A research team has built a system that predicts how proteins interact even when those proteins were absent from training data entirely.
The model, called MKGR, combines two types of information: region-aware encoding of protein sequences and four separate biomedical knowledge graphs linking proteins to drugs, diseases, miRNA, and lncRNA. Graph attention encoders pull modality-specific embeddings from sparse associations, while a gating module at the pair level decides how much weight to give sequence evidence versus graph evidence for each candidate pair. A bridge reconstruction objective keeps the graph learning grounded by forcing the model to recover shared protein-entity associations. Tests on two benchmark datasets — covering both novel-old and novel-novel cold-start scenarios — showed MKGR outperforming sequence, network, and knowledge-graph baselines across five standard metrics.
The cold-start problem is a genuine bottleneck in drug discovery. Most protein interaction models lean heavily on network topology, which means they fail the moment a newly characterized protein shows up with no known connections. A model that handles those unknowns better could shorten the gap between identifying a disease mechanism and finding a molecule that disrupts it.
Protein interaction prediction has attracted a wave of deep learning approaches over the past few years, but most gains have come from proteins that are already well-documented. MKGR's value, if it holds up beyond these benchmarks, is in the long tail — the understudied proteins where the biology is murkiest and the clinical stakes are often highest.