Introduction
Accurate prediction of cancer drug response (CDR) is critical for advancing precision oncology. Current machine learning and deep learning approaches often fail to fully capture the complex spatial structures and chemical bond interactions of drug compounds, and they struggle to integrate multi-omics data, such as gene expression and DNA methylation, from the same cell line. There is a need for modeling frameworks that can jointly learn from molecular graphs and multi-omics profiles in an end-to-end manner.
We present GramDRP, a Graphormer-based multi-modal framework for CDR prediction. GramDRP incorporates shortest distance path (SDP) and edge encodings into the attention mechanism to capture spatial and chemical features of drugs, while a multi-view interactive encoder integrates gene expression and DNA methylation data to model inter-omics relationships.On multiple benchmark datasets, GramDRP consistently outperforms state-of-the-art methods in root mean square error (RMSE), Pearson correlation, and Spearman’s rank correlation. Moreover, GramDRP identifies potential therapeutic candidates and key regulatory genes, providing mechanistic insights and supporting applications in precision oncology and drug discovery.
Figure 1. The framework of GramDRP.