Interpretable dose-dependent cancer-drug response prediction for precision medicine

Abstract

Personalized medicine represents a transformative approach in healthcare by tailoring treatment strategies to an individual’s unique genetic profile, unlike traditional one-size-fits-all approaches. This strategy aims to optimize therapeutic outcomes by aligning medical treatments with the specific characteristics of each patient. However, the extensive anti-cancer drug space poses significant challenges, as the number and diversity of compounds make it difficult to identify the most proper therapy for an individual patient. Recent advances in deep learning offer a promising solution by predicting which anti-cancer drugs are likely to be most beneficial, potentially overcoming the limitations imposed by the large number of therapeutic options. However, although deep learning models can suggest anti-cancer drugs that may be effective for individual patients, there are unsolved limitations. One significant challenge is the lack of interpretability inherent in many deep learning approaches, which can obscure the rationale behind their predictions. Furthermore, while these models can indicate promising drug candidates, they fall short of providing practical treatment specifics, such as optimal drug dosing regimens. In this dissertation, I addressed these limitations by proposing advanced drug response prediction models with biological knowledge. First, I proposed a deep learning model, HiDRA, to predict the response of drug on given cancer cell using gene expressions of the cancer cell, structural fingerprint of the compound, and biological pathway. Compared to state-of-the-art (SOTA) models, the proposed model showed the best performance in the Half maximal inhibitory concentration (IC50) prediction task. Through attention analysis, the proposed model demonstrated its interpretability by identifying drug target genes and target pathways. The predictive power of HiDRA was also proved by the in vitro experiments. In the second study, I focused on developing a model named DD-PRiSM, that predicts the efficacy of combination therapy at arbitrary concentrations. To predict the efficacy of combination therapy, I enhanced HiDRA to develop a model that predicts the efficacy of monotherapy at arbitrary concentrations based on Hill equation. Using the predicted efficacies and mechanisms of action (MoA) of the monotherapies, the proposed model successfully predicted the efficacy of combination therapy with high performance. DD-PRiSM demonstrated its interpretability through the additional analyses that confirm the characteristics of drug combinations prone to synergy effects and the differences in synergy effects across various cancer types. In conclusion, the model proposed in this dissertation for predicting the efficacy of anticancer treatment at arbitrary concentrations is expected to make a significant contribution to the advancement of personalized medicine.