String kernels for protein sequence comparisons: improved fold recognition

Abstract

BackgroundThe amino acid sequence of a protein is the blueprint from which its structure and ultimately function can be derived. Therefore, sequence comparison methods remain essential for the determination of similarity between proteins. Traditional approaches for comparing two protein sequences begin with strings of letters (amino acids) that represent the sequences, before generating textual alignments between these strings and providing scores for each alignment. When the similitude between the two protein sequences to be compared is low however, the quality of the corresponding sequence alignment is usually poor, leading to poor performance for the recognition of similarity.ResultsIn this study, we develop an alignment free alternative to these methods that is based on the concept of string kernels. Starting from recently proposed kernels on the discrete space of protein sequences (Shen et al, Found. Comput. Math., 2013,14:951-984), we introduce our own version, SeqKernel. Its implementation depends on two parameters, a coefficient that tunes the substitution matrix and the maximum length of k-mers that it includes. We provide an exhaustive analysis of the impacts of these two parameters on the performance of SeqKernel for fold recognition. We show that with the right choice of parameters, use of the SeqKernel similarity measure improves fold recognition compared to the use of traditional alignment-based methods. We illustrate the application of SeqKernel to inferring phylogeny on RNA polymerases and show that it performs as well as methods based on multiple sequence alignments.ConclusionWe have presented and characterized a new alignment free method based on a mathematical kernel for scoring the similarity of protein sequences. We discuss possible improvements of this method, as well as an extension of its applications to other modeling methods that rely on sequence comparison.

A weighted string kernel for protein fold recognition

“…The weighted string kernel considered here, referred to as WSeqKernel, is inspired by the convolution string kernels introduced by D. Haussler [ 36 ], the local alignment kernel presented by Saigo et al [ 27 ], and the string kernel of Smale and co-workers [ 28 ]. An unweighted version was presented in details in Nojoomi and Koehl [ 29 ]. We provide here the key elements of its construction, emphasizing the differences with those kernels.…”

“… is the sequence kernel considered in this paper. Following [ 28 , 29 , 36 ], we make the following remarks: The input kernel matrix G is not a traditional substitution matrix, as it does not involve applying the logarithm function on the probability measures. While the latter is needed to make scores additive, a necessary condition to enable the use of dynamic programming algorithms to generate pairwise sequence alignment, it is not needed for the string kernel we use here.…”

“…A solution to this limitation was proposed, the so-called spaced seeds methods that defines patterns with match and possible don’t care positions [ 18 – 21 ]. Another class of alignment-free methods for comparing protein sequences that are directly relevant to this work are the string kernel based methods [ 22 – 29 ].…”

“…In this paper we describe a new weighted string kernel that attempts to combine the benefits of the local string kernels [ 27 , 28 ] that use a substitution matrix and of the weighted degree kernels that consider weighted sums of kernels obtained with fixed length k-mers [ 25 ]. It is an extension of a preliminary study in which we introduced an unweighted kernel, SeqKernel, and showed its applications to protein fold recognition [ 29 ]. In this preliminary study, we have shown that the kernel values computed by SeqKernel show dependencies on sequence length, and that those dependencies can be minimized by changing the values of its parameters.…”

Algorithms to compute the Burrows-Wheeler Similarity Distribution

“…They are, however, computationally intensive to evaluate. With current databases of biological sequences at the order of hundreds of gigabytes, alternatives have been proposed both as faster, heuristic algorithms and as easier to compute similarity measures [32,33,3,25,11].…”