Reproduce “GeoPep: A geometry-aware masked language model for protein-peptide binding site prediction” on OpenResearch

About this paper

Dian Chen, Yunkai Chen, Tong Lin, Sijie Chen, Xiaolin Cheng

GeoPep is a computational model leveraging a multimodal protein foundation model (ESM3) and Kolmogorov-Arnold Networks (KANs) to predict protein-peptide binding sites, achieving superior performance with a precision of 0.8496, recall of 0.9567, F1-score of 0.90, and accuracy of 0.9604. This method accurately localizes binding interfaces and accommodates induced-fit conformational changes, outperforming previous specialized and general protein-protein prediction techniques.

Key insights

• Multimodal protein foundation models (like ESM3) providing geometric reasoning are essential for accurate peptide binding site prediction, significantly surpassing sequence-only representations.
• Kolmogorov-Arnold Networks (KANs) offer a parameter-efficient and expressive architecture for fine-tuning large models, demonstrating faster convergence and superior performance on specialized tasks.
• Integrating distance-based geometric constraints into the loss function substantially improves the spatial coherence and physical realism of predicted binding sites.

Summary by alphaXiv.