Recalibrating AlphaFold2 Confidence Metrics for Dynamic Peptide Ensembles Using NMR Experimental Data

AlphaFold2 (AF2) has transformed protein structure prediction by producing highly accurate three-dimensional structures alongside residue-level confidence estimates (pLDDT). However, pLDDT reflects predicted structural accuracy rather than conformational stability or dynamics, potentially leading to discrepancies with experimentally observed ensembles. This project aims to investigate whether AF2 confidence estimates can be systematically recalibrated using local structural environments and biochemical context. We propose to leverage a dataset of approximately 6,000 proteins and 500,000 residues with paired AF2 predictions and NMR-derived structures. Incorporating secondary-structure features, structural variability, and residue interactions into a GENConv-based graph neural network we regress NMR-derived ground truth LDDT scores. The goal is to develop a physically informed correction framework that improves the reliability of AF2 confidence estimates, particularly for flexible and dynamic protein regions.

Team: Sahadot Hossen, Fahim Ahmed, Amitava Roy, PhD, Prof. Amin Ahsan Ali, PhD.