Search Results for « Protein sequences »

Abstract This paper presents a comprehensive analysis of MMseqs2 clusters and traditional machine learning (ML) clustering algorithms, including KMeans and Hierarchical clusterings, in terms of protein sequences. The analyses are validated experimentally. The cluster analyses have been performed in the A stral Compendium protein sequences dataset hosted in the SCOPe database. The dataset is embedded using two pre-trained transformer models using Evolutionary Scale Modeling (ESM) to perform KMeans and Hierarchical clustering algorithms. Afterward, those four clusters are compared with MMseqs2/Linclust and MMseqs2/easy-cluster methods. After performing the experiment, MMseqs2/Linclust and MMseqs2/easy-cluster outperform traditional machine learning cluster algorithms by a considerable margin. This analysis demonstrates the superiority of the MMseqs2 clustering techniques over conventional machine learning clustering algorithms. The source code of the experiment is publicly available and readily accessible through: https://github.com/mrzResearchArena/protein-clustering.

Abstract Human mitochondrial DNA (mtDNA) contains 13 protein-coding subunits of the oxidative phosphorylation pathway, 22 tRNA and two rRNA genes. However, accumulating evidence suggests that mtDNA encodes additional overlapping genetic elements, including mitochondrial-derived peptides (MDPs) and alternative reading frames. Here, we assessed signatures of selection across 66 328 human mtDNAs and studied the potential impact of disease-causing mutation within these mtDNA overlapping sequences. By employing frame-specific dN/dS analysis for the overlapping reading frames, and codon position-specific diversity calculations we found that SHLP6 and SHLP3 (within 16S rRNA) display significant signatures of purifying selection. Mutational asymmetry analysis revealed purifying selection in both frames and strands of GAU/COX1, while other alternative reading frames show asymmetric patterns, supporting negative selection primarily on the encompassing canonical gene. Analysis of mito-ribosome profiling (HEK293 cells) revealed translation initiation signatures only for SHLP6 and ALTND4, providing functional support for their translation in HEK293 cells. Analysis of disease-causing mutations revealed that several such mutations have predicted deleterious effects on both canonical and alternative sequences, though canonical genes tend to be more frequently affected. Taken together, we provide evolutionary and functional evidence supporting biological relevance of certain MDPs and underline the need to re-evaluate the functionality of mutations in such sequences. This article is part of the theme issue ‘Evolutionary genetics of mitochondria: on diverse and common evolutionary constraints across eukarya’.

Abstract— Amino acid encoding genes contain character state information that may be useful for phylogenetic analysis on at least two levels. The nucleotide sequence and the translated amino acid sequences have both been employed separately as character states for cladistic studies of various taxa, including studies of the genealogy of genes in multigene families. In essence, amino acid sequences and nucleic acid sequences are two different ways of character coding the information in a gene. Silent positions in the nucleotide sequence (first or third positions in codons that can accrue change without changing the identity of the amino acid that the triplet codes for) may accrue change relatively rapidly and become saturated, losing the pattern of historical divergence. On the other hand, non‐silent nucleotide alterations and their accompanying amino acid changes may evolve too slowly to reveal relationships among closely related taxa. In general, the dynamics of sequence change in silent and non‐silent positions in protein coding genes result in homoplasy and lack of resolution, respectively. We suggest that the combination of nucleic acid and the translated amino acid coded character states into the same data matrix for phylogenetic analysis addresses some of the problems caused by the rapid change of silent nucleotide positions and overall slow rate of change of non‐silent nucleotide positions and slowly changing amino acid positions. One major theoretical problem with this approach is the apparent non‐independence of the two sources of characters. However, there are at least three possible outcomes when comparing protein coding nucleic acid sequences with their translated amino acids in a phylogenetic context on a codon by codon basis. First, the two character sets for a codon may be entirely congruent with respect to the information they convey about the relationships of a certain set of taxa. Second, one character set may display no information concerning a phylogenetic hypothesis while the other character set may impart information to a hypothesis. These two possibilities are cases of non‐independence, however, we argue that congruence in such cases can be thought of as increasing the weight of the particular phylogenetic hypothesis that is supported by those characters. In the third case, the two sources of character information for a particular codon may be entirely incongruent with respect to phylogenetic hypotheses concerning the taxa examined. In this last case the two character sets are independent in that information from neither can predict the character states of the other. Examples of these possibilities are discussed and the general applicability of combining these two sources of information for protein coding genes is presented using sequences from the homeobox region of 46 homeobox genes from Drosophila melanogaster to develop a hypothesis of genealogical relationship of these genes in this large multigene family.

One of the major challenges and opportunities for protein function prediction is how to integrate diverse function-relevant biological data, such as protein sequences, protein function annotations, spatial gene-gene interactions, and protein-protein interactions, to improve the accuracy of protein function prediction. Given a target protein sequence, our method generated an initial list of Gene Ontology (GO) function terms based on the known function annotations of the homologous proteins found by PSI-BLAST, the set of GO terms were then expanded according to the association rules between GO terms learned by mining the SWISS-Prot database, and finally the GO terms were further augmented by the function annotations of the neighboring proteins of the target protein found in protein-protein interaction networks of 23 species and the novel spatial gene-gene interaction networks of the human genome constructed from the Hi-C chromosomal conformation data of the genome. The individual probabilistic scores of these GO terms derived from these different sources were statistically combined into final scores to rank them. The GO terms ranked at top were then used as the function prediction of the target protein.

Abstract Understanding how proteins interact and how mutations affect these interactions is crucial for unraveling the complexities of biological systems and their evolution. Mutations can significantly alter protein behavior, impacting stability, interactions, and activity, thereby affecting cellular functions and influencing disease development and treatment effectiveness. Experimental methods for examining protein interactions are often slow and costly, highlighting the need for efficient computational strategies. We present MuLAN, a groundbreaking deep learning method that leverages light attention networks and the power of pre-trained protein language models to infer protein interactions, predict binding affinity changes, and reconstruct mutational landscapes for proteins involved in binary interactions, starting from mutational changes and directly using sequence data only. Unlike previous methods that depend heavily on structural information, MuLAN’s sequence-based approach offers faster and more accessible predictions. This innovation allows for variations in predictions based on specific partners, opening new possibilities for understanding protein behavior through their sequences. The potential implications for disease research and drug development mark a significant step forward in the computational analysis of protein interactions.

Abstract Public databases of protein sequences, such as the National Center for Biotechnology Information (NCBI) Protein repository and UniProt, contain millions of proteins identified in samples from specific species but named as uncharacterised, hypothetical or unclassified due to a lack of information about their function. It has been demonstrated previously that many such sequences show high similarity to genes from RNA viruses, either due to viral infection of the original sample, contamination or endogenous viral elements (EVEs) integrated into the genome of the sample species. Many proteins from RNA virus discovery research are also deposited into these repositories but, for various reasons, can only be labelled as uncharacterised and classified taxonomically at a superkingdom or realm level. Sequences from protein repositories not labelled specifically as being derived from the RNA viral RNA dependent RNA polymerase (RdRp) protein are often used as negative controls when looking to identify viral RdRp sequences, so the presence of unlabelled viruses amongst these datasets is problematic. In this study, we screened uncharacterised proteins from two large public protein repositories - NCBI Protein and UniProt - to identify sequences likely to be derived from RNA viral RdRp. 3,560 such sequences were identified, many derived from EVEs. Many previously unknown EVEs were identified and led to characterisation of additional, related sequences. For example, a group of orbivirus-like viruses infecting nematodes was uncovered which appears to have both ancient endogenous and circulating exogenous members. Many recent integrations of mito-like viruses into plant genomes were identified, indicative of current or recent RNA viral activity. In several taxonomic groups, the first example of an EVE, and in some cases the first example of any RNA virus, was uncovered. The large number of EVEs uncovered by this relatively small-scale search suggests that only a fraction of the true diversity of EVEs is currently known. We also explore uncharacterised proteins further by providing provisional taxonomic annotations for RdRps which are currently only listed as members of the Riboviria realm. A number of sequences are identified which are indistinguishable from known, pathogenic viruses but are labelled as bacteria, seemingly as a result of mislabelling or contamination. Sequences which are not RNA viral but show some similarity to RdRp are also analysed, as a potential source of false positives in virus discovery research. Finally, recommendations are made for generating useful negative controls.