MUSI: an integrated system for identifying multiple specificity from very large peptide or nucleic acid data sets

Abstract

Peptide recognition domains and transcription factors play crucial roles in cellular signaling. They bind linear stretches of amino acids or nucleotides, respectively, with high specificity. Experimental techniques that assess the binding specificity of these domains, such as microarrays or phage display, can retrieve thousands of distinct ligands, providing detailed insight into binding specificity. In particular, the advent of next-generation sequencing has recently increased the throughput of such methods by several orders of magnitude. These advances have helped reveal the presence of distinct binding specificity classes that co-exist within a set of ligands interacting with the same target. Here, we introduce a software system called MUSI that can rapidly analyze very large data sets of binding sequences to determine the relevant binding specificity patterns. Our pipeline provides two major advances. First, it can detect previously unrecognized multiple specificity patterns in any data set. Second, it offers integrated processing of very large data sets from next-generation sequencing machines. The results are visualized as multiple sequence logos describing the different binding preferences of the protein under investigation. We demonstrate the performance of MUSI by analyzing recent phage display data for human SH3 domains as well as microarray data for mouse transcription factors.

Unsupervised HLA Peptidome Deconvolution Improves Ligand Prediction Accuracy and Predicts Cooperative Effects in Peptide–HLA Interactions

“…As such, we anticipated that naturally processed HLA ligands that come from up to six different HLA molecules could be ideally modeled with a mixture of PWMs (24)(25)(26). In this probabilistic framework, a total log-likelihood function is defined as follows:…”

“…M k is the PWM representing the kth motif (with M k X n l ;l the PWM entry corresponding to the lth residue in peptide X n ), and w k is the contribution of each motif in the total likelihood. Priors were included, as described previously (24,25).…”

“…As such, we postulated that HLA peptidome data consist of a superposition of different motifs that represent the up-to-six class I alleles expressed in any given sample. From a computational point of view, this kind of superposition can be deconvoluted using the machine-learning framework of mixture models (25,26). In this work, we implemented such an algorithm for the special case of unsupervised deconvolution of HLA class I peptidome (see Materials and Methods).…”

“…Therefore, there is a growing interest in using these data toward a better understanding of (neo-)Ag presentation and peptide-HLA interactions. In this study, we use a novel unsupervised computational approach to identify HLA-binding motifs based on the machine-learning framework of mixture models (24)(25)(26). Our work shows that HLA peptidome data can be incorporated to train predictors with improved accuracy compared with training only with publicly available binding data.…”

Deep sequencing analysis of phage libraries using Illumina platform

“…Remarkably, a rare amino acid W appeared as a consensus amino acid in many sub-sequences, and it was present as the N-terminal amino acid in 50 out of 150 peptides. Our simple clustering analysis could be potentially replaced by more advanced software packages, such as MUltiple Specificity Identifier (MUSI) [31], which was designed to identify distinct families of consensus sequence motifs within deep sequencing data. The analysis could potentially identify conserved peptide motifs emerging as the results of growth-induced selection.…”