#34 · ML / AI methods

Graph-based methods extend gene, protein, and pathway network prediction

The dominant thread is the explicit encoding of biological structure into graph-based models. GO4CD builds Gene Ontology partitions at multiple granularities for cluster-level causal discovery in gene regulatory networks. Reactome co-membership is reweighted using pathway hierarchy and sequence relations before random-walk gene prioritization. CLARA applies two-stage hierarchical selection to class imbalance in protein binding site prediction, and LTDA adds a Lorentz-manifold adapter with tangent-space diffusion for multi-label PPI type prediction. MOSAIC learns unsupervised embeddings from Krylov analysis of accumulated diffusion operators, while CDVGL-PDA integrates multi-modal knowledge through contrastive dual-view graph learning. What is new is treating prior knowledge hierarchies and geometry as structural components of the learning problem, rather than auxiliary features. MOSAIC and CDVGL-PDA rely more on learned representations than on explicit hierarchies.