Integrated Framework for Probing Multimodal Protein Foundation Models with Structure-Functional Interpretability...

Area of Science:

• Computational biology
• Protein structure and function analysis
• Machine learning in bioinformatics

Background:

• Allosteric regulation is key to protein function, but computationally distinguishing allosteric from orthosteric binding sites is difficult.
• Multimodal protein foundation models integrate diverse biological data (sequence, structure, dynamics) but their effectiveness in allosteric site detection is unclear.

Purpose of the Study:

• To develop a framework for evaluating multimodal protein foundation models in distinguishing allosteric and orthosteric binding sites.
• To understand how different biological signals (evolutionary, structural, functional, dynamical) contribute to model performance across varying binding site separability.

Main Methods:

• Introduced a computational framework using modality embedding ablations, encoder architecture comparisons, and variance decomposition.
• Evaluated the OneProt multimodal model across four datasets representing different biological complexities and binding site contexts.
• Analyzed performance across low-, intermediate-, and high-separability regimes based on dataset properties.

Main Results:

• Model performance is primarily driven by intrinsic dataset properties (63.7% of variance), not model complexity.
• Identified three separability regimes: low (poor discrimination), intermediate (multimodal integration improves performance), and high (near-ceiling performance).
• The contribution of different data modalities (pocket geometry, sequence, text, structure, dynamics) is regime-dependent.

Conclusions:

• Dataset characteristics, rather than architectural complexity, are the main determinants of multimodal model performance in binding site classification.
• The effectiveness of multimodal models relies on aligning available data modalities with the biological signatures that differentiate binding sites within specific datasets and regimes.
• Structural and molecular dynamics encoders are most beneficial in intermediate- and high-separability settings.