From memorization to generalization: Why physics will improve machine learning -based prediction of protein...

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• AlphaFold-like models excel at predicting single protein structures.
• Reliable prediction of protein-protein interactions (PPIs), especially antibody-antigen docking, remains a significant challenge.
• Current limitations often arise from data scarcity and reliance on pattern memorization.

Purpose of the Study:

• To review emerging physics-integrated machine learning approaches for improving PPI prediction generalization.
• To categorize strategies enhancing the physical plausibility of predictions.
• To outline a path from memorization-based models to physically generalizable models.

Main Methods:

• Enriching model inputs with physics-based sampling (e.g., molecular dynamics, fast Fourier transform ensembles).
• Designing neural network architectures with strict geometric inductive biases (e.g., SE(3)-equivariance).
• Constraining generative models using physical energy functions or potentials.

Main Results:

• Hybrid strategies combining machine learning with physics principles show promise for overcoming data scarcity limitations.
• These approaches aim to enforce physical plausibility throughout the prediction pipeline.
• Improved generalization capabilities are observed for out-of-distribution targets compared to standard models.

Conclusions:

• Physics-integrated machine learning offers a promising direction for advancing protein-protein interaction prediction.
• Hybrid strategies represent a crucial step towards achieving true physical generalization in structural biology predictions.
• This transition from memorization to physical generalization is key for tackling complex biological problems like antibody-antigen docking.