Congrats!!!!

In sum, we present the first de novo benchmark of protein-carbohydrate using AF3 and similar methods to identify the protein-sugar interactome at a structural level!

Once again, would like to thank my coauthors, who were integral to every step!

Although our work shows the impressive capabilities of these models; our work is limited in that we have few structures in our set, and those structures are mostly globular single-domain proteins bound to short carbohydrates - similar to our work on CAPSIF.

Previously, in PiCAP we uncovered the protein-sugar interactome, identifying over 7.5k human proteins that likely bind to carbohydrates. Since AF3 and other methods do well 85% of the time - can we use AF3 to predict these structures? Simple answer: No, not just yet on a high throughput scale

But, can these models tell us when they don't perform well - using their confidence metrics like ipTM or pLDDT? The simple answer - kinda. AF3 is incredibly overconfident, but there is a moderate trend in most models from prediction confidence and prediction accuracy.

Across the board, all models perform great at protein-carbohydrate docking, achieving 85% of top-5 structures with at least acceptable quality! However, the largest limitation is that overall performance decreases as saccharide length increases.

In this work, we developed a unique test set of 20 proteins to evaluate the performance of AF3, Boltz-1, Chai-1, DiffDock, and RFAA at protein-carbohydrate docking. We additionally made a new metric: DockQC, inspired by protein-protein docking to evaluate their performances across many test cases!

I’d like to thank my PI Dr @jeffreyjgray.bsky.social
and co-mentor Dr. Ronald L Schnaar for their advise on this project. I’d also like to thank Sergey Lyskov for implementing the ROSIE server!

I would also like to thank the @rosettacommons.bsky.social at large, especially Yijie Luo for providing us with their designed carbohydrate non-binder dataset (which they used to train their CLIMBS algorithm, also just published on BioRxiv!).

If you’d like to use CAPSIF2 or PiCAP, you can do so freely at our github link: github.com/Graylab/picap or you can use our server on ROSIE: r2.graylab.jhu.edu/apps/ !

We also provide an updated model of our CAPSIF, a model for predicting carbohydrate binding regions of proteins. Comparing CAPSIF2 with @parthbibekar.bsky.social’s PeSTo-Carbs, finding their model outperforms CAPSIF2 on the TS90 dataset, but, CAPSIF2 modestly outperforms P-C on a larger dataset.

Now that the stage is set: what proteins bind to carbohydrates? We used PiCAP on three proteomes: E. Coli, mice, and humans and found that PiCAP predicts 35-40% of proteins bind to carbohydrates. And we provide a table of all proteins and their predictions for open scientific use and validation!

PiCAP does well on the test set -- but how does it do on other datasets that aren’t manually sculpted? We compare PiCAP to LectomeXplore finding a 99.5% agreement between the methods. We also interrogated the ganglioside interactome finding that PiCAP has a strong correlation with the experiments!

To do this we developed a deep learning (DL) model named Protein interaction of CArbohydrate Predictor (PiCAP). We find that PiCAP has approximately a 90% accuracy on identifying carbohydrate binding proteins, with limitations primarily on proteins that have reduced evolutionarily evidence (Abs).

It all started by asking just a simple question: “What proteins bind carbohydrates?” Lectins bind carbs, but, there are myriads of other proteins that bind carbohydrates (e.g. RTKs). And so we set out to find that out: can we find all the proteins in the human proteome that bind carbohydrates.