Repurposing AI for protein interactions and dynamics: opportunities, limitations, and lessons

Abstract

Understanding protein interactions and dynamics of biological systems is central in drug discovery. Advances in artificial intelligence (AI) have expanded the scope of predictive learning for complex biological systems. Repurposing current gold-standard AI algorithms for structural and biological applications illustrates how flexible and powerful these approaches can be. In this mini-review, we examine how AI models are repurposed across domains and analyze how inductive biases, learning objectives, and representation choices inherited from their original applications shape performance in protein interaction and dynamics tasks. We discuss where AI approaches succeed, where they systematically fail, and how their behavior differs from physics-based modeling. We further highlight unresolved biological challenges, data and benchmarking limitations, and emerging opportunities for hybrid AI-physics workflows that balance efficiency with physical realism. By framing recent developments through a cross-domain adaptation framework, this review aims to provide practical guidance for selecting, evaluating, and integrating AI models in protein interaction and dynamics studies, and to support more reliable and biologically meaningful applications of AI in computational protein science.