C-P.35: From PMGen to PMpan: Fast and Accurate Peptide-MHC Modeling, Sampling and Binding Prediction
While accurate structural modeling of pMHC interactions is critical for immunotherapy design, current tools struggle with narrow class coverage, length restrictions, and a lack of structure-aware sampling. Because most binding prediction models rely entirely on sequence data, they tend to overfit to specific MHC alleles, failing to generalize to rare variants. To bridge this generalization gap, we hypothesized that incorporating 3D structural properties could enable cross-allele transfer learning. To this end, we distilled our previous high-quality but slow prediction model (PMGen) into PMGen2, achieving ultra-fast structure prediction. We integrated this with new tools for motif clustering (PepCluster2), profile sampling (PMscore), and binding prediction (PMpan). Ultimately, our results show that structure-based modeling successfully generalizes performance from human-trained models to non-human species.
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826