#20 · Structural bio & protein design

Fused protein language model embeddings predict enzyme, variant, and kinase properties

The shared thread is moving beyond a single sequence embedding toward multimodal fusion or context-aware representation. AltereGO combines score-level predictions from four protein representations for Gene Ontology molecular function, and KINBERT pairs a protein language model with phospho-PCA pooled screening and Bayesian hierarchical modeling for kinase-substrate interactions. SAEKP fuses enzyme and substrate representations with a site attention module that reweights them to predict kcat, Km, and Ki. MIL-BC instead represents proteins as bags of automatically detected domains for EC number prediction, improving F1 by 0.03-0.17 over a sequence baseline. AdventML and ESMRank are less representation-centric: AdventML predicts optimal temperature from transformer embeddings with resampling for extremes, and ESMRank applies learning-to-rank with overlap-aware rank fusion across about 1100 deep mutational scans. The items do not share a single method; what is new is the emphasis on combining modalities or explicit site and domain structure.