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

Area of Science:

• Computational protein science
• Drug discovery
• Artificial intelligence applications

Background:

• Understanding protein interactions and dynamics is crucial for drug discovery.
• Artificial intelligence (AI) offers advanced predictive learning for complex biological systems.
• Repurposing AI algorithms from other domains shows their flexibility in structural and biological applications.

Purpose of the Study:

• To examine AI model repurposing across domains for protein interaction and dynamics tasks.
• To analyze how AI performance is shaped by inherited characteristics from original applications.
• To provide guidance on selecting, evaluating, and integrating AI models in computational protein science.

Main Methods:

• Cross-domain adaptation framework for AI models.
• Analysis of inductive biases, learning objectives, and representation choices in AI.
• Comparison of AI approaches with physics-based modeling.

Main Results:

• AI models show success and systematic failures in protein interaction and dynamics tasks.
• Differences in AI behavior compared to physics-based modeling are identified.
• Limitations in data, benchmarking, and emerging hybrid AI-physics workflows are highlighted.

Conclusions:

• AI offers powerful tools for protein science, but careful selection and evaluation are needed.
• Hybrid AI-physics workflows can balance efficiency with physical realism.
• This review supports more reliable and biologically meaningful AI applications in drug discovery.