Residual Multi-Modal Learning for Pan-Breast-Cancer Drug Response Prediction

Predicting drug sensitivity across diverse cancer cell lines remains a fundamental challenge in precision oncology, particularly for data-scarce cell lines where per-cell-line models overfit and lookup-table approaches cannot generalise to unseen biological contexts. We present DL4DR, a Two Tower Residual Late Fusion deep learning model that addresses this challenge through content-based, identity-free genomic conditioning. The Cell Line Tower encodes each cell line as a 3 x 139 x 139 genomic image - encoding gene expression, mutation severity, and copy-number variation as RGB channels - using a convolutional encoder that maps directly from biological content, never from a cell line ID. The Compound Tower combines three complementary molecular representations: D-MPNN graph message passing, ORNN octave convolutional image features, and an ECFP hard-memorization head that preserves activity-cliff resolution. Predictions are composed as a residual sum: f = fhard + {lambda}(zc). fresidual, where the learned gate $\lambda$ modulates how much interaction signal supplements the memorization baseline. Evaluated across 51 breast cancer cell lines(136,342 records), Residual Fusion outperforms the ECFP-Only baseline in 48/51 cell lines (94.1%), with {Delta}R2 > 0.02 in 26/51 (51.0%). On the leave-cell-line-out split - the decisive test of genomic generalisation - the mean {Delta} R2 = 0.016 across all 51 lines demonstrates that the genomic encoder learns transferable biological signal beyond cell line identity. External validation on 601 cell lines across 27 cancer tissue types (CellTiter-Glo dataset; 0 cell line overlap with training) achieves median R2 = 0.627, within the range of the internal random-split performance (R2 = 0.61--0.69), confirming pan-cancer generalisation. GradCAM interpretability on the Cell Line Tower recovers TP53 among the top-five cross-cell-line genomic activators (5/51 cell lines) alongside several uncharacterised candidate genes (e.g.FSIP2, 6/51) - without any prior pathway annotation - providing partial biological validation of the learned representation, while also indicating that a substantial share of the encoder's top-ranked signal corresponds to genes with no current annotation as breast cancer drivers. Code and data are available at https://github.com/bayjuan5/DL4DR.