FUSED: Cross-Domain Integration of Foundation Models for Cancer Drug Response Prediction

Till Rössner, Jonas Balke, Ming Tang

Keywords: drug response prediction, single cell foundation model, molecular foundation model for drugs, cross domain foundation model integration, multi-head attention

TL;DR: We present FUSED, to our knowledge the first architecture that integrates Foundation Models (FMs) from distinct domains (molecular FMs for drugs and single-cell FMs for cell lines), to enable more accurate and robust cancer drug response prediction.

Abstract:

AI-driven methods for predicting drug responses hold promise for advancing
personalized cancer therapy, but cancer heterogeneity and the high cost of data
generation pose substantial challenges. Here we explore the transfer learning capa
bility and introduce FUSED (Fusion of Foundation Model Embeddings for Drug
Response Prediction), a novel architecture for cross-domain foundation model (FM)
integration. By systematically benchmark FMs across two domains– molecular
FMfor drugs and single-cell FM for cell lines, we demonstrate that integrating
single-cell FMs substantially reduces the number of input features required for
cell line representation. Among FMs, Molformer significantly outperforms Chem
BERTa, and scGPT surpasses scFoundation in predictive accuracy and training
stability. Moreover, integrating single-cell FMs improves performance in both
drug-known and leave-one-drug-out scenarios. These findings highlight the poten
tial of cross-domain FM integration for more efficient and robust drug response
prediction.

Submission Number: 30