#2 · ML / AI methods

Learned multimodal and graph representations for protein and RNA prediction

The dominant thread is learned representations that combine heterogeneous biological inputs, often with graph structure as a recurring component. TriFit fuses sequence, AlphaFold2 structure, and GNM protein-dynamics embeddings through a mixture-of-experts module, reporting AUROC 0.897 on ProteinGym. Hybrid Gated Fusion combines sequence, structure, text, and interaction-network features via bilinear gating. Graph ML comparisons on protein contact networks span spectral embeddings, simplicial complexes, kernels, and GNNs for enzyme and EC class prediction. PNABPred adds biophysical and evolutionary priors to a protein language model for binding-site prediction. What is new relative to single-modality graph models is explicit multimodal fusion, exemplified by TriFit's protein-dynamics channel and Hybrid Gated Fusion's gating. A leakage-free Parkinson's benchmark, however, finds weak discriminative signal in sequence-based representations, with best F1 0.704.