SPIRED-Fitness: an end-to-end framework for the prediction of protein structure and fitness from single sequence

Abstract

Significant research progress has been made in the field of protein structure and fitness prediction. Particularly, single-sequence-based structure prediction methods like ESMFold and OmegaFold achieve a balance between inference speed and prediction accuracy, showing promise for many downstream prediction tasks. Here, we propose SPIRED, a novel single-sequence-based structure prediction model that exhibits comparable performance to the state-of-the-art methods but with approximately 5-fold acceleration in inference and at least one order of magnitude reduction in training consumption. By integrating SPIRED with downstream neural networks, we compose an end-to-end framework named SPIRED-Fitness for the rapid prediction of both protein structure and fitness from single sequence. SPIRED-Fitness and its derivative SPIRED-Stab achieve state-of-the-art performance in predicting the mutational effects on protein fitness and stability metrics, respectively.

GDFold2: a fast and parallelizable protein folding environment with freely defined objective functions

“…Hence, the non-statistical-potential-style loss function in GDFold2 successfully complements the 2D geometric prediction results by statistics derived from high-resolution protein structures, nearly without causing information loss. More GDFold2 folding cases could be found in the Supplementary Materials of our SPIRED paper 21 . Unlike SPIRED, Cerebra prediction results involve full information required for all-atom structural reconstruction.…”

“…We employed the highly efficient GDFold2 pipeline to generate a large amount of decoy structures based on randomly masked SPIRED and Cerebra predictions for a number of proteins in their training sets (see Methods ) and used these decoys to train the QA model. Consequently, to avoid information leaking, during the evaluation of our QA model, we chose proteins in the CAMEO testing dateset 21 that are absent in the SPIRED/Cerebra training sets, and selected RoseTTAFold predictions as constraints to generate 24 structural models by GDFold2 for each protein target. Generally, the structural models produced for each individual protein target present certain degree of structural dissimilarity, with an average deviation ( i .…”

“…The second point, particularly, restricts its application in modeling protein structures using diversified constraints. For instance, in our recent work, we developed a lightweight protein structure prediction method SPIRED 21 , which utilizes 2D information to predict L (i.e. the number of residues) sets of C α atom coordinates, each taking the local coordinate system of one individual residue as reference.…”