Biology subjects

Green, A. G.

Publications and source records attributed to Green, A. G..

14 recordsLinked to original sources

bioRxiv2026

Towards Sparse Causal Features for Zero-shot Mutation Effect Prediction in a Protein Language Model

Protein language models (pLMs) such as ESM-2 achieve strong zero-shot mutation-effect prediction, yet the internal computations supporting these predictions remain poorly understood. We introduce a sparse feature circuit framework that combines sparse autoencoders, integrated-gradients attribution, and activation patching to identify the latent features that causally mediate zero-shot mutation effect prediction in ESM-2 650M. We evaluate this framework over 67 mutations ranging from strongly deleterious to weakly deleterious in the DNAJA1 J-domain, where ESM-2 predictions agree strongly with deep mutational scanning measurements. We find that circuits selected by indirect effect recover the model's predictions more efficiently and provide more informative biological explanations than those selected by raw activation changes, showing that activation magnitude does not necessarily reflect causal importance. We find that related substitutions reuse substantial portions of their recovered circuits, ranging from 40% to 75%, and that the shared features often represent residues in three-dimensional contact with the mutation site. To our knowledge, our work provides the first causal, feature-level account of zero-shot mutation effect prediction in a pLM.

bioinformatics↗

bioRxiv2026

FARM: Forecasting Antibiotic Resistance in Mycobacterium tuberculosis using biophysics and machine learning

Antibiotic-resistant tuberculosis remains a major public health challenge, and rapid diagnosis of resistant infections based on genomic markers holds promise for improving time to effective treatment. However, the vast majority of clinically observed variants in resistance-associated genes remain of uncertain significance, limiting the utility of predictors. Here we develop a multimodal forecasting framework, FARM (Forecasting Antibiotic Resistance in Mycobacterium tuberculosis) to determine whether a newly observed mutation in a resistance gene may indeed cause resistance. Our framework combines structural context, biophysical energy features, protein language model features, and mutational AAIndex physicochemical descriptors. Using 345 labeled mutations from the World Health Organization 2021 catalogue, we train interpretable models that distinguish resistance-associated from non-resistance-associated variants with holdout AUCs of 0.843-0.943. In a novel temporal evaluation of 62 mutations reclassified after the training data was released, the selected Combined model achieved 80.7% recall of resistant reclassifications (resistant-class F1=86.8; AUC=0.735). Applied to 4,525 current uncertain-significance mutations, the framework prioritizes 696 candidate resistance mutations, including genes associated with the new antibiotics bedaquiline, delamanid, and pretomanid. These forecasts are intended to support future catalogue updates and experimental follow-up.

bioinformatics↗

bioRxiv2026

BIG-TB: A benchmark for evaluating prediction and interpretability of sequence-based machine learning using Mycobacterium tuberculosis genomes

Foundation models aim to learn useful representations of biological sequences. However, the applicability of these representations for a wide range of tasks, including phenotype prediction and variant discovery, is still in question, in large part due to the relatively small set of benchmark tasks. To this end, we present the Benchmarks for Interpretable prediction from Genomes of Tuberculosis, BIG-TB. We curate over 17,000 genomes with high-quality short read sequencing data and experimentally measured antibiotic resistance phenotypes, combined with a curated list of canonical resistance-conferring variants, and provide these data in an ML-ready format. BIG-TB defines two tasks for interrogating the utility of foundation models: (1) predictive performance of antibiotic resistance phenotypes, and (2) attribution of predictions to known resistance-conferring variants from an expert-curated dataset. Using our benchmark, we show that DNA-based foundation models do not yet outperform simple machine learning baselines (mean test AUC = 0.888 vs. 0.846 for best CNN variant vs. best DNABERT variant across drugs). Our benchmark also supports protein-based models, where performance is worse than DNA-based models due to the loss of representation of non-coding variants, but where foundation models are more competitive with simple ML. We show that the models with highest predictive performance do not necessarily perform the best at canonical resistance variant discovery - indicating that in many cases the improved performance may be due to non-causal associations between variants and phenotype. Finally, we show that the choice of embedding representation has a major impact on foundation model performance, and that representations that average over sequence position perform poorly at both prediction and canonical resistance variant discovery. Overall, BIG-TB provides a new type of benchmark for foundation models of biological sequences, facilitating comparison of representations across multiple tasks. Code available at https://github.com/SAGE-Lab-UMass/Big-TB-benchmark.

bioinformatics↗

bioRxiv2025

A frameshift mutation in the methyltransferase rlmN is associated with increased linezolid resistance in Mycobacterium tuberculosis

Linezolid is a key component of treatment regimens for multidrug-resistant and extensively drug-resistant tuberculosis, which is caused by the pathogen Mycobacterium tuberculosis (MTB). Resistance to linezolid in MTB has traditionally been attributed to mutations in the 23S rRNA (rrl) and ribosomal protein L3 (rplC), but only a fraction of clinically observed linezolid resistance is explained by mutations in these two genes. We report that an analysis of strains with paired whole-genome sequencing and linezolid minimum inhibitory concentration (MIC) phenotyping from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC) reveals that a relatively common frameshift mutation in MTB methyltransferase rlmN (5.3% of assessed isolates) is significantly associated with increased linezolid MIC. In additional to statistical associations, we provide evolutionary evidence of homology to an established linezolid resistance mechanism in Staphylococcus aureus, and structural evidence that the frameshift mutation likely ablates rlmN methyltransferase functionality.

bioinformatics↗

bioRxiv2025

The structural context of mutations in proteins predicts their effect on antibiotic resistance

In Mycobacterium tuberculosis, a prevalent and deadly pathogen, resistance to antibiotics evolves primarily through non-synonymous mutations in proteins. Sequence-based analyses are currently used to understand the genetic basis of antibiotic resistance, either via genotype-phenotype association, or via signals of convergent evolution. These methods focus on primary sequence and often neglect other biological signals such as protein structural information. We hypothesize that integrating the structural context of mutations improves the prediction of effects on function and phenotype. We curate high confidence structural annotations for the M. tuberculosis proteome from 1,371 crystallography and 2,316 AlphaFold predictions, and combine the structures with mutations from over 31,000 clinical M. tuberculosis isolates. We demonstrate that mutations in proteins known to cause resistance are clustered in 3D space, even in proteins where inactivating mutations at any position are thought to cause resistance. We develop a statistic to search the M. tuberculosis proteome for signal of clustered mutations, finding over 450 proteins that display this signal, many of which have a known relationship with antibiotic resistance. We show that a supervised classifier trained on 3D distance to known resistance sites alone has an F1 score of 94.6% at classifying mutations as resistance-conferring across proteins. This work demonstrates that protein structure provides useful information for categorizing which variants may cause antibiotic resistance, even when the majority of structures are AI-predicted.

bioinformatics↗

bioRxiv2025

Mechanistic evidence that motif-gated domain recognition drives contact prediction in protein language models

Protein language models (pLMs) achieve state-of-the-art performance on protein structure and function prediction tasks, yet their internal computations re-main opaque. Sparse autoencoders (SAEs) have been used to recover sparse features, called latents, from pLM layer representations, whose activations cor-relate with known biological concepts. However, prior work has not established which model concepts are causally necessary for pLM performance on down-stream tasks. Here, we adapt causal activation patching to the pLM setting and perform it in SAE latent space to extract the minimal circuit responsible for accuracy in a contact prediction task for two case study proteins. We observe that preserving only a tiny fraction of latent-token pairs (0.022% and 0.015%) is sufficient to retain contact prediction accuracy in a residue unmasking experiment. Our circuit indicates a two-step computation in which early-layer motif detectors respond to short local sequence patterns, gating mid-to-late domain detectors which are selective for protein domains and families. Path-level ab-lations confirm the causal dependence of domain latents on upstream motif latents. To evaluate these components quantitatively, we introduce two diagnostics: a Motif Conservation Test and a Domain Selectivity Framework that supports hypothesis-driven tests. All candidate motif-detector latents pass the conservation test, and 18/23 candidate domain-detector latents achieve AUROC [≥]0.95. To our knowledge, this is the first circuits-style causal analysis for pLMs, pin-pointing the motifs, domains, and motif-domain interactions that drive contact prediction in two specific case studies. The framework introduced herein will enable future mechanistic dissection of protein language models. Code available at https://github.com/NainaniJatinZ/plm_circuits

bioinformatics↗

bioRxiv2025

AMR-GNN: A multi-representation graph neural network framework to enable genomic antimicrobial resistance prediction

Whole-genome sequencing (WGS) data are an invaluable resource for understanding antimicrobial resistance (AMR) mechanisms. However, WGS data are high-dimensional and the lack of standardized genomic representations is a key barrier to AMR prediction. To fully explore these high-resolution data, we propose AMR-GNN, a graph deep learning-based framework that integrates multiple genomic representations with graph neural networks (GNN) to enable AMR prediction from genomic sequence data. We tested AMR-GNN with Pseudomonas aeruginosa, a clinically relevant Gram-negative bacterial pathogen known for its complex AMR mechanisms. We demonstrate that AMR-GNN addresses several key problems in AMR prediction with data-driven machine learning (ML) approaches, including using multiple genomic representations to enhance performance, mitigate the influence of clonal relationships, and identify informative biomarkers to provide explainability and generate novel hypotheses. Follow-up validation on the largest publicly available dataset spanning both Gram-negative and Gram-positive pathogens highlights AMR-GNNs broad applicability in detecting AMR in diverse and clinically relevant pathogen-drug combinations.

genomics↗

bioRxiv2025

Identification of chemical features that influence mycomembrane permeation and antitubercular activity

Tuberculosis (TB), caused by Mycobacterium tuberculosis (Mtb), is the deadliest single-agent infection worldwide. Current antibiotic treatment for TB takes a minimum of four months, underscoring the need for better therapeutics. The unique mycobacterial cell envelope, particularly the outermost mycomembrane, has long been thought to promote intrinsic antibiotic resistance by limiting compound entry into Mtb. Understanding chemical features that influence permeation across the mycomembrane may enable more accurate predictions of whole cell anti-Mtb activity, leading to development of more effective antibacterials. Here we query the mycomembrane permeation of over 1500 azide-tagged compounds in live Mtb with the bioorthogonal click chemistry-based assay PAC-MAN. We use cheminformatics and machine learning to identify chemical features associated with mycomembrane permeation and show that they have predictive value via systematic modification of two small molecule series. Additionally, we find that chemical features that influence mycomembrane permeation correlate with anti-Mtb activity in large compound libraries. These findings suggest that the mycomembrane is indeed a significant barrier to whole cell activity in Mtb and provide a rational framework for designing or modifying compounds to overcome this barrier.

microbiology↗

bioRxiv2022

The discovery of genome-wide mutational dependence in naturally evolving populations

BackgroundEvolutionary pressures on bacterial pathogens can result in phenotypic change including increased virulence, drug resistance, and transmissibility. Understanding the evolution of these phenotypes in nature and the multiple genetic changes needed has historically been difficult due to sparse and contemporaneous sampling. A complete picture of the evolutionary routes frequently travelled by pathogens would allow us to better understand bacterial biology and potentially forecast pathogen population shifts. MethodsIn this work, we develop a phylogeny-based method to assess evolutionary dependency between mutations. We apply our method to a dataset of 31,428 Mycobacterium tuberculosis complex (MTBC) genomes, a globally prevalent bacterial pathogen with increasing levels of antibiotic resistance. ResultsWe find evolutionary dependency within simultaneously- and sequentially-acquired variation, and identify that genes with dependent sites are enriched in antibiotic resistance and antigenic function. We discover 20 mutations that potentiate the development of antibiotic resistance and 1,003 dependencies that evolve as a consequence antibiotic resistance. Varying by antibiotic, between 9% and 80% of resistant strains harbor a dependent mutation acquired after a resistance-conferring variant. We demonstrate that mutational dependence can not only improve prediction of phenotype (e.g. antibiotic resistance), but can also detect sequential environmental pressures on the pathogen (e.g. the pressures imposed by sequential antibiotic exposure during the course of standard multi-antibiotic treatment). Taken together, our results demonstrate the feasibility and utility of detecting dependent events in the evolution of natural populations. Data and code available at: https://github.com/farhat-lab/DependentMutations

genomics↗

bioRxiv2021

A convolutional neural network highlights mutations relevant to antimicrobial resistance in Mycobacterium tuberculosis

Long diagnostic wait times hinder international efforts to address multi-drug resistance in M. tuberculosis. Pathogen whole genome sequencing, coupled with statistical and machine learning models, offers a promising solution. However, generalizability and clinical adoption have been limited in part by a lack of interpretability and verifiability, especially in deep learning methods. Here, we present a deep convolutional neural network (CNN) that predicts the antibiotic resistance phenotypes of M. tuberculosis isolates. The CNN performs with state-of-the-art levels of predictive accuracy. Evaluation of salient sequence features permits biologically meaningful interpretation and validation of the CNNs predictions, with promising repercussions for functional variant discovery, clinical applicability, and translation to phenotype prediction in other organisms.

genomics↗

bioRxiv2021

Dormant spores sense amino acids through the B subunits of their germination receptors

Bacteria from the orders Bacillales and Clostridiales differentiate into stress-resistant spores that can remain dormant for years, yet rapidly germinate upon nutrient sensing. How spores monitor nutrients is poorly understood but in most cases requires putative membrane receptors. The prototypical receptor from Bacillus subtilis consists of three proteins (GerAA, GerAB, GerAC) required for germination in response to L-alanine. GerAB belongs to the Amino Acid-Polyamine-Organocation superfamily of transporters. Using evolutionary co-variation analysis, we provide evidence that GerAB adopts a structure similar to an L-alanine transporter from this superfamily. We show that mutations in gerAB predicted to disrupt the ligand-binding pocket impair germination, while mutations predicted to function in L-alanine recognition enable spores to respond to L-leucine or L-serine. Finally, substitutions of bulkier residues at these positions cause constitutive germination. These data suggest that GerAB is the L-alanine sensor and that B subunits in this broadly conserved family function in nutrient detection.

microbiology↗

bioRxiv2021

Coevolution of interacting proteins through non-contacting and non-specific mutations

Proteins often accumulate neutral mutations that do not affect current functions1 but can profoundly influence future mutational possibilities and functions2-4. Understanding such hidden potential has major implications for protein design and evolutionary forecasting5-7, but has been limited by a lack of systematic efforts to identify potentiating mutations8,9. Here, through the comprehensive analysis of a bacterial toxin-antitoxin system, we identified all possible single substitutions in the toxin that enable it to tolerate otherwise interface-disrupting mutations in its antitoxin. Strikingly, the majority of enabling mutations in the toxin do not contact, and promote tolerance non-specifically to, many different antitoxin mutations, despite covariation in homologs occurring primarily between specific pairs of contacting residues across the interface. In addition, the enabling mutations we identified expand future mutational paths that both maintain old toxin-antitoxin interactions and form new ones. These non-specific mutations are missed by widely used covariation and machine learning methods10,11. Identifying such enabling mutations will be critical for ensuring continued binding of therapeutically relevant proteins, such as antibodies, aimed at evolving targets12-14.

evolutionary biology↗

bioRxiv2019

Proteome-scale discovery of protein interactions with residue-level resolution using sequence coevolution

The majority of protein interactions in most organisms are unknown, and experimental methods for determining protein interactions can yield divergent results. Here we use an orthogonal, purely computational method based on sequence coevolution to discover protein interactions at large scale. In the model organism Escherichia coli, 53% of protein pairs in the proteome are eligible for our method given currently available sequenced genomes. When assaying the entire cell envelope proteome, which is understudied due to experimental challenges, we found 620 likely interactions and their predicted structures, increasing the space of known interactions by 529. Our results show that genomic sequencing data can be used to predict and resolve protein interactions to atomic resolution at large scale. Predictions and code are freely available at https://marks.hms.harvard.edu/ecolicomplex and https://github.com/debbiemarkslab/EVcouplings

bioinformatics↗

bioRxiv2018

The EVcouplings Python framework for coevolutionary sequence analysis

SummaryCoevolutionary sequence analysis has become a commonly used technique for de novo prediction of the structure and function of proteins, RNA, and protein complexes. This approach requires extensive computational pipelines that integrate multiple tools, databases, and data processing steps. We present the EVcouplings framework, a fully integrated open-source application and Python package for coevolutionary analysis. The framework enables generation of sequence alignments, calculation and evaluation of evolutionary couplings (ECs), and de novo prediction of structure and mutation effects. The application has an easy to use command line interface to run workflows with user control over all analysis parameters, while the underlying modular Python package allows interactive data analysis and rapid development of new workflows. Through this multi-layered approach, the EVcouplings framework makes the full power of coevolutionary analyses available to entry-level and advanced users.\n\nAvailabilityhttps://github.com/debbiemarkslab/evcouplings\n\nContactsander.research@gmail.com, debbie@hms.harvard.edu

bioinformatics↗