Deep Learning-Based Methods for Drug Synergy Prediction
Halil İbrahim Kuru
PhD. Student
(Supervisor: Assoc.Prof.Ercüment Çiçek) Computer Engineering Department
Bilkent University
Abstract: Combination therapy involves administering multiple drugs concurrently and offers enhanced efficacy, reduced toxicity, and mitigation of drug resistance. However, as the number of possible drug combinations scales exponentially with the size of the drugs, the identification of drugs that would work synergistically remains an experimental challenge. This thesis presents a structured progression of three deep learning frameworks that advance in-silico drug–drug synergy prediction by improving predictive accuracy, enhancing clinical relevance, and deepening interpretability.
The first model, MatchMaker, learns drug representations conditioned on gene expression profiles to capture cellular context jointly with the chemical structure. Benchmarking on the DrugComb dataset demonstrates that MatchMaker reduces the mean squared error of predicting drug pair synergy scores on cell lines by over one-third relative to the state-of-the-art models. Matchmaker establishes a robust and scalable baseline for high-throughput combinatorial screening on cell lines. The models trained on cell lines like MatchMaker can predict synergistic drug pairs, but do not model patient heterogeneity. Building upon its foundation, the second model—the Personalized Deep Synergy Predictor (PDSP)—addresses the critical need for patient-specific drug synergy prediction for the first time in the literature. Patient-level synergy data is scarce and prohibits training deep learning models. PDSP employs a multi-task learning strategy, coupling drug synergy prediction with a monotherapy response prediction task for which patient-level data is more abundant. PDSP applies transfer learning to fine-tune model parameters using such ex-vivo monotherapy data from individual patients. In leukemia samples, PDSP achieves a 27% increase in accuracy over state-of-the-art baselines such as MatchMaker with a low false-positive rate. PDSP demonstrated the feasibility of personalized combination prediction with limited patient-specific measurement data for the first time.
Both Matchmaker and PDSP use aggregated synergy scores that are calculated over the dose-response matrix, where rows and columns corresponds to different dosages of the given drug combination, and each entry in the matrix contains a measurement of the biological response (i.e., cell viability or inhibition) when we combine the two drugs at specific doses. The third model, DeepSynBa, goes beyond the single exaggerated synergy scores to the entire dose–response landscape. By embedding a two-dimensional generalized Hill equation within a deep learning architecture, DeepSynBa decouples the distinct dimensions of synergistic potency and efficacy. It enables flexible, post hoc computation of a range of drug interaction metrics and delivers mechanistic insight into the nature and extent of the interaction. On the NCI-ALMANAC dataset, DeepSynBa reduces dose-surface prediction error by 40–60% relative to the state-of-the-art tensor factorization methods and accurately localizes dose regions of maximal synergy. Collectively, these three models present a coherent trajectory—from scalable general drug combination prediction models to personalized predictors, to dose-resolved and mechanistically grounded inference. Beyond oncology, the underlying methodological innovations—spanning multi-task learning, transfer learning, and mechanistic parameterization—are readily generalizable to other therapeutic domains such as infectious, metabolic, and neurological disorders, where combination regimens are standard care. The approaches developed in this work represent a significant step toward accelerating the high-throughput screening efforts in identifying synergistic drug pairs. These models also pave the way to the clinical deployment of AI-driven decision-support systems capable of recommending what to combine, for whom, and at which doses. In doing so, they promise to accelerate the rational design of combination therapies that are more effective, safer, and more precisely tailored to individual patients.
DATE: June 03, Tuesday @ 13:30 Place: EA 409