Illuminating the functional landscape of the dark proteome across the Animal Tree of Life through natural language processing models

Abstract

Understanding how coding genes and their functions evolve over time is a key aspect of evolutionary biology. Protein coding genes poorly understood or characterized at the functional level may be related to important evolutionary innovations, potentially leading to incomplete or inaccurate models of evolutionary change, and limiting the ability to identify conserved or lineage-specific features. Homology-based methodologies often fail to transfer functional annotations in a large fraction of the coding gene repertoire in non-model organisms. This is particularly relevant in animals, where a high number of their coding genes yield no functional annotation. Here, we leverage machine learning and natural language processing models to investigate functional annotation in the ‘dark proteome’ (defined as the unknown functional landscape’) of ca. 1,000 gene repertoires of virtually all animal phyla, totaling ca. 23.2 million coding genes. Gene ontology annotations were transferred to virtually all genes, with the model ProtT5 outperforming both homology-based and other machine learning-based models. We then explored the ‘dark proteome’ of all animal phyla revealing an enrichment in functions related to immune response, viral infection, response to stimuli, development, or signaling, among others. Furthermore, we provide an open-access pipeline - FANTASIA - to implement and benchmark these methodologies in any dataset. Our results uncover the putative functions of poorly understood protein-coding genes across the Animal Tree of Life and contribute to a more comprehensive understanding of the molecular basis of animal evolution.

Decoding functional proteome information in model organisms using protein language models

“…These results may reflect the focus of different scientific communities on particular genes or gene families, with different groups of genes having been studied more intensely, while others remain relatively overlooked ( 2 ). Overall, the LMs (GOPredSim-T5 and GOPredSim-SeqVec) provided a larger and more consistent coverage of the predicted functions for proteins across organisms, regardless of their annotation stage, making them promising tools to address functional studies in non-model organisms as already indicated in the animal phyla ( 40 ).…”

“…Our study demonstrates that LMs may be particularly useful to examine poorly annotated genomes, as exemplified here in C. elegans , both at the genome level and in terms of the functional retrieval of DEGs. We also performed large-scale testing of LMs, in particular applying the GOPredSim-T5 annotation of 2.4 million to genes in various non-model organisms in the animal phyla, where we found interesting biological insights [see ( 40 )].…”

Decoding functional proteome information in model organisms using protein language models

“…Our study demonstrates that LMs may be particularly useful to examine poorly annotated genomes, as exemplified here in C. elegans, both at the genome level and in terms of the functional retrieval of DEGs. We also performed largescale testing of LMs, in particular applying the GOPredSim-T5 annotation of 2.4 million to genes in various non-model organisms in the animal Phyla, where we found interesting biological insights (see 40).…”

A punctuated burst of massive genomic rearrangements by chromosome shattering and the origin of non-marine annelids

“…Gene repertoire evolutionary dynamics across nodes were estimated with pyHam 83 and curated with custom scripts following the workflow available at the GitHub repository (https://github.com/MetazoaPhylogenomicsLab/Vargas-Chavez_et_al_2024_Chromosome_shat tering_clitellate_origins). Longest isoforms in protein mode were functionally annotated with both homology-based methods (eggNOG-mapper v2 84 ) and natural language processing ones (FANTASIA 85 ). Using the Gene Ontology (GO) annotations inferred with FANTASIA, GO enrichment analysis of genes loss in each internode were calculated using the Fisher test and the elim algorithm as implemented in the topGO package 86 .…”