Decoding functional proteome information in model organisms using protein language models

Abstract

The gap between sequence information and experimental determination of function in proteins is currently unsolvable. The computational and automatic implementation of function prediction is not significantly improving function assignment, so there is a pressing need to find alternative methods, since standard approaches are not able to bridge the gap. Modern machine learning methods have been recently developed to predict function, being deep neural convolutional networks a popular choice. Protein language models have been also tested and proved reliable in curated datasets, but have not been applied yet to full collections of curated proteomes. We have tested how two different machine learning based methods perform when decoding the functional information from proteomes of selected model organisms. We found that the protein Language Models are more precise and informative across gene ontology categories for all the species, recovering functional information from transcriptomics experiments. These results indicate that these methods emerge as suitable alternatives for large scale annotation contexts and further downstream analyses, from which we propose a guide of use.

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

“…This is why faster sequence-based functional inference is a good alternative for research questions involving a high number of species (proteomes) with diverse types of proteins. In a previous study, we assessed the performance of several AI-based functional annotation methods when inferring functions in model organisms and found that protein language models provided a good 11 alternative in terms of coverage, informativeness, and precision when compared to deep learning-based and profile-based tools 17 . Here, we extended that comparison to non-model organisms across the animal kingdom and assessed how AI-based methods behaved when compared to the widely used homology-based eggNOG-mapper.…”

“…For the following analyses, we decided to proceed with the unfiltered annotated set of GOs inferred by GOPredSim-ProtT5. This decision was based on 4 factors: (1) reliability index thresholds were estimated based on a very reduced dataset strongly biased towards model organisms, and their application may impact considerably the amount of information we retrieve for non-model species, (2) reducing the reliability index values in model organisms by removing those species from the lookup dataset did not change the functional annotation, (3) information content of GO terms predicted by GOPredSim-ProtT5 is independent of their reliability index 17 , and (4) we identified genes with GO terms with low reliability index aligning well with the results of GOPredSim-ProtT5 and showing experimentally confirmed functions. Firstly, the human pseudogene LOFF_G0001716 (or AL627230.1), found to be differentially expressed in a bone mechanical response experiment 20 , was inferred to have a role in wound healing after annotation with GOPredSim-ProtT5 (GO:0042060), a term related to angiogenesis, a prerequisite for bone formation.…”

“…The best-performing method for GO prediction as highlighted by this study and our previous in model organisms 17 , GOPredSim-ProtT5, is publicly available as a web server to explore functional annotation of single sequences (https://embed.protein.properties/). However, its automatic application for whole proteomes has not been implemented yet, nor does it include the necessary steps for the necessary data preprocessing or output configuration.…”

“…23.5 million genes. To do so, we compared the performance of 3 functional annotation methods: one based on homology 2 , one based on deep learning and sequence similarity 10 , and one based on protein language models 12 (with two models tested), using the parameters previously benchmarked in controlled functional datasets and empirically tested in model organisms by our team 17 . We show that protein language model-based annotations outperformed homology-and deep learning-based ones, and discuss the implications of our findings in furthering our understanding of the functional landscape of the 'dark proteome' across the Animal Tree of Life.…”

MATEdb2, a collection of high-quality metazoan proteomes across the Animal Tree of Life to speed up phylogenomic studies

“…To help alleviate these issues, we previously published the Metazoan Assemblies from Transcriptomic Ensembles database (MATEdb) containing high-quality transcriptome assemblies for 335 arthropods and mollusks (Fernández et al, 2022). Here, we present its second version, MATEdb2, that differs from the previous one in three main aspects: (1) we have increased the taxonomic sampling to all animal phyla with high-quality data publicly available, and provide the first transcriptomic sequences for some animal taxa; (2) we include a standardized pipeline for obtaining the protein sequences from GFF genomic files instead of adding the precomputed protein files publicly available, making it easier to replicate and combine with the associated genomic sequence; (3) we provide the functional annotation of all proteins using a language-based new methodology that outperforms traditional methods (Barrios-Núñez et al, 2024). We hope that this newer version of MATEdb accelerates research on animal evolution by providing a wider taxonomic resource of high-quality proteomes across the Animal Tree of Life.…”