Srivaishnavi Cherukuvada
Research Mentor: Yang Li
Mentor Department: Ecology and Evolutionary Biology, LSA
Author(s): Srivaishnavi Cherukuvada , Yang Li
Session: Session 1 (9:00 AM – 9:50 AM)
Presentation Type: Poster 121
Abstract
Protein folding and three-dimensional structure are fundamental to protein stability and biological function, yet accurately predicting domain-level conformations remains challenging. Deep mutational scanning (DMS) provides large-scale experimental measurements of amino acid substitution effects, offering potential constraints for improving structural modeling. Here, we developed MutFold, a mutation-informed folding framework that integrates mega-scale mutational data to refine protein structural predictions. MutFold incorporates large-scale DMS datasets and provides an interactive visualization interface for exploring residue-level mutation effects and structural prediction confidence. To evaluate the impact of mutational constraints on structure prediction, we analyzed a dataset of over 500,000 human missense variants across 522 protein domains. Structural predictions were assessed using pLDDT as a measure of local confidence and TM-score for similarity to experimentally resolved PDB structures. Across all domains, MutFold increased model confidence in 62.6% of cases compared with baseline AlphaFold2 predictions. However, among 361 domains with available experimental structures, structural similarity improved in only 44% of targets, with TM-score increases reaching up to 0.17 in individual domains. These results indicate that while large-scale mutational data can improve model confidence, such improvements do not necessarily translate into more accurate global folds. Our findings suggest that mega-scale mutational constraints provide selective benefits for protein structure prediction and highlight the need for improved strategies to integrate experimental mutational data into structural modeling frameworks.
