PPI Candidate Ranking: Large-Scale Evaluation of a Domain Knowledge–Guided Pipeline

Computational approaches have become central to Protein–Protein Interaction (PPI) research, complementing experimental techniques that remain costly and incomplete. While modern deep learning methods capture diverse biological signals and hold promise in expanding the known interactome, empirical validation remains a critical bottleneck due to its long and expensive procedures. To address this challenge, we introduce the problem of PPI candidate ranking, aiming to prioritize interactions for experimental testing. We propose a novel framework that leverages domain knowledge through interpretability-guided ranking and further refines prioritization by integrating complementary sources of evidence, including interaction scores, structural plausibility, and biomedical language features. Evaluations on a large-scale dataset constructed from successive STRING releases demonstrate that our approach yields significant improvements over two state-of-the-art PPI prediction models, providing more accurate and biologically coherent rankings.

Understanding which proteins work together inside our cells is crucial for discovering new medicines and understanding diseases like cancer. Scientists can now use AI to predict these partnerships, but the challenge is that the AI often produces enormous lists of candidates, and the truly promising ones tend to be buried deep inside, making them impractical for lab testing. We developed a method that looks inside the AI's internal representations to understand why it recognized known protein partnerships, then uses those patterns to bring the most promising new candidates to the top of the list. Think of it like a detective studying past cases to recognize similar clues in new ones. We further refined this shortlist using biological knowledge, such as where proteins live in the cell and what functions they share. Tested on two successive releases of a major protein database, our approach improved early-ranking performance by up to a hundredfold compared to standard AI predictions, directly helping researchers decide which experiments to run first.