Explainable AI for drug response prediction using molecular and omics data: a transformer-based approach

Abstract

Understanding the mechanisms of drug action is essential to accurately predict drug responses and advance precision medicine. Although significant progress has been made with machine learning and deep learning techniques, many existing approaches focus primarily on achieving high accuracy. This emphasis often comes at the expense of interpretability, limiting our ability to understand the complex interactions between drug molecules and their biological targets.

This study aims to investigate the applicability of transformer-based models specifically on BERT architectures to evaluate their potential effectiveness in drug response prediction using molecular and omics data. The motivation behind this investigation lies in the growing interest in applying transformer models to biomedical data, despite the limited evidence on their comparative performance in this specific domain. Although prior approaches such as Graph Neural Networks (GNN) and Convolution Neural Networks (CNN) have shown promise, it remains unclear whether transformers can offer improved performance.

Using pre-trained HuggingFace BERT models, this research will explore the findings of the predictions of drug response levels measured by IC50 values using drug molecular structures, gene expression profiles, and single nucleotide polymorphism (SNP) data. Furthermore, explainable artificial intelligence (XAI) techniques are employed to enhance model interpretability and identify biologically relevant features.