Unveiling the true impact of drug and cell linerepresentations in drug synergy prediction

Drug combination therapy holds promise as an effective strategy for treating complexdiseases such as cancer. However, due to the vast combinatorial space of drugcombinations, experimental screening of all of them is not feasible. Computationalmodels have been developed to prioritize drug pairs that could work synergistically toaccelerate experimental screening efforts. These models are trained on large datasetsof previously reported drug combination measurements and use rich representationsof drugs and cell lines that encode chemical, structural, and biological properties.In this thesis, we first aimed to improve upon our previous synergy predictor, Match-Maker, by incorporating richer biological information such as pathways and mechanismof action or alternative drug representations. Despite all our efforts, none ofthe models could perform better. Motivated by these findings, we tested a morestraightforward approach by replacing detailed feature representations with one-hotencodings of drugs and cell lines. Surprisingly, these models stripped of chemical andbiological information can come very close to the results trained with rich biologicaland chemical information.Here, in this thesis, we systematically experimented with published synergy predictionmodels by replacing drug representations and cell line features with a simpleone-hot encoding of drugs and cell lines in various evaluation settings. Regardlessof the drug input feature or the architecture, we observe that the simple one-hotencoding baseline performs similarly in all models. This unexpected result suggests) that the representations serve as simple identifiers and models that capture generalco-variation patterns of synergy measurements rather than learning chemical or biologicalinformation. This could be why the models do not generalize well to newdrugs and cell lines. While synergy prediction models are still beneficial in decidingon what pairs to test within a panel of drugs and cell lines, these results demonstratethat alternative approaches are needed for developing synergy prediction models thatcould work across new drugs, cell lines, and patients.