Enhanced HACBLalign Method using Transitional Pattern Search and Pre-Trained Classification Model for Protein Remote Homology Detection and Fold Recognition

One of the most important tasks to predict the structure of proteins is Protein Remote Homology Detection and Fold Recognition.To do this, a Hierarchical Attention-based Convolutional Neural Network with Bidirectional Long Short-Term Memory called the HACBLalign algorithm was proposed by the authors, which performs Multiple Sequence Alignments (MSAs), extracts features, and recognizes protein homologies.But, when the quantity of Protein Sequences (PSs) increases, the number of times the decision-making system runs also increases.To avoid this issue, this article proposes an Enhanced HACBLalign (EHACBLalign) method using Transitional Pattern Search (TPS) and pre-trained classification for Protein Remote Homology Detection and Fold Recognition.During the alignment stage, the intermediate sequences such as Hit Regions (HRs) are identified by the TPS.Then, the HRs are extended in middle layers and utilized as a query in all TPS iterations.Besides, the HACBLalign algorithm is applied in all intermediate layers for generating pairwise alignments.Moreover, each pairwise alignment between intermediate sequences is merged to get the final alignment.Further, various characteristics are obtained from the chosen alignment and learned by the pre-trained Convolutional Neural Network (CNN) with a softmax function for recognizing protein remote homologies precisely.This enhances the performance of the decision-making system for largescale PS databases.Finally, the test outcomes exhibit that the EHACBLalign realizes a 94.6%, 94.1%, and 93.4% accuracy on SCOP 1.53, SCOP 1.67, and superfamily corpora, respectively in Protein Remote Homology Detection and Fold Recognition.