Paper by: Samuel W. Canner, Lei Lu,  Sho S. Takeshita, Jeffrey J. Gray

Read more on bioRxiv

This study builds the first benchmark for testing AI models on protein–carbohydrate docking, showing that current tools work reasonably well but need improvement, a key step toward decoding how carbohydrates influence health, disease, and drug design.

What this paper is about

Proteins in our body don’t just interact with other proteins; they also bind to sugars (carbohydrates). These protein–carbohydrate interactions play critical roles in immunity, cell signaling, infection, and many diseases. But predicting how a protein docks with a carbohydrate is extremely tricky because carbohydrates are flexible, can take many forms, and are harder to model than small drug-like molecules.

The authors of this paper created the BCAPIN (Benchmark of Carbohydrate Protein Interactions) dataset, a high-quality dataset of known protein–carbohydrate complexes. They then tested modern AI-based protein structure prediction tools (like AlphaFold3, Boltz-1, Chai-1, DiffDock, RosettaFold-All Atom) on this dataset to see how well they predict protein–carbohydrate docking.

Why it’s scientifically important

• Fills a gap: Until now, no benchmark existed for testing AI models specifically on protein–carbohydrate interactions. This paper sets the standard.
• New evaluation metric: They introduced DockQC, a scoring system to judge docking quality more consistently.
• Findings: AI models can predict protein–carbohydrate docking reasonably well (~85% acceptable accuracy) but struggle with longer, more complex carbohydrates.

How this plays into real life

Understanding protein–carbohydrate docking is crucial because these interactions influence how viruses infect cells, how our immune system recognizes threats, and how diseases like cancer progress. By creating the first benchmark for testing AI models on protein–carbohydrate interactions, this study helps researchers better predict these bindings, paving the way for improved drug design, vaccine development, and a deeper understanding of how carbohydrates control biology.

Abstract

Advances in deep learning have produced a range of models for predicting the protein-sugar interactome; however, structural docking of noncovalent protein-carbohydrate complexes remains largely unexplored. Although all-atom structure prediction models like AlphaFold3 (AF3), Boltz-1, Chai-1, DiffDock, and RosettaFold-All Atom (RFAA) were validated on protein-small molecule complexes, no benchmark or evaluation exists specifically for noncovalent protein-carbohydrate docking. To address this, we developed a high-quality dataset of experimental structures – Benchmark of CArbohydrate Protein Interactions (BCAPIN). Using BCAPIN and a novel evaluation metric, DockQC, we assessed the performance of all-atom structure prediction models on non-covalent protein-carbohydrate docking. We found all methods achieved comparable results, with an 85% success rate for structures of at least acceptable quality. However, we found that the predictive power of all models declined with increasing carbohydrate polymer length. With the capabilities and limitations assessed, we evaluated AF3’s ability to predict binding for a set of putative human carbohydrate binding and carbohydrate non-binding proteins. While current models show promise, further development is needed to enable high-confidence, high-throughput prediction of the complete protein-sugar interactome.