Predicting Drug Responses by Machine Learning

Abstract

Machine learning (ML) has revolutionized the pharmaceutical industry in recent decades, influencing molecule design, drug target identification, biomarker discovery, and various stages of drug development. This transformation, driven by the synergy between ML and high-throughput drug screening technologies, has broadened the scope for novel treatments and therapeutic indications. This dissertation explores the application of ML algorithms in surmounting fundamental challenges in drug development, including stabilizing high-throughput screening outcomes and transforming initial discoveries into clinical practices. The first part of the dissertation enhances the generalizability of drug-based experimental results. Our first project in this part assesses the reproducibility across experimental batches in vitro, using data from DrugComb, the most extensive public portal for combination treatment currently available. A critical experimental variable identified is the concentration selection for dose-response matrices. To address this, a concentration imputation method is implemented during feature preparation, markedly improving the predictive transferability of ML algorithms across datasets. The next project shifts focus to the transferability of results between different biological contexts (in vivo and in vitro). I present the winning algorithm from the Malarian DREAM Challenge, which predicts artemisinin resistance in laboratory isolates using models trained on transcriptome and response data from Plasmodium falciparum strains. This project tackles challenges arising from different microarray platforms, response evaluation methods, and biological backgrounds. A rank normalization method is employed to mitigate platform discrepancies, and model visualization highlighted key genes and pathways indicative of artemisinin resistance in both in vivo and in vitro settings. The second part discusses ML's role in discovering new treatments, using DNA damage response (DDR) targeted combination therapy as a case study. An original high-throughput screening dataset featuring 87 anti-cancer drugs and 12 cancer tissues is introduced for DDR combination therapy. Effective and synergistic treatments were identified in combination with ATM, ATR, or DNAPK inhibitors. An ML model is developed, incorporating molecular readouts, synthetic lethality, drug-target interaction, biological networks, chemical structure, and drugs' modes of action, to predict DDR combination treatment responses in new biological contexts. This model shows promise in prescribing optimal DDR treatments based on the patient's biological characteristics, enhancing treatment responses. Furthermore, a core gene panel of only 40 genes was found to be more efficient in predicting DDR combination treatment responses than using full genomic or transcriptomic profiles, leading to the development of a rapid-selection interface for DDR combination treatments in pharmaceutical and clinical applications.