Correct prediction of protein tertiary buildings from amino acid sequences stays a basic problem in computational biology. Though AlphaFold2 represents a serious advance, systematic discrepancies persist between its predictions and experimentally decided buildings. On condition that particular person residues contribute differentially to protein operate, we hypothesize that incorporating residue-specific significance metrics improves prediction accuracy. Right here, we develop i-Fold (significance Fold), an enhanced neural structure that builds upon the AlphaFold2 framework by integrating protein language model-derived residue significance scores as dynamic positional weights throughout coaching. Analysis on a benchmark check set of 3599 protein buildings reveals that the prediction error of i-Fold is lowered by 0.3 Å on common, bettering the prediction success fee by 7.6%, and constant outcomes are obtained on an unbiased check set of 167 not too long ago launched protein buildings. Notably, i-Fold demonstrates explicit enhancements for targets which might be sometimes difficult for AlphaFold2, together with ribosomal proteins, membrane proteins, and orphan proteins. Our findings point out that specific integration of evolutionary residue significance can advance the state-of-the-art in protein construction prediction, producing extra correct and generalizable fashions with out considerably growing computational value.
