#38 · Structural bio & protein design

Graph and embedding models infer molecular interactions from sequence and network data

The common thread across these items is inferring molecular relationships from sequence, structure, or network data rather than direct measurement. A relation-aware graph-hypergraph network with SE(3)-equivariant encoders addresses antibody CDR sequence-structure co-design and reports gains over the compared approaches. A binding-site study pairs ProtT5 embeddings of ligand-contacting residues with mutual nearest-neighbor networks across 121,673 binding sites to map multiscale organization. A separate preprint introduces antagonistic-cooperative polarization, a target-wise sign principle with a node-separable polarization index, applied through null-model analysis to reconstructed pathway-associated regulatory networks. What is new is the extension of network inference from physical to regulatory interactions, paired with explicit structural constraints and null-model testing. The items are heterogeneous, so this thread is dominant rather than shared by every item; the antibody co-design work is primarily a structural modeling method.