Protein folding stability estimation with explicit consideration of unfolded states

Abstract

Folding stability is crucial for the vast majority of proteins. Computational methods suggested to date for the absolute folding stability (ΔG) prediction, including those driven from protein structure prediction AIs, show clear limitations in reproducing quantitative experimental values. Here we present IFUM, a deep neural network that jointly estimates ΔG and the equilibrium ensemble of folded and unfolded states represented by residue-pair distance probability distributions. This joint learning considerably enhances prediction accuracy compared to learning ΔG alone. Trained on a dataset including Mega-scale small proteins, disordered proteins, and wild-type natural proteins, IFUM is robust to various protein types and can accurately predict complex mutational effects like insertions or deletions. Here, we show that IFUM effectively guides real-world design challenges, exhibiting strong correlation with experimental melting temperatures in protein engineering and outperforming AlphaFold-based metrics in de novo design selection.