Token-Level Transformer improves breast cancer prediction by jointly analyzing genomic and clinical data

In a new paper on arXiv, researchers Suxing Liu and Byungwon Min introduce a Token-Level Cross-Modal Transformer with Contrastive Multi-Task Learning for breast cancer subtype classification and survival prediction. Existing methods in precision oncology treat genomic and clinical modalities as monolithic feature vectors, preventing fine-grained token-level interactions. They also rely on simple linear weighting or late averaging for cross-modal fusion, and optimize survival and classification objectives independently—missing a joint regularization signal.

The proposed model addresses all three limitations. It processes each modality at the token level, allowing structured token exchange between genomic and clinical data. A contrastive multi-task learning framework jointly optimizes classification and survival prediction, sharing representations across tasks. This approach enables more nuanced patient stratification and survival risk assessment, moving beyond black-box fusion. The work is available on arXiv and represents a step toward more precise, multimodal AI in cancer care. By integrating heterogeneous data at a finer granularity, the model could improve clinical decision-making and personalized treatment planning.