Predictions and experimental validation of fold-switching proteins

National Library Of Medicine

Investigators

Linked publications, trials & patents

Abstract

No introductory-level biochemistry textbook is complete without a chapter about how a proteins primary sequence of amino acids determines its fold. So far, nearly all computational work has focused on predicting a single protein structure from the proteins amino acid sequence. Our research challenges the one-sequence-one-structure paradigm. In 2018, we found nearly 100 examples of proteins that can adopt more than one stable fold. The structural heterogeneity of these fold-switching proteins allows them either to perform more than one function or to be highly regulated in cells. These functional changes affect human health as several fold-switching proteins are associated with human diseases such as cancer, autoimmune dysfunction, and bacterial infections. Now we are taking this research to the next level by (1) developing computational approaches to predict which amino acid sequences can switch folds (2) testing our predictions experimentally, and (3) collaborating with others to understand their functional importance. This year, we developed CF-random, a generalizable approach that predicts fold switching from genomic sequences with known homologs in the PDB. This method outperforms the powerful algorithms AlphaFold2, RoseTTAFold, and ESMFold in predicting fold switching. It also outperforms SPEACH-AF and AF-Cluster, competing methods for predicting conformational heterogeneity in proteins. We published a Comment in Nature reporting these results (PMID: 39972235, could not be added to publication list). We ran CF-random on the E. coli genome and estimate that ~5% of these proteins switch folds. We are collaborating with Brian Volkman to test these predictions experimentally. We have also found that CF-random is capable of predicting new fold switchers that have not been deposited in the Protein Data Bank (PDB). Our studies of CF-random have indicated fundamental limitations in AI-based models in predicting fold-switching proteins. With the eventual goal of developing AI-based models that more accurately predict fold switching and other conformational changes important to human disease, we are investigating why AI-based models, such as the Nobel-Prize winning AlphaFold2 (AF2), struggle to predict fold switching. In collaboration with Matthew Coudron at NIST, we have found that AF2 associates sequence patterns with specific folds. Some of these associations result from spurious correlations that lead to inaccurate predictions. We are devising ideas to generate new models that rely on more robust sequence information. Such models will allow for more robust predictions of mutational effects on protein structure and function, as well as effects of alternative splicing, critical and understudied topics related to human disease. We are collaborating with G. Marius Clore (NIDDK) to mechanistically characterize the fold switching of a bacterial protein, RfaH, which regulates the expression of virulence genes associated with human infections. This mechanistic characterization aims to reveal principles of fold switching that have not been discovered yet. Finally, we have been working with David Levens (NCI) to model an alternative conformational state of the human cancer protein FUBP1, which may explain its alternative topoisomerase activity.