From Scarce Functional Labels to Label-Aware Generation in Homologous Protein Families
The study addresses difficulty in annotating protein function when labels are scarce in large homologous families. They tested a two-stage light-supervision strategy: first, using a Restricted Boltzmann Machine (RBM) to infer family-specific labels from limited data, and second, using these inferred labels to generate new, functionally specific protein sequences. The findings show that family-specific RBM embeddings outperform other representations for this task.