Fast Model-based Protein Homology Discovery without Alignment

Abstract

The need for quick gene categorization tools is growing as more genomes are sequenced. To evaluate a newly sequenced genome, the genes must first be identified and translated into amino acid sequences, which are then categorized into structural or functional classes. Protein homology detection using sequence alignment algorithms is the most effective way for protein categorization. Discriminative approaches such as support vector machines (SVMs) and position-specific scoring matrices (PSSM) derived from PSI-BLAST have recently been used to improve alignment algorithms. However, if a fresh sequence is being aligned, alignment algorithms take time. must be compared to a large number of previously published sequences — the same is true for SVMs. Building a PSSM for the PSSM is even more time-consuming than a fresh order It would take roughly 25 hours to implement the best-performing approaches to classify the sequences on today's computers. Describing a novel genome (20, 000 genes) as belonging to one single organism. There are hundreds of classes to choose from, though. Another flaw with alignment algorithms is that they do not construct a model of the positive class, instead of measuring the mutual distance between sequences or profiles. Only multiple alignments and hidden Markov models are common classification approaches for creating a positive class model, but they have poor classification performance. A model's advantage is that it may be evaluated for chemical features that are shared by all members of the class to get fresh insights into protein function and structure. We used LSTM to solve a well-known remote protein homology detection benchmark, in which a protein must be categorized as a member of the SCOP superfamily. LSTM achieves state-of-the-art classification performance while being significantly faster than other algorithms with similar classification performance. LSTM is five orders of magnitude quicker than the quickest SVM-based approaches and two orders of magnitude faster than methods that perform somewhat better in classification (which, however, have lower classification performance than LSTM). We applied LSTM to PROSITE classes and analyzed the derived patterns to test the modeling capabilities of the algorithm. Because it does not require established similarity metrics like BLOSUM or PAM matrices, LSTM is complementary to alignment-based techniques. The PROSITE motif was retrieved by LSTM in 8 out of 15 classes. In the remaining seven examples, alternative motifs are developed that, on average, outperform the PROSITE motifs in categorization.