An Effective Framework for Protein Fold Prediction through the Fusion of Evolutionary Information and Attention Mechanism with Deep Neural Networks.

Abstract

Proteins are essential for the vital functions of living organisms, playing key roles in various cellular processes. Accurate prediction of protein structures is crucial, with an important step involving the identification of the structural fold of an unknown protein by comparing its sequence to known protein folds. These known folds, particularly those experimentally validated, serve as templates for modeling the structure of the unknown protein. This process, known as Protein Fold Recognition (PFR), has remained a longstanding challenge. Existing PFR methods typically use mathematical techniques that incorporate evolutionary information through Evolutionary Profiles. In this study, we propose enhancing PFR performance by integrating evolutionary information with the attention mechanism from transformer-based models. We evaluate our approach using four widely-used protein datasets: DD, EDD, TG, and the latest SCOPe version (2.08), with four proposed models. Among these, the best-performing model achieves impressive accuracies of 93.10%, 96.40%, 93.30%, and 92.24% for the respective datasets. Three of the models incorporate ESM-2, a large language model pre-trained on protein data. A comprehensive performance analysis demonstrates that the fusion of attention-based features with evolutionary features yields the most effective classification results. Notably, the inclusion of evolutionary features boosts the accuracy of ESM-2 by 6-10% on the DD, EDD, and TG datasets, while a significant increase of over 5% is observed for the SCOPe 2.08 dataset. This work marks a notable progression in Protein Fold Recognition by combining evolutionary features with attention mechanisms, leading to improved prediction accuracy.