. While high-throughput drug combination screening is effective for identifying synergistic drug combinations, measuring all possible combinations is impractical due to the vast space of therapeutic agents and cell lines. In this study, we propose a biologically-motivated deep learning approach to identify pathway-level features from drug and cell lines' molecular data for predicting drug synergy and quantifying the interactions in synergistic drug pairs. This method obtained an MSE of 70.6 ± 6.4, significantly surpassing previous approaches while providing potential candidate pathways to explain the prediction. We further demonstrate that drug combinations tend to be more synergistic when their top contributing pathways are closer to each other on a protein interaction network, suggesting a potential strategy for combination therapy with topologically interacting pathways. Our computational approach can thus be utilized both for prescreening of potential drug combinations and for designing new combinations based on proximity of pathways associated with drug targets and cell lines. Implications: Our computational framework may be translated in the future to clinical scenarios where synergistic drugs are tailored to the patient and additionally, drug development could benefit from designing drugs that target topologically close pathways.

Gene expression plays a key role in health and disease. Estimating the genetic components underlying gene expression can thus help understand disease etiology. Polygenic models termed “transcriptome imputation” are used to estimate the genetic component of gene expression, but these models typically consider only the cis regions of the gene. However, these cis-based models miss large variability in expression for multiple genes. Transcription factors (TFs) that regulate gene expression are natural candidates for looking for additional sources of the missing variability. We developed a hypothesis-driven approach to identify second-tier regulation by variability in TFs. Our approach tested two models representing possible mechanisms by which variations in TFs can affect gene expression: variability in the expression of the TF and genetic variants within the TF that may affect the binding affinity of the TF to the TF-binding site. We tested our TF models in whole blood and skeletal muscle tissues and identified TF variability that can partially explain missing gene expression for 1035 genes, 76% of which explains more than the cis-based models. While the discovered regulation patterns were tissue-specific, they were both enriched for immune system functionality, elucidating complex regulation patterns. Our hypothesis-driven approach is useful for identifying tissue-specific genetic regulation patterns involving variations in TF expression or binding.

Transcriptional regulation is associated with a broad range of diseases. Methods associating genetic polymorphism with gene transcription levels offer key insights for understanding the transcriptional regulation plan. The majority of gene imputation methods focus on modeling polymorphism in the cis regions of the gene, partially owing to the large genetic search space. We hypothesize that polymorphism within transcription factors (TFs) may help explain transcription levels of their transcribed genes. Here, we test this hypothesis by developing TF-TWAS: imputation models that integrate transcription factor information with transcription-wide association study methodology. By comparing TF-TWAS models to base models that use only gene cis information, we are able to estimate possible mechanisms of the TF polymorphism effect – TF expression or binding affinity within four tissues – whole blood, liver, brain hippocampus and coronary artery. We identified 48 genes where the TF-TWAS models explain significantly better their expression than cis models alone in at least one of the four tissues. Sixteen of these genes are associated with various diseases, including cancer, neurological, psychiatric and rare genetic diseases. Our method is a new expansion to transcriptome-wide association studies and enables the identification of new associations between polymorphism in transcription factor and gene transcription levels.

Computational approaches to predict drug sensitivity can promote precision anticancer therapeutics. Generalizable and explainable models are of critical importance for translation to guide personalized treatment and are often overlooked in favor of prediction performance. Here, we propose a pathway-based model for drug sensitivity prediction that integrates chemical structure information with enrichment of cancer signaling pathways across drug-associated genes, gene expression, mutation and copy number variation data to predict drug response on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset. Using a deep neural network, we outperforming state-of-the-art deep learning models, while demonstrating good generalizability a separate dataset of the Cancer Cell Line Encyclopedia (CCLE) as well as provide explainable results, demonstrated through case studies that are in line with current knowledge. Additionally, our pathway-based model achieved a good performance when predicting unseen drugs and cells, with potential utility for drug development and for guiding individualized medicine.

Supplementary Data from SynPathy: Predicting Drug Synergy through Drug-Associated Pathways Using Deep Learning