#9 · ML / AI methods

Contrastive and graph encoders for biomolecular interaction and expression prediction

The dominant thread is the use of pretrained or external biological structure to shape learned representations for prediction tasks. AbAgKer draws on pre-trained representations and epitope pseudo-structure priors, DynaTCR uses dynamic hard-negative graph ensembles for TCR-epitope binding, and DFCE-KanT applies contrastive learning to predict spatial gene expression from histology images. LFFA reuses frozen foundation-model features for few-shot pathology biomarker prediction, while MRGBMDAT and HMA-GCA apply graph encoding, bilinear fusion, and gated cross-attention to miRNA-disease and circRNA-miRNA associations. Not all items share this emphasis: TRACE reports near-chance results under an audited protocol, and a knowledge graph benchmark shows that train-test redundancy and degree bias inflate link prediction performance. The notable addition across this set is closer attention to evaluation rigor alongside architectural design.