PhD Defence of David Hostallero - Applications of Deep Learning and Graph Representation Learning in Precision Cancer Medicine

Abstract

Computationally designing personalized treatment plans to increase a cancer patient's chances of recovery using their molecular profiles has been one of the major objectives of precision cancer medicine. Despite the advancement of high-throughput sequencing and artificial intelligence, drug response prediction has remained a challenging task. This thesis presents novel methodologies for predicting responses to drug treatments, addressing challenges such as limited clinical data and drug-specific biases. Leveraging available datasets, we explored the utility of different information modalities into predictive models.

First, we focused on clinical drug response prediction using only preclinical data. This stemmed from the current situation of cancer drug response datasets, wherein drug responses for preclinical cancer cell line (CCL) samples treated with hundreds of drugs are widely available, while clinical drug responses of tumors are only available in small patient cohorts for a handful of drugs. We developed a deep learning pipeline that leverages tissue information to bridge discrepancies between CCL and tumor samples, enabling models to distinguish between sensitive and resistant patients.

We then ventured towards improving drug representation using knowledge graphs composed of CCLs, drugs, and genes. Unlike previous methods that solely rely on the structural properties of drug molecules, we integrated additional response-relevant information, such as molecular profiles of extremely sensitive/resistant CCLs, CRISPR gene effects, and drug targets. Our analyses demonstrated superior performance compared to existing methods and baseline approaches.

Beyond drug response prediction, we also identified potential biomarkers of drug response for each model that we presented. This not only enhances model interpretability, but also produces data-driven hypotheses. Many implicated genes and pathways were supported by literature, and in some cases, experimentally validated. We introduced a graph-based interpretation method to provide further insights and visualize the prediction process at a high level.

The content of this thesis not only improve drug response prediction but also sheds light on potential therapeutic targets, contributing to the advancement of precision cancer medicine.