Biology subjects

Bromberg, Y.

Publications and source records attributed to Bromberg, Y..

18 recordsLinked to original sources

bioRxiv2026

In Other Words: Learning When to Trust Generated Protein Function Descriptions

Generative protein-to-text models write function descriptions in free-form English but give no indication of which to believe. Indeed, free text is hard to evaluate. We introduce a large language model judge of equivalence of predicted vs. ''true'' function descriptions, validated against blinded expert curation. We also present ProtTale, our sequence-to-text model, that learns from the judge's verdicts to estimate its own reliability, pJay. On unseen Swiss-Prot enzymes, ProtTale and three other leading methods tie on accuracy but are right about different proteins. Here, ProtTale's unique correct predictions outnumber all other methods' combined. Scored from sequence alone, pJay is a plug-in filter, raising all models' performance. On short proteins, it raises precision for every tested method nearly twofold. We argue that generative annotation becomes useful not by being right more often, but by knowing when it can be trusted---a standard that every generative model in biology should meet.

bioinformatics↗

bioRxiv2026

REFinder: Mining new enzymes from metagenomes

Microbial metagenomes encode vast catalytic diversity, but recovering enzymes of novel, yet-undescribed functionality typically requires whole-community assembly plus a means of identifying active proteins ab initio. The first step of the task is compute-intensive and misses low-abundance sequences. The second is complicated by our inability to predict novel functions. REBEAN, our DNA language model, sidesteps the latter by assigning each sequencing read a high-level Enzyme Commission (EC) class or a non-enzyme label without alignment. Here we build REFinder, a pipeline that addresses both steps by routing REBEAN-annotated reads to assemble only the putative enzymatic reads that have identified catalytic signatures. In our evaluation of 50 microbiome metagenomes, REFinder identified 1.2 to 2.1 fold more enzymes than could be recovered via homology-based annotation of the proteins from the corresponding full assemblies. Moreover, it was as much as 6.4-fold cheaper computationally than full assembly. Across all samples, REFinder identified at least three fourths and as many as 90% of the homology-accessible enzymes identified via full assembly of the complete metagenomes. Notably, a fraction of these, 22% to 45% per EC class, carried no similarity to Swiss-Prot proteins, i.e. a set of enzymes homology cannot annotate. For roughly two fifths of the over thirteen thousand such novel oxidoreductases from two saliva samples, ESMFold predicted structures aligned with TM-score[≥]0.7 to a characterized enzyme structure in the PDB - a substantial structural similarity without sequence homology. These results illustrate that targeted, alignment-free assembly turns even well-mapped microbiomes into a source of thousands of previously invisible but credible novel enzymes, raising our expectations for exploration of environmental microbiomes.

bioinformatics↗

bioRxiv2026

Fusion2AI: functional re-annotation of the bacterial transparent matter proteins

Determining the biological function of bacterial proteins remains a central challenge in microbiology. We previously developed Fusion, a reference-free scheme that clustered the tens of millions of known, unique, bacterial proteins into about four hundred thousand groups of functionally similar proteins, i.e. functions. For any newly identified proteins, Fusion function assignment could be performed using function-aware sequence alignment to each function's representatives. However, Fusion functions lack human-readable descriptions of their functionality. Here, we leverage large language models (LLMs) to synthesize scattered, partial annotations of functions of individual proteins into coherent English-language descriptions for ~100,000 clusters, i.e. a quarter of all Fusion functions. From these descriptions, we extract Gene Ontology (GO) terms and Enzyme Commission (EC) numbers. We validate these against annotations inferred from UniProtKB/Swiss-Prot. These annotations immediately expand functional coverage for both genome and microbiome -level analyses. For organisms in FusionDB, we annotate 50% more proteins than alignment to Swiss-Prot alone. For other bacteria, the gain is smaller but still reaches 30% for some organisms. For microbiomes, annotation coverage increases nearly ten-fold. Our work thus enables the analysis of microbial "transparent matter," i.e. functions already described in earlier research that have not yet made it into usable annotations. Our annotated Fusion function database is freely available at https://services.bromberglab.org/fusion2ai and as a reference DB for metagenome annotation https://services.bromberglab.org/mifaser.

bioinformatics↗

bioRxiv2026

16S rRNA k-mer composition encodes microbial functional potential

16S rRNA amplicon sequencing is widely used for microbiome profiling, but most methods rely on reference databases of characterized organisms, limiting its accuracy in function prediction for underrepresented environments. We discovered that 16S rRNA k-mer composition carries substantial functional signal: (i) whole-genome k-mer profiles predict genome-encoded functions, and (ii) 16S rRNA k-mer profiles reflect their source genomes composition. Building on these relationships, we developed embeRNA, a neural network framework that predicts functions directly from 16S rRNA k-mer embeddings without requiring taxonomy assignment or phylogenetic placement. embeRNA outputs per-function probability scores, enabling users to tune decision thresholds to balance precision and recall or account for community novelty. In a stringent "novel microbes" benchmark - where all test sequences shared <97% identity with training data - embeRNA outperformed reference-based methods, particularly for hard-to-label functions. Applied to soil metagenomes with paired 16S and whole metagenome shotgun sequencing (WMS) data, embeRNA recovered most WMS-inferred functions and produced abundance profiles strongly correlated with WMS results, attaining better performance than a reference-based approach. Our findings demonstrate that 16S rRNA directly captures functional potential, and 16S amplicon sequencing data can complement WMS-based inference to broaden functional characterization of microbiomes, especially in understudied environments.

bioinformatics↗

bioRxiv2025

Whole Metagenome Sequencing: not Deep Enough for Complete Microbial Function Recovery

BackgroundWhole metagenome shotgun sequencing (WMS) is widely used to profile microbial function. However, technical variability in sequencing and analysis often obscures true biological patterns. Large-scale studies are particularly susceptible to batch effects, such as differences in sequencing depth and platform and annotation strategies, as well as sample-to-flow-cell assignments. However, the relative effects of these factors on functional inference in such studies have yet to be systematically evaluated. We analyzed oral-rinse WMS data from a study cohort including 671 Nigerian youths aged 9-18, sequenced on two Illumina platforms. Microbial molecular functionality encoded in these data were annotated using the mi-faser/Fusion pipeline, to capture the broad functional repertoire, and HUMAnN 3/EC numbers pipeline to characterize curated enzymatic activities. We then quantified how technical factors and batch effects shaped the recovery of microbial functionality. ResultsThree findings of our work were most salient. First, we observed that the choice of annotation strategy traded off between breadth and specificity of functional coverage. Second, we found that low-prevalence functions were disproportionately lost at shallow sequencing depths, indicating that in e.g. case-control studies with few representatives of the minor class, sequencing depth could critically impact study resolution. Finally, using our newly developed model relating sequencing depth to functional recovery, we demonstrated that increasing sequencing depth does not directly or proportionally improve functional recall. That is, at as little as 10% of this studys sequencing depth, 30% of the estimated complete microbiome functional repertoire was detectable. However, even at the full depth used in this study, we were only able to recover an estimated 60% of that complete functional repertoire. ConclusionsTogether, these findings and our depth-to-function mapping framework provide practical guidelines for the design and interpretation of WMS studies. Coordinating sequencing depth planning with annotation strategy, experimental design, and rigorous batch control is thus essential for robust detection of microbial functions and for ensuring reproducible microbiome insights.

bioinformatics↗

bioRxiv2025

Which pLM to choose?

AO_SCPLOWBSTRACTC_SCPLOWProtein-language models (pLMs) provide a novel means for mapping the protein space. Which of these new maps best advances specific biological analyses, however, is not obvious. To elucidate the principles of model selection, we benchmarked fourteen pLMs, spanning several orders of magnitude in parameter count, across a hundred million protein pairs, to assess how well they capture sequence, structure, and function similarity. For each model, we distinguish inherent information, i.e. signal recoverable from raw-embedding distances, and extractable information, i.e. signal revealed through additional supervised training. Three key results emerge. First, pLM protein representation space is inherently different from the space of biological protein representations, i.e. sequences or structures. Here, a size-performance paradox is salient - mid-scale foundation models are as good as much larger ones in reflecting all tested biological properties. Second, pLM representations compress and store biological information in proportion to model size. That is, a lightweight feed-forward network can be trained on embedding pairs to predict said biological properties well - a capacity dividend. Finally, we observe that a task-specific learning radically reshapes the embedding space, gaining inherent understanding of the task, but garbling any further extractions. In other words, smaller pLMs can provide efficient and compute-light general insight. Larger models are advantageous only when fine-tuning is planned to accomplish a specific task. Furthermore, representations generated by "specialist" models are not immediately generalizable throughout protein biology. Thus, for pLMs, bigger isnt always better.

bioinformatics↗

bioRxiv2025

Quantifying uncertainty in Protein Representations Across Models and Task

Embeddings, derived by language models, are widely used as numeric proxies for human language sentences and structured data. In the realm of biomolecules, embeddings serve as efficient sequence and/or structure representations, enabling similarity searches, structure and function prediction, and estimation of biophysical and biological properties. However, relying on embeddings without assessing the models confidence in its ability to accurately represent molecular properties is a critical flaw--akin to using a scalpel in surgery without verifying its sharpness. In this study, we propose a means to evaluate the ability of protein language models to represent proteins, assessing their capacity to encode biologically relevant information. Our findings reveal that low-quality embeddings often fail to capture meaningful biology, displaying vector properties indistinguishable from those of randomly generated sequences. A key contributor to this performance issue is the models failure to learn the underlying biology from unevenly distributed sequence spaces in the training data. Our novel, model-agnostic scoring framework is, to the best of our knowledge, the first to quantify protein sequence embedding reliability. We believe that our robust approach to screening embeddings prior to making biological inferences, stands to significantly enhance the reliability of downstream applications.

bioinformatics↗

bioRxiv2024

Deciphering enzymatic potential in metagenomic reads through DNA language model

Microbial communities drive essential global processes, yet much of their functional potential remains unexplored. Metagenomics stands to elucidate this microbial "dark matter" by directly sequencing the microbial community DNA from environmental samples. However, the exploration of metagenomic sequences is mostly limited to establishing their similarity to curated reference sequences. A paradigm shift - language model (LM) -based methods - offer promising avenues for reference-free analysis of metagenomic reads. Here, we introduce two LMs, a pretrained foundation model REMME, aimed at understanding the DNA context of metagenomic reads, and the fine-tuned REBEAN for predicting the enzymatic potential encoded within the read-corresponding genes. By emphasizing function recognition over gene identification, REBEAN labels gene-encoded molecular functions of previously explored and new (orphan) sequences. Even though it was not trained to do so, REBEAN identifies the genes function-relevant parts. It thus expands enzymatic annotation of unassembled metagenomic reads. Here, we present novel enzymes discovered using our models, highlighting model impact on our understanding of microbial communities.

bioinformatics↗

bioRxiv2024

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 KEGG. 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.

bioinformatics↗

bioRxiv2023

Functional profiling of the sequence stockpile: a review and assessment of in silico prediction tools

In silico functional annotation of proteins is crucial to narrowing the sequencing-accelerated gap in our understanding of protein activities. Numerous function annotation methods exist, and their ranks have been growing, particularly so with the recent deep learning-based developments. However, it is unclear if these tools are truly predictive. As we are not aware of any methods that can identify new terms in functional ontologies, we ask if they can, at least, identify molecular functions of new protein sequences that are non-homologous to or far-removed from known protein families. Here, we explore the potential and limitations of the existing methods in predicting molecular functions of thousands of such orphan proteins. Lacking the ground truth functional annotations, we transformed the assessment of function prediction into evaluation of functional similarity of orphan siblings, i.e. pairs of proteins that likely share function, but that are unlike any of the currently functionally annotated sequences. Notably, our approach transcends the limitations of functional annotation vocabularies and provides a platform to compare different methods without the need for mapping terms across ontologies. We find that most existing methods are limited to identifying functional similarity of homologous sequences and are thus descriptive, rather than predictive of function. Curiously, despite their seemingly unlimited by-homology scope, novel deep learning methods also remain far from capturing functional signal encoded in protein sequence. We believe that our work will inspire the development of a new generation of methods that push our knowledge boundaries and promote exploration and discovery in the molecular function domain.

bioinformatics↗

bioRxiv2022

Rethinking bacterial relationships in light of their molecular abilities

Determining the repertoire of a microbes 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.

bioinformatics↗

bioRxiv2021

mebipred: identifying metal-binding potential in protein sequences.

Metal-binding proteins have a central role in maintaining life processes. Nearly one-third of known protein structures contain metal ions that are used for a variety of needs, such as catalysis, DNA/RNA binding, protein structure stability, etc. Identifying metal-binding proteins is thus crucial for understanding the mechanisms of cellular activity. However, experimental annotation of protein metal-binding potential is severely lacking, while computational techniques are often imprecise and of limited applicability. We developed a novel machine learning-based method, mebipred, for identifying metal-binding proteins from sequence-derived features. This method is nearly 90% accurate in recognizing proteins that bind metal ions and ion containing ligands. Moreover, the identity of ten ubiquitously present metal ions and ion-containing ligands can be annotated. mebipred is reference-free, i.e. no sequence alignments are involved, and outperforms other prediction methods, both in speed and accuracy. mebipred can also identify protein metal-binding capabilities from short sequence stretches and, thus, may be useful for the annotation of metagenomic samples metal requirements inferred from translated sequencing reads. We performed an analysis of microbiome data and found that ocean, hot spring sediments and soil microbiomes use a more diverse set of metals than human host-related ones. For human-hosted microbiomes, physiological conditions explain the observed metal preferences. Similarly, subtle changes in ocean sample ion concentration affect the abundance of relevant metal-binding proteins. These results are highlight mebipreds utility in analyzing microbiome metal requirements. mebipred is available as a web server at services.bromberglab.org/mebipred and as a standalone package at https://pypi.org/project/mymetal/

bioinformatics↗

bioRxiv2021

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.

bioinformatics↗

bioRxiv2021

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). PredictProteins infrastructure has moved to the LCSB increasing throughput; the use of MMseqs2 sequence search reduced runtime five-fold; 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. AvailabilityFreely accessible webserver PredictProtein.org; Source and docker images: github.com/rostlab

bioinformatics↗

bioRxiv2020

Shedding Light on Microbial Dark Matter with A Universal Language of Life

The majority of microbial genomes have yet to be cultured, and most proteins predicted from microbial genomes or sequenced from the environment cannot be functionally annotated. As a result, current computational approaches to describe microbial systems rely on incomplete reference databases that cannot adequately capture the full functional diversity of the microbial tree of life, limiting our ability to model high-level features of biological sequences. The scientific community needs a means to capture the functionally and evolutionarily relevant features underlying biology, independent of our incomplete reference databases. Such a model can form the basis for transfer learning tasks, enabling downstream applications in environmental microbiology, medicine, and bioengineering. Here we present LookingGlass, a deep learning model capturing a "universal language of life". LookingGlass encodes contextually-aware, functionally and evolutionarily relevant representations of short DNA reads, distinguishing reads of disparate function, homology, and environmental origin. We demonstrate the ability of LookingGlass to be fine-tuned to perform a range of diverse tasks: to identify novel oxidoreductases, to predict enzyme optimal temperature, and to recognize the reading frames of DNA sequence fragments. LookingGlass is the first contextually-aware, general purpose pre-trained "biological language" representation model for short-read DNA sequences. LookingGlass enables functionally relevant representations of otherwise unknown and unannotated sequences, shedding light on the microbial dark matter that dominates life on Earth. AvailabilityThe pretrained LookingGlass model and the transfer learning-derived models demonstrated in this paper are available in the LookingGlass release v1.01. The open source fastBio Github repository and python package provides classes and functions for training and fine tuning deep learning models with biological data2. Code for reproducing analyses presented in this paper are available as an open source Github repository3.

bioinformatics↗

bioRxiv2020

Snow Microbiome Functional Analyses Reveal Novel Microbial Metabolism of Complex Organic Compounds

Microbes active in extremely cold environments 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 twenty snow samples collected during early and late spring in Svalbard, Norway using our computational read-based microbiome function annotation tool, mi-faser. Our results revealed a more diverse microbiome functional capacity and activity in the early compared to in the late spring samples. The dissimilarity between the metagenomes and metatranscriptomes of the same samples was also significantly higher in the early spring. These findings suggest that early spring samples may contain a larger fraction of DNA of dormant organisms, while late spring samples reflect a new community that is metabolically active. We additionally showed that the abundance of the sequencing reads mapping to the fatty acid synthesis-related microbial pathways was significantly positively correlated with organic acid levels, in both our late spring metagenomes and metatranscriptomes. Moreover, the geraniol degradation pathway and the styrene degradation pathway read abundances correlated and inversely correlated, respectively, with the organic acid levels. These results suggest a possible nutrient switch. Our study thus highlights the activity of microbial degradation pathways of complex organic compounds previously unreported at low temperatures.

bioinformatics↗

bioRxiv2019

fuNTRp: Identifying protein positions for variation driven functional tuning

Evaluating the impact of non-synonymous genetic variants is essential for uncovering disease associations and mechanisms of evolution. Understanding corresponding sequence changes is also fundamental for synthetic protein design and stability assessments. However, the performance gain of variant effect predictors observed in recent years is not in line with the increased complexity of new methods. One likely reason for this might be that most approaches use similar sets of gene/protein features for modeling variant effect, often emphasizing sequence conservation. While high levels of conservation highlight residues essential for protein activity, much of the in vivo observable variation is arguably weaker in its impact and, thus, requires evaluation at a higher level of resolution. Here we describe function Neutral/Toggle/Rheostat predictor (funtrp), a novel computational method that categorizes protein positions based on the position-specific expected range of mutational impacts: Neutral (weak/no effects), Rheostat (function-tuning positions), or Toggle (on/off switches). We show that position types do not correlate strongly with familiar protein features such as conservation or protein disorder. We also find that position type distribution varies across different protein functions. Finally, we demonstrate that position types reflect experimentally determined functional effects and can thus improve performance of existing variant effect predictors and suggest a way forward for the development of new ones.

bioinformatics↗

bioRxiv2017

Identifying Crohns disease signal from variome analysis

BackgroundAfter many years of concentrated research efforts, the exact cause of Crohns disease remains unknown. Its accurate diagnosis, however, helps in management and even preventing the onset of disease. Genome-wide association studies have identified 140 loci associated with CD, but these carry very small log odds ratios and are uninformative for diagnoses.\n\nResultsHere we describe a machine learning method - AVA,Dx (Analysis of Variation for Association with Disease) - that uses whole exome sequencing data to make predictions of CD status. Using the person-specific variation in these genes from a panel of only 111 individuals, we built disease-prediction models informative of previously undiscovered disease genes. In this panel, our models differentiate CD patients from healthy controls with 71% precision and 73% recall at the default cutoff. By additionally accounting for batch effects, we are also able to predict individual CD status for previously unseen individuals from a separate CD study (84% precision, 73% recall).\n\nConclusionsLarger training panels and additional features, including regulatory variants and environmental factors, e.g. human-associated microbiota, are expected to improve model performance. However, current results already position AVA,Dx as both an effective method for highlighting pathogenesis pathways and as a simple Crohns disease risk analysis tool, which can improve clinical diagnostic time and accuracy.

bioinformatics↗