Protein folding stability estimation with an explicit consideration of unfolded states

Abstract

Folding stability is a critical requirement for the vast majority of proteins. Computational methods suggested to date for the absolute folding stability ({Delta}G) prediction - including those driven from protein structure prediction AIs - show clear limitations on reproducing quantitative experimental values. Here we present IFUM, a deep neural network that jointly estimates {Delta}G and the equilibrium ensemble of folded and unfolded states represented by their residue-pair distance probability distributions. This joint learning considerably enhances the {Delta}G prediction accuracy against the scenario where {Delta}G prediction was learned alone. To improve the model, we extend the dataset beyond previous related works to include the Mega-scale small proteins and disordered proteins for training as well as wild-type natural proteins with sizes up to 869 residues for validation. We show that IFUM is robust to various protein types and sizes, and is capable of accurately predicting more general types of mutational effects such as sequence insertions or deletions. The applicability of IFUM is demonstrated through two real-world design challenges. First, for blind-tested protein engineering scenarios containing many sequence substitutions and insertions, good correlation with experimental melting temperatures (Tm) is observed. Second, for the de novo design selection, IFUM shows considerably improved performance over broadly used AlphaFold-based metrics. IFUM is a free software available at github.com/HParklab/IFUM and also through Google Colab.