When bacteria become resistant to antibiotics, they often do it by making more of a pump — a multi-protein machine that spans the bacterial cell envelope and expels the drug. But these pumps are assembled from several different protein parts that must arrive in matched quantities: extra copies of one part do not help, and are wasteful or actively toxic. Yeast faces the identical problem with its ribosome, and solves it not by gene copy number but by promoter architecture — the physical layout of control sites upstream of each gene encodes how much of each subunit gets made, so that duplicated genes get weaker promoters and single-copy genes get stronger ones. This proposal asks whether the same two-layer logic — a spatial promoter grammar operating under a stoichiometric constraint — governs bacterial resistance genes, and whether building that logic into predictive models improves quantitative resistance prediction on clinical data.
The opportunity is concrete because the clinical data now exists. A large-scale analysis of 12,350 Mycobacterium tuberculosis isolates found that mutations in the 100 base pairs upstream of resistance genes shift quantitative drug susceptibility as much as coding mutations do, including 492 variants that raise the minimum inhibitory concentration (MIC — the lowest drug concentration that stops growth) and 91 that make bacteria more sensitive Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis reveals genetic determinants of resistance and susceptibility in a target gene approach. Yet that study treated each gene and its promoter window in isolation. Meanwhile, work on yeast ribosomes shows promoters hardcoding dosage compensation through motif spacing Compensation for differences in gene copy number among yeast ribosomal proteins is encoded within their promoters, and work on protein interaction networks shows why subunits of stable complexes cannot tolerate isolated overproduction What properties characterize the hub proteins of the protein-protein interaction network of Saccharomyces cerevisiae?. Nobody has connected these: we ask whether the promoter mutations that drive clinical resistance — and those that backfire into hypersensitivity — are shaped by the stoichiometric demands of the protein complexes the resistance genes belong to.
Core Claim
The claim has two parts. First, mechanistic: a measurable fraction of non-coding resistance variation is explained not by how much a promoter mutation changes expression in isolation, but by whether the resulting expression change preserves or breaks the required subunit ratio of the multiprotein assembly the gene participates in. Promoter mutations that raise one efflux-pump subunit without its partners should be systematically overrepresented among the hypersensitizing and low-fitness variants, because unassembled subunits impose the same misassembly costs that make stable-complex hubs dosage-sensitive in yeast. Second, predictive: encoding (a) a spatial distance-decay model of promoter influence and (b) a stoichiometric coupling constraint into a genome-based MIC predictor will improve quantitative resistance prediction over the strongest existing model, particularly on isolates whose resistance is promoter-driven.
This is a claim about bacteria under drug selection, which distinguishes it from the closest existing work. Zeevi et al. established promoter-encoded dosage compensation, but only for the eukaryotic ribosome in normal growth, with no antimicrobial context, and left 60% of promoter-activity variance unexplained Compensation for differences in gene copy number among yeast ribosomal proteins is encoded within their promoters. The CRyPTIC analysis connected promoter mutations to MIC but used a fixed 100 bp window, one gene at a time, with no interaction network and no spatial model Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis.... GeneBac coupled sequence CNNs to protein-interaction graphs but its edges are generic confidence scores, not assembly stoichiometries, and it has no promoter distance grammar at all Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction. The aneuploidy work in Candida shows drug selection can exploit gross dosage shifts, but at the cost of instability — copy-number bludgeoning rather than promoter tuning Chromosome 1 trisomy confers resistance to aureobasidin A in Candida albicans. Our proposal sits precisely in the empty cell: fine-grained spatial promoter grammar, stoichiometric constraints, and clinical resistance phenotypes.
Research Program
The program builds a model we call SGN (Stoichiometric Grammar Network) and uses it both as a predictor and as a testable hypothesis about resistance evolution. Three components.
First, annotation. For each resistance-relevant gene in the CRyPTIC M. tuberculosis dataset and the P. aeruginosa clinical dataset, we extend the promoter window from the 100 bp used in the CRyPTIC study to roughly ±2 kb, motivated by the finding that pioneer transcription factors act within a tight 1 kb window A synergistic DNA logic predicts genome-wide chromatin accessibility and by bacterial spatial logic in which binding directly over the core promoter represses by physically blocking RNA polymerase while upstream binding activates OpaR Controls the Metabolism of c-di-GMP in Vibrio parahaemolyticus. We annotate complex membership and required subunit stoichiometry for multi-subunit assemblies using STRING interactions as GeneBac does Sequence-based modelling..., but we upgrade the relevant edges — for tripartite resistance-nodulation-division pumps such as MexAB-OprM and MexGHI-OpmD Rewiring of Gene Expression in Pseudomonas aeruginosa During Diauxic Growth Reveals an Indirect Regulation of the MexGHI-OpmD Efflux Pump by Hfq, Disinfection of Hospital Sink Drains Enriches Pseudomonadota and Efflux Pump-Mediated Antibiotic Resistance in Reestablished Biofilms — to carry assembly stoichiometry rather than generic confidence.
Second, the model. The promoter module parameterizes regulatory influence as a learned function of physical distance to the transcription start site, following the B-spline distance-decay approach that beat conservation-based weighting with a median correlation of 0.69 to 0.74 against observed expression differences A Quantitative Model of Transcriptional Regulation Reveals the Influence of Binding Location on Expression, implemented with dense convolutional blocks plus self-attention over the promoter window, the architecture that reached 84.79% average AUC predicting expression from histone marks Predicting gene expression from histone modifications with self-attention based neural networks and transfer learning. The stoichiometric layer couples predicted subunit expression within a complex through a relative-imbalance penalty:
$$\mathcal{L}_{\text{stoich}} = \sum_{C} \sum_{i \in C} \left( \log \hat{e}_i - \log s_i - \mu_C \right)^2$$
where $\hat{e}_i$ is the model's predicted expression of subunit $i$, $s_i$ is the required copy number of subunit $i$ in complex $C$, and $\mu_C$ is the mean of $\log \hat{e}_i - \log s_i$ across the complex, so the penalty fires only on relative imbalance, never on a coordinated increase — a complex-wide pump overexpression is legitimate resistance, a single-subunit overshoot is not. This mirrors the yeast result that duplicate-pair promoter activities sum to a stabilized equimolar output Compensation for differences in gene copy number....
Third, the evolutionary analysis. We test whether the 91 hypersensitizing and 492 MIC-elevating variants from the CRyPTIC study segregate by predicted stoichiometric consequence: promoter mutations predicted to raise a single subunit's expression while breaking the assembly ratio should be enriched among hypersensitizing variants, whereas coordinated or target-proximal mutations should dominate the resistance set. We also exploit the paired RNA-seq compendium of 386 P. aeruginosa strains, where gene expression is predictable from DNA sequence with nearly 0.9 Pearson correlation Sequence-based modelling..., to check predicted expression imbalance against measured imbalance directly.
A known tension must be addressed rather than ignored. Protein language models for resistance find their attention concentrated on conserved residues and binding pockets Prediction of antibiotic resistance mechanisms using a protein language model, while the distance-decay work shows physical position near the transcription start site predicts regulatory output far better than conservation scores, with non-conserved near-TSS sites functionally active A Quantitative Model of Transcriptional Regulation.... The CRyPTIC data already hints at the spatial paradigm: seven novel rifampicin-resistance mutations fell outside the classic conserved resistance-determining region. Our spatial module is built to adjudicate this on non-coding variation: we fit both a conservation-weighted and a distance-weighted promoter influence model and compare which predicts variant MIC effects.
Baselines
The naive reference point is catalog lookup: querying isolates against the WHO M. tuberculosis Mutation Catalogue and homology search against resistance gene databases. It is the do-nothing baseline for clinical practice and it captures everything already known, but it has no mechanism at all for non-coding effects and fails on divergent sequences — the corpus reports that BLAST accuracy falls to 0.917 at the 40% sequence-identity threshold where a protein language model holds 0.927 Prediction of antibiotic resistance mechanisms using a protein language model. The strongest existing method is GeneBac, a CNN over sequence variants coupled to a graph neural network over STRING interactions, which predicts continuous MIC with Pearson correlations over 0.87 on held-out M. tuberculosis strains and over 0.76 on P. aeruginosa Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction; we will re-run it on identical CRyPTIC and P. aeruginosa splits with an equal tuning budget, and beating it means a consistent Pearson improvement across drugs on the same held-out isolates, with a pre-registered larger margin on the subset of isolates whose resistance is promoter-mutation-driven. The third baseline is the target-gene linear mixed model from the CRyPTIC analysis itself, which linked 100 bp promoter variants to quantitative MIC and identified the 492 MIC-elevating and 91 hypersensitizing variants Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis...; no reported result found on Lacuna for a predictive-correlation metric in that paper, so the comparison is variant recovery — SGN should recover its 492 variants and additionally find resistance-associated promoter variants outside its 100 bp window. Finally, two ablations isolate the key ingredients: the full architecture with the stoichiometric loss removed, and with the spatial distance-decay module replaced by a flat window encoder. If the full model beats GeneBac but the ablations do not beat the ablated versions, the ingredients, not the capacity, are doing the work.
Why not just…?
Why not just BLAST against resistance catalogs? Because the entire non-coding half of the problem is invisible to it, and even the coding half degrades under divergence: at 40% sequence identity BLAST drops to 0.917 accuracy against 0.927 for a protein language model Prediction of antibiotic resistance mechanisms using a protein language model, and structure-aware methods add 3.1% macro accuracy on a low-homology benchmark Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks. Catalog lookup cannot, even in principle, predict that a promoter mutation hypersensitizes a strain.
Why not just widen the linear mixed model's promoter window? This is the most serious alternative, and we treat it as a live threat rather than a strawman. It is cheap, interpretable, and the existing model already found that promoter mutations matter. Two things limit it. It fits each gene independently, so it cannot represent a stoichiometric constraint that couples genes to each other — the effect of raising one subunit depends on the expression state of its partners, which is exactly the non-additive epistasis the CRyPTIC authors themselves observed for rifampicin and ethambutol Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis.... And linear models were already benchmarked and beaten on this data: GeneBac substantially outperformed Lasso and ElasticNet on MIC prediction Sequence-based modelling.... Still, we will run exactly this baseline — the mixed model with windows widened to match SGN's — and if it matches SGN's variant recovery and MIC correlation, the spatial and stoichiometric machinery is not earning its complexity and we will say so.
Why not just make GeneBac bigger? Because the hypothesis is about representation, not capacity. The decisive control is parameter-matched: we take the GeneBac architecture, hold its parameter count and training data fixed, and swap in only the distance-decay promoter module and the stoichiometric loss. If the gains survive that control, capacity is not the explanation; if they vanish, the simpler explanation wins and the mechanistic claim is weakened accordingly. We are honest that this outcome is possible — the rival mechanism is already documented: horizontally transferred resistance genes carry costs through codon composition and ribosomal sequestration, costs that intensify at sub-lethal antibiotic levels The Fitness Effects of Codon Composition of the Horizontally Transferred Antibiotic Resistance Genes Intensify at Sub-lethal Antibiotic Levels. Translational stoichiometry could dominate transcriptional stoichiometry; experiment 3 below is designed to detect this rather than assume it away.
Why not just use binary resistant/susceptible labels? Because the phenomena at the heart of the proposal live below the clinical breakpoint. The 91 hypersensitizing variants are invisible to binary classification, and sub-breakpoint MIC elevations are precisely what precede treatment failure Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis.... We report Pearson correlation on continuous MIC as the primary metric throughout.
Proposed Experiments
Experiment 1 tests the spatial grammar directly. We map the positions of the 492 MIC-elevating and 91 hypersensitizing promoter variants relative to the transcription start site and fit B-spline distance-decay influence functions, then compare against a conservation-score-weighted model, replicating in a bacterial resistance context the comparison that distance decisively won in mouse tissues A Quantitative Model of Transcriptional Regulation.... The decisive check: does variant position predict MIC effect better than conservation, and do effect sizes decay with distance as the grammar predicts?
Experiment 2 tests the predictive claim. SGN is trained and evaluated on the CRyPTIC dataset (12,350 evaluated isolates, 13 drugs) and the P. aeruginosa dataset (1,109 strains, 14 drugs), on identical splits to GeneBac, which reports Pearson correlations over 0.87 (MTB) and over 0.76 (PA) on held-out strains Sequence-based modelling.... Success is a consistent improvement across drugs, a larger margin on the promoter-driven subset, and recovery of the known 492 MIC-elevating variants plus resistance variants outside the 100 bp window.
Experiment 3 tests the stoichiometric mechanism against measured data and against the rival translational-cost explanation. Using the 386 P. aeruginosa strains with paired RNA-seq, we ask whether promoter variants predicted by SGN to create subunit imbalance show the corresponding measured expression imbalance across the Mex pump components, and whether imbalance predicts deviation from expected MIC. If expression ratios stay balanced whenever expression rises — that is, if promoters appear already tuned to coordinated expression, as in the yeast ribosome — the stoichiometric lens adds nothing and we report the negative result. A small wet-lab confirmation closes the loop: LacZ reporter assays in P. aeruginosa for a handful of Mex promoter mutations, using the same assay logic that established activation-versus-repression by binding position OpaR Controls the Metabolism of c-di-GMP in Vibrio parahaemolyticus, to verify that the model's positional predictions match single-promoter reality.
The payoff is two-directional. If the stoichiometric grammar is real, clinical sequencing gains a way to flag promoter variants that will hypersensitize or burn out — the bacterial analogue of the unstable aneuploid resistance that reverts when drug pressure lifts Chromosome 1 trisomy confers resistance to aureobasidin A in Candida albicans — and combination regimens could exploit stoichiometric fragility deliberately. If it is not real, we will have cleanly bounded where promoter-encoded dosage compensation ends, and the spatial grammar module alone will still have extended MIC prediction beyond the 100 bp frontier.
References
- Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis reveals genetic determinants of resistance and susceptibility in a target gene approach
- Compensation for differences in gene copy number among yeast ribosomal proteins is encoded within their promoters
- What properties characterize the hub proteins of the protein-protein interaction network of Saccharomyces cerevisiae?
- Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction
- Chromosome 1 trisomy confers resistance to aureobasidin A in Candida albicans
- A synergistic DNA logic predicts genome-wide chromatin accessibility
- OpaR Controls the Metabolism of c-di-GMP in Vibrio parahaemolyticus
- Rewiring of Gene Expression in Pseudomonas aeruginosa During Diauxic Growth Reveals an Indirect Regulation of the MexGHI-OpmD Efflux Pump by Hfq
- Disinfection of Hospital Sink Drains Enriches Pseudomonadota and Efflux Pump-Mediated Antibiotic Resistance in Reestablished Biofilms
- A Quantitative Model of Transcriptional Regulation Reveals the Influence of Binding Location on Expression
- Predicting gene expression from histone modifications with self-attention based neural networks and transfer learning
- Prediction of antibiotic resistance mechanisms using a protein language model
- Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks
- The Fitness Effects of Codon Composition of the Horizontally Transferred Antibiotic Resistance Genes Intensify at Sub-lethal Antibiotic Levels