In the twilight zone of protein sequence homology: do protein language models learn protein structure?

Motivation: Protein language models based on the transformer architecture are increasingly improving performance on protein prediction tasks, including secondary structure, subcellular localization, and more. Despite being trained only on protein sequences, protein language models appear to implicitly learn protein structure. This paper investigates whether sequence representations learned by protein language models encode structural information and to what extent. Results: We address this by evaluating protein language models on remote homology prediction, where identifying remote homologs from sequence information alone requires structural knowledge, especially in the "twilight zone" of very low sequence identity. Through rigorous testing at progressively lower sequence identities, we profile the performance of protein language models ranging from millions to billions of parameters in a zero-shot setting. Our findings indicate that while transformer-based protein language models outperform traditional sequence alignment methods, they still struggle in the twilight zone. This suggests that current protein language models have not sufficiently learned protein structure to address remote homology prediction when sequence signals are weak. Availability and implementation: We believe this opens the way for further research both on remote homology prediction and on the broader goal of learning sequence- and structure-rich representations of protein molecules. All code, data, and models are made publicly available.

Assembling bacterial puzzles: piecing together functions into microbial pathways.

Functional metagenomics enables the study of unexplored bacterial diversity, gene families, and pathways essential to microbial communities. However, discovering biological insights with these data is impeded by the scarcity of quality annotations. Here, we use a co-occurrence-based analysis of predicted microbial protein functions to uncover pathways in genomic and metagenomic biological systems. Our approach, based on phylogenetic profiles, improves the identification of functional relationships, or participation in the same biochemical pathway, between enzymes over a comparable homology-based approach. We optimized the design of our profiles to identify potential pathways using minimal data, clustered functionally related enzyme pairs into multi-enzymatic pathways, and evaluated our predictions against reference pathways in the KEGG database. We then demonstrated a novel extension of this approach to predict inter-bacterial protein interactions amongst members of a marine microbiome. Most significantly, we show our method predicts emergent biochemical pathways between known and unknown functions. Thus, our work establishes a basis for identifying the potential functional capacities of the entire metagenome, capturing previously unknown and abstract functions into discrete putative pathways.

Variant Effect Prediction in the Age of Machine Learning.

Over the years, many computational methods have been created for the analysis of the impact of single amino acid substitutions resulting from single-nucleotide variants in genome coding regions. Historically, all methods have been supervised and thus limited by the inadequate sizes of experimentally curated data sets and by the lack of a standardized definition of variant effect. The emergence of unsupervised, deep learning (DL)-based methods raised an important question: Can machines learn the language of life from the unannotated protein sequence data well enough to identify significant errors in the protein "sentences"? Our analysis suggests that some unsupervised methods perform as well or better than existing supervised methods. Unsupervised methods are also faster and can, thus, be useful in large-scale variant evaluations. For all other methods, however, their performance varies by both evaluation metrics and by the type of variant effect being predicted. We also note that the evaluation of method performance is still lacking on less-studied, nonhuman proteins where unsupervised methods hold the most promise.

Session Introduction: Precision Medicine: Innovative methods for advanced understanding of molecular underpinnings of disease.

Precision medicine, also often referred to as personalized medicine, targets the development of treatments and preventative measures specific to the individual's genomic signatures, lifestyle, and environmental conditions. The series of Precision Medicine sessions in PSB has continuously highlighted the advances in this field. Our 2024 collection of manuscripts showcases algorithmic advances that integrate data from distinct modalities and introduce innovative approaches to extract new, medically relevant information from existing data. These evolving technology and analytical methods promise to bring closer the goals of precision medicine to improve health and increase lifespan.

Learning from the unknown: exploring the range of bacterial functionality.

Determining the repertoire of a microbe's molecular functions is a central question in microbial biology. Modern techniques achieve this goal by comparing microbial genetic material against reference databases of functionally annotated genes/proteins or known taxonomic markers such as 16S rRNA. Here, we describe a novel approach to exploring bacterial functional repertoires without reference databases. Our Fusion scheme establishes functional relationships between bacteria and assigns organisms to Fusion-taxa that differ from otherwise defined taxonomic clades. Three key findings of our work stand out. First, bacterial functional comparisons outperform marker genes in assigning taxonomic clades. Fusion profiles are also better for this task than other functional annotation schemes. Second, Fusion-taxa are robust to addition of novel organisms and are, arguably, able to capture the environment-driven bacterial diversity. Finally, our alignment-free nucleic acid-based Siamese Neural Network model, created using Fusion functions, enables finding shared functionality of very distant, possibly structurally different, microbial homologs. Our work can thus help annotate functional repertoires of bacterial organisms and further guide our understanding of microbial communities.

A New Paradigm for Pandemic Preparedness.

Purpose of Review: Preparing for pandemics requires a degree of interdisciplinary work that is challenging under the current paradigm. This review summarizes the challenges faced by the field of pandemic science and proposes how to address them. Recent Findings: The structure of current siloed systems of research organizations hinders effective interdisciplinary pandemic research. Moreover, effective pandemic preparedness requires stakeholders in public policy and health to interact and integrate new findings rapidly, relying on a robust, responsive, and productive research domain. Neither of these requirements are well supported under the current system. Summary: We propose a new paradigm for pandemic preparedness wherein interdisciplinary research and close collaboration with public policy and health practitioners can improve our ability to prevent, detect, and treat pandemics through tighter integration among domains, rapid and accurate integration, and translation of science to public policy, outreach and education, and improved venues and incentives for sustainable and robust interdisciplinary work.

Inferring Potential Cancer Driving Synonymous Variants.

Synonymous single nucleotide variants (sSNVs) are often considered functionally silent, but a few cases of cancer-causing sSNVs have been reported. From available databases, we collected four categories of sSNVs: germline, somatic in normal tissues, somatic in cancerous tissues, and putative cancer drivers. We found that screening sSNVs for recurrence among patients, conservation of the affected genomic position, and synVep prediction (synVep is a machine learning-based sSNV effect predictor) recovers cancer driver variants (termed proposed drivers) and previously unknown putative cancer genes. Of the 2.9 million somatic sSNVs found in the COSMIC database, we identified 2111 proposed cancer driver sSNVs. Of these, 326 sSNVs could be further tagged for possible RNA splicing effects, RNA structural changes, and affected RBP motifs. This list of proposed cancer driver sSNVs provides computational guidance in prioritizing the experimental evaluation of synonymous mutations found in cancers. Furthermore, our list of novel potential cancer genes, galvanized by synonymous mutations, may highlight yet unexplored cancer mechanisms.

Predicting embryonic aneuploidy rate in IVF patients using whole-exome sequencing.

Infertility is a major reproductive health issue that affects about 12% of women of reproductive age in the United States. Aneuploidy in eggs accounts for a significant proportion of early miscarriage and in vitro fertilization failure. Recent studies have shown that genetic variants in several genes affect chromosome segregation fidelity and predispose women to a higher incidence of egg aneuploidy. However, the exact genetic causes of aneuploid egg production remain unclear, making it difficult to diagnose infertility based on individual genetic variants in mother's genome. In this study, we evaluated machine learning-based classifiers for predicting the embryonic aneuploidy risk in female IVF patients using whole-exome sequencing data. Using two exome datasets, we obtained an area under the receiver operating curve of 0.77 and 0.68, respectively. High precision could be traded off for high specificity in classifying patients by selecting different prediction score cutoffs. For example, a strict prediction score cutoff of 0.7 identified 29% of patients as high-risk with 94% precision. In addition, we identified MCM5, FGGY, and DDX60L as potential aneuploidy risk genes that contribute the most to the predictive power of the model. These candidate genes and their molecular interaction partners are enriched for meiotic-related gene ontology categories and pathways, such as microtubule organizing center and DNA recombination. In summary, we demonstrate that sequencing data can be mined to predict patients' aneuploidy risk thus improving clinical diagnosis. The candidate genes and pathways we identified are promising targets for future aneuploidy studies.

Quantifying structural relationships of metal-binding sites suggests origins of biological electron transfer.

Biological redox reactions drive planetary biogeochemical cycles. Using a novel, structure-guided sequence analysis of proteins, we explored the patterns of evolution of enzymes responsible for these reactions. Our analysis reveals that the folds that bind transition metal­containing ligands have similar structural geometry and amino acid sequences across the full diversity of proteins. Similarity across folds reflects the availability of key transition metals over geological time and strongly suggests that transition metal­ligand binding had a small number of common peptide origins. We observe that structures central to our similarity network come primarily from oxidoreductases, suggesting that ancestral peptides may have also facilitated electron transfer reactions. Last, our results reveal that the earliest biologically functional peptides were likely available before the assembly of fully functional protein domains over 3.8 billion years ago.Thus, life is a special, very complex form of motion of matter, but this form did not always exist, and it is not separated from inorganic nature by an impassable abyss; rather, it arose from inorganic nature as a new property in the process of evolution of the world. We must study the history of this evolution if we want to solve the problem of the origin of life. [A. I. Oparin (1)]

Decoding the effects of synonymous variants.

Synonymous single nucleotide variants (sSNVs) are common in the human genome but are often overlooked. However, sSNVs can have significant biological impact and may lead to disease. Existing computational methods for evaluating the effect of sSNVs suffer from the lack of gold-standard training/evaluation data and exhibit over-reliance on sequence conservation signals. We developed synVep (synonymous Variant effect predictor), a machine learning-based method that overcomes both of these limitations. Our training data was a combination of variants reported by gnomAD (observed) and those unreported, but possible in the human genome (generated). We used positive-unlabeled learning to purify the generated variant set of any likely unobservable variants. We then trained two sequential extreme gradient boosting models to identify subsets of the remaining variants putatively enriched and depleted in effect. Our method attained 90% precision/recall on a previously unseen set of variants. Furthermore, although synVep does not explicitly use conservation, its scores correlated with evolutionary distances between orthologs in cross-species variation analysis. synVep was also able to differentiate pathogenic vs. benign variants, as well as splice-site disrupting variants (SDV) vs. non-SDVs. Thus, synVep provides an important improvement in annotation of sSNVs, allowing users to focus on variants that most likely harbor effects.

PredictProtein - Predicting Protein Structure and Function for 29 Years.

Since 1992 PredictProtein (https://predictprotein.org) is a one-stop online resource for protein sequence analysis with its main site hosted at the Luxembourg Centre for Systems Biomedicine (LCSB) and queried monthly by over 3,000 users in 2020. PredictProtein was the first Internet server for protein predictions. It pioneered combining evolutionary information and machine learning. Given a protein sequence as input, the server outputs multiple sequence alignments, predictions of protein structure in 1D and 2D (secondary structure, solvent accessibility, transmembrane segments, disordered regions, protein flexibility, and disulfide bridges) and predictions of protein function (functional effects of sequence variation or point mutations, Gene Ontology (GO) terms, subcellular localization, and protein-, RNA-, and DNA binding). PredictProtein's infrastructure has moved to the LCSB increasing throughput; the use of MMseqs2 sequence search reduced runtime five-fold (apparently without lowering performance of prediction methods); user interface elements improved usability, and new prediction methods were added. PredictProtein recently included predictions from deep learning embeddings (GO and secondary structure) and a method for the prediction of proteins and residues binding DNA, RNA, or other proteins. PredictProtein.org aspires to provide reliable predictions to computational and experimental biologists alike. All scripts and methods are freely available for offline execution in high-throughput settings.

Low Diversity of Human Variation Despite Mostly Mild Functional Impact of De Novo Variants.

Non-synonymous Single Nucleotide Variants (nsSNVs), resulting in single amino acid variants (SAVs), are important drivers of evolutionary adaptation across the tree of life. Humans carry on average over 10,000 SAVs per individual genome, many of which likely have little to no impact on the function of the protein they affect. Experimental evidence for protein function changes as a result of SAVs remain sparse - a situation that can be somewhat alleviated by predicting their impact using computational methods. Here, we used SNAP to examine both observed and in silico generated human variation in a set of 1,265 proteins that are consistently found across a number of diverse species. The number of SAVs that are predicted to have any functional effect on these proteins is smaller than expected, suggesting sequence/function optimization over evolutionary timescales. Additionally, we find that only a few of the yet-unobserved SAVs could drastically change the function of these proteins, while nearly a quarter would have only a mild functional effect. We observed that variants common in the human population localized to less conserved protein positions and carried mild to moderate functional effects more frequently than rare variants. As expected, rare variants carried severe effects more frequently than common variants. In line with current assumptions, we demonstrated that the change of the human reference sequence amino acid to the reference of another species (a cross-species variant) is unlikely to significantly impact protein function. However, we also observed that many cross-species variants may be weakly non-neutral for the purposes of quick adaptation to environmental changes, but may not be identified as such by current state-of-the-art methodology.

Impact of vitamin A transport and storage on intestinal retinoid homeostasis and functions.

Lecithin:retinol acyltransferase and retinol-binding protein enable vitamin A (VA) storage and transport, respectively, maintaining tissue homeostasis of retinoids (VA derivatives). The precarious VA status of the lecithin:retinol acyltransferase-deficient (Lrat-/-) retinol-binding protein-deficient (Rbp-/-) mice rapidly deteriorates upon dietary VA restriction, leading to signs of severe vitamin A deficiency (VAD). As retinoids impact gut morphology and functions, VAD is often linked to intestinal pathological conditions and microbial dysbiosis. Thus, we investigated the contribution of VA storage and transport to intestinal retinoid homeostasis and functionalities. We showed the occurrence of intestinal VAD in Lrat-/-Rbp-/- mice, demonstrating the critical role of both pathways in preserving gut retinoid homeostasis. Moreover, in the mutant colon, VAD resulted in a compromised intestinal barrier as manifested by reduced mucins and antimicrobial defense, leaky gut, increased inflammation and oxidative stress, and altered mucosal immunocytokine profiles. These perturbations were accompanied by fecal dysbiosis, revealing that the VA status (sufficient vs. deficient), rather than the amount of dietary VA per se, is likely a major initial discriminant of the intestinal microbiome. Our data also pointed to a specific fecal taxonomic profile and distinct microbial functionalities associated with VAD. Overall, our findings revealed the suitability of the Lrat-/-Rbp-/- mice as a model to study intestinal dysfunctions and dysbiosis promoted by changes in tissue retinoid homeostasis induced by the host VA status and/or intake.

Snow microbiome functional analyses reveal novel aspects of microbial metabolism of complex organic compounds.

Microbes active in extreme cold are not as well explored as those of other extreme environments. Studies have revealed a substantial microbial diversity and identified cold-specific microbiome molecular functions. We analyzed the metagenomes and metatranscriptomes of 20 snow samples collected in early and late spring in Svalbard, Norway using mi-faser, our read-based computational microbiome function annotation tool. Our results reveal a more diverse microbiome functional capacity and activity in the early- vs. late-spring samples. We also find that functional dissimilarity between the same-sample metagenomes and metatranscriptomes is significantly higher in early than late spring samples. These findings suggest that early spring samples may contain a larger fraction of DNA of dormant (or dead) organisms, while late spring samples reflect a new, metabolically active community. We further show that the abundance of sequencing reads mapping to the fatty acid synthesis-related microbial pathways in late spring metagenomes and metatranscriptomes is significantly correlated with the organic acid levels measured in these samples. Similarly, the organic acid levels correlate with the pathway read abundances of geraniol degradation and inversely correlate with those of styrene degradation, suggesting a possible nutrient change. Our study thus highlights the activity of microbial degradation pathways of complex organic compounds previously unreported at low temperatures.

Amino acid encoding for deep learning applications.

BACKGROUND: The number of applications of deep learning algorithms in bioinformatics is increasing as they usually achieve superior performance over classical approaches, especially, when bigger training datasets are available. In deep learning applications, discrete data, e.g. words or n-grams in language, or amino acids or nucleotides in bioinformatics, are generally represented as a continuous vector through an embedding matrix. Recently, learning this embedding matrix directly from the data as part of the continuous iteration of the model to optimize the target prediction - a process called 'end-to-end learning' - has led to state-of-the-art results in many fields. Although usage of embeddings is well described in the bioinformatics literature, the potential of end-to-end learning for single amino acids, as compared to more classical manually-curated encoding strategies, has not been systematically addressed. To this end, we compared classical encoding matrices, namely one-hot, VHSE8 and BLOSUM62, to end-to-end learning of amino acid embeddings for two different prediction tasks using three widely used architectures, namely recurrent neural networks (RNN), convolutional neural networks (CNN), and the hybrid CNN-RNN. RESULTS: By using different deep learning architectures, we show that end-to-end learning is on par with classical encodings for embeddings of the same dimension even when limited training data is available, and might allow for a reduction in the embedding dimension without performance loss, which is critical when deploying the models to devices with limited computational capacities. We found that the embedding dimension is a major factor in controlling the model performance. Surprisingly, we observed that deep learning models are capable of learning from random vectors of appropriate dimension. CONCLUSION: Our study shows that end-to-end learning is a flexible and powerful method for amino acid encoding. Further, due to the flexibility of deep learning systems, amino acid encoding schemes should be benchmarked against random vectors of the same dimension to disentangle the information content provided by the encoding scheme from the distinguishability effect provided by the scheme.

Fingerprinting cities: differentiating subway microbiome functionality.

BACKGROUND: Accumulating evidence suggests that the human microbiome impacts individual and public health. City subway systems are human-dense environments, where passengers often exchange microbes. The MetaSUB project participants collected samples from subway surfaces in different cities and performed metagenomic sequencing. Previous studies focused on taxonomic composition of these microbiomes and no explicit functional analysis had been done till now. RESULTS: As a part of the 2018 CAMDA challenge, we functionally profiled the available ~ 400 subway metagenomes and built predictor for city origin. In cross-validation, our model reached 81% accuracy when only the top-ranked city assignment was considered and 95% accuracy if the second city was taken into account as well. Notably, this performance was only achievable if the similarity of distribution of cities in the training and testing sets was similar. To assure that our methods are applicable without such biased assumptions we balanced our training data to account for all represented cities equally well. After balancing, the performance of our method was slightly lower (76/94%, respectively, for one or two top ranked cities), but still consistently high. Here we attained an added benefit of independence of training set city representation. In testing, our unbalanced model thus reached (an over-estimated) performance of 90/97%, while our balanced model was at a more reliable 63/90% accuracy. While, by definition of our model, we were not able to predict the microbiome origins previously unseen, our balanced model correctly judged them to be NOT-from-training-cities over 80% of the time. Our function-based outlook on microbiomes also allowed us to note similarities between both regionally close and far-away cities. Curiously, we identified the depletion in mycobacterial functions as a signature of cities in New Zealand, while photosynthesis related functions fingerprinted New York, Porto and Tokyo. CONCLUSIONS: We demonstrated the power of our high-speed function annotation method, mi-faser, by analysing ~ 400 shotgun metagenomes in 2 days, with the results recapitulating functional signals of different city subway microbiomes. We also showed the importance of balanced data in avoiding over-estimated performance. Our results revealed similarities between both geographically close (Ofa and Ilorin) and distant (Boston and Porto, Lisbon and New York) city subway microbiomes. The photosynthesis related functional signatures of NYC were previously unseen in taxonomy studies, highlighting the strength of functional analysis.