FASTCAR: Rapid alignment-free prediction of sequence alignment identity scores using self-supervised general linear models

Abstract

Pairwise alignment has been the predominant algorithm in the field of bioinformatics since its beginning. Several applications have been made in order to speed up this algorithm using heuristics, but almost all of these methods still depend on the slow quadratic alignment algorithm. Many applications utilize sequence identity scores without the corresponding alignments, e.g. scanning a database for similar sequences to a query sequence or sequence clustering. For these applications, we propose FASTCAR, which is the first machine-learning application that predicts alignment identity scores using completely alignment-free methods. Training data are produced from the input database by a generative method, mutating sequences to generate known alignment identity scores, thereby bypassing alignment algorithms. We evaluated FASTCAR, USEARCH, and BLAST by using them to scan three large-scale databases consisting of millions of sequences. FASTCAR is faster -up to 100 times -than USEARCH and BLAST. FASTCAR has reasonable sensitivity and accuracy while achieving the highest specificity, precision, and F-measure. Identity scores produced by FASTCAR are closer to the scores of the pure alignment algorithm than those produced by USEARCH and BLAST. This is the first time when the identity scores can be obtained in linear time and linear space.

Look4TRs: A de-novo tool for detecting simple tandem repeats using self-supervised hidden Markov models

“…The labels in the case at hand are the locations of known MS. A self-supervised algorithm can generate its own labels. Earlier, we successfully invented self-supervised systems for predicting enhancers 27 , masking repeats 21 , clustering DNA sequences 28 , and predicting the identity score of two sequences in linear time 29 . Next, we discuss (i) how the labels are generated, (ii) how the HMM is trained, and (iii) how the module auto-calibrates itself.…”

“…GLMs have been applied broadly in bioinformatics. Previously, we have applied GLMs to ranking the quality of predicted protein structures [31][32][33] , filtering out spurious MS 22 , and predicting the similarity between two DNA sequences 28,29 . We have devised a similar GLM-based classifier in MeShClust 28 , which is a tool for clustering DNA sequences, and a similar GLM-based regression model in FASTCAR 29 , which is a tool for approximating the identity score between two DNA sequences in linear time.…”

“…The feature selection algorithm uses a greedy approach that selects the best feature at each step. This feature-selection strategy is the same approach we utilized in FASTCAR 29 , whereas MeShClust's classifier has four predetermined features and does not utilize this feature selection algorithm. Doing these steps accumulates features that improve the mean error (the absolute difference between the predicted identity score and that due to the alignment algorithm) at every step on the testing set, which is different from the training set.…”

LtrDetector: A modern tool-suite for detecting long terminal repeat retrotransposons de-novo on the genomic scale

“…Horner's rule is used for efficiently converting the number from its quaternary to its decimal representation [56]. We have used similar data structures successfully in other software tools [57][58][59][60]. For example, the 5-mer ACCTG is transformed to 01132 (base 4) and then to 94 (base 10), mapping it to the 94 th cell in the array.…”

MeShClust2: Application of alignment-free identity scores in clustering long DNA sequences

“…This means that the identity score can be calculated without actually calculating the alignment itself. We have applied this idea successfully in FASTCAR, which is a search tool for approximating the alignment identity score in linear time 27 . This alignment-free adaptation will allow for much faster clustering with comparable accuracy while extending the ability of MeShClust to cluster long DNA sequences.…”