Predicting Bacterial Heteroresistance from Genomes via Single-Cell Transcriptomics

When a clinical lab sequences a bacterial isolate's whole genome to predict drug resistance, it grinds up millions of cells and reads the averaged DNA. That average is blind to the rare subpopulations that matter most: in one cohort of Enterobacter cloacae complex isolates, 24.4% of strains labelled colistin-sensitive by standard testing in fact harboured resistant subclones, present at frequencies as low as 1 in $10^6$ cells, with subpopulation MICs at or above 128 mg/L (Clinical characteristics, molecular epidemiology and mechanisms of colistin heteroresistance in Enterobacter cloacae complex). Single-cell RNA sequencing — sequencing the messenger RNA of individual cells one at a time — was invented precisely to resolve such rare and transient states, but it has matured almost entirely in mammalian tissue, because bacterial mRNA lacks the poly(A) tails that standard droplet chemistry captures, decays within minutes, and sits behind a rigid cell wall. The result is an odd stalemate: the genomic side has fast, accurate models that cannot see rare resistant cells or non-mutational resistance, and the single-cell side has the resolution but no foothold in bacteria.

This proposal builds that foothold and uses it computationally. We will develop a paired single-cell bacterial assay that measures, in each individual bacterium, its transcriptome together with targeted genotypes at the resistance loci nominated by bulk whole-genome sequencing; and we will train a model that consumes an isolate's whole-genome features and predicts the distribution of single-cell transcriptional states — including the fraction of cells that switch into a resistance program — under a given antibiotic and concentration. The clinical payoff is a test that flags heteroresistant isolates and attributes the mechanism (for example PhoPQ-driven arnA induction versus carriage of a resistance plasmid) from genome data alone, in the isolates where today's genomic pipelines are confidently wrong.

Core Claim

The claim is that an isolate's whole-genome sequence carries enough information to predict which cells within it will transcriptionally bifurcate into a resistant state under drug pressure, once a model has been trained on paired single-cell genotype-plus-transcriptome data. Formally, for an antibiotic $a$ at concentration $c$, the model predicts the mixture

$$\hat{p}(x \mid g, a, c) = \big(1-\hat\theta(g,a,c)\big)\,q_0(x) \;+\; \hat\theta(g,a,c)\,q_R(x;\,g),$$

where $x$ is a single-cell expression vector, $g$ is a fixed embedding of the isolate's whole-genome features, $q_0$ is the transcriptional program of the dying majority, $q_R$ the induced-resistance program, and $\hat\theta$ the predicted persister fraction. Heteroresistance is called when $\hat\theta$ exceeds the detection floor of population analysis profiling. The architecture follows chemCPA (Predicting single-cell perturbation responses for unseen drugs), which pre-trained on ~1.3 million bulk perturbation profiles (L1000) and fine-tuned on multiplexed single-cell data to predict out-of-distribution drug responses with median $r^2 = 0.79$ across differentially expressed genes — versus median $r^2 \approx 0.00$ without bulk pre-training. We transplant that recipe to bacteria: pre-train on the ~704 bulk transcriptomic datasets across 27 genera hosted in PATRIC (PATRIC, the bacterial bioinformatics database and analysis resource), replace chemCPA's human gene space and drug embeddings with operon-aware bacterial gene units and antibiotic-class embeddings, and condition everything on the WGS-derived genotype embedding $g$ — the ingredient no prior perturbation model has.

This is distinct from each neighbouring line of work. Single-cell sequencing of patient microbiomes via single amplified genomes (Elucidation of population-based bacterial adaptation to antimicrobial treatment by single-cell sequencing analysis of the gut microbiome of a hospital patient) reads individual bacteria's DNA — it mapped 29 resistance-gene subtypes and 309 horizontal-transfer events in one patient — but cannot measure expression. The FAST acoustic-flow-cytometry assay (Rapid susceptibility profiling of carbapenem-resistant Klebsiella pneumoniae) detects surviving single cells optically in 158 minutes but produces zero sequence data. And the pairing of WGS with bulk RNA-seq in phage-resistance studies (Investigation of the therapeutic efficacy and resistance mechanisms of lytic phages targeting ST218 KL57 CR-hvKP) showed that strains can resist therapy with no coding mutation at all — by transcriptionally silencing the target capsular locus — precisely the class of resistance that both bulk methods average away. Our contribution sits in the empty cell of that grid: per-bacterium genotype linked to per-bacterium expression, then distilled into a genome-only predictor.

Research Program

The program has three parts. First, the assay: individual bacteria are separated by microfluidic partitioning, cell walls are weakened enzymatically, and mRNA is reverse-transcribed from random hexamers (not poly(dT)) carrying a well barcode; split-pool rounds of barcoding then index each cell's cDNA. In the same compartments, multiplexed amplicons of the resistance loci identified from the isolate's bulk WGS — resistance genes, plasmid markers, and candidate regulatory SNPs such as mgrB or phoPQ lesions — are co-amplified so each cell carries both a transcriptome and a targeted genotype. This is the highest-risk element, and we set an explicit kill criterion: if, at an intermediate milestone of 10,000 cells profiled from a reference K. pneumoniae strain, fewer than 300 genes are detected per cell on average and known operons (co-transcribed gene groups) do not co-vary within cells, the wet-lab route is abandoned and the computational program proceeds on simulated bifurcations calibrated to bulk data, with the claim correspondingly narrowed.

Second, the benchmark: we will assemble roughly 60 clinical E. cloacae complex and K. pneumoniae isolates chosen to over-sample the heteroresistance rate of 24.4% reported for the ECC cohort, each with bulk WGS, gold-standard PAP quantification of resistant-subpopulation frequency across a colistin gradient, and single-cell profiles at three drug concentrations and two time points. Third, the model: an adversarial autoencoder in the chemCPA style, conditioned on $g$, pre-trained on PATRIC bulk transcriptomes, fine-tuned on the benchmark. The conditioning on $g$ is the key ingredient — the same model without it can predict a population-average response but cannot distinguish isolates with identical drug exposure histories and different heteroresistance outcomes, which is exactly what a genomic test must do.

Baselines

The naive reference point is current standard of care plus its best genomic upgrade: call susceptibility from the bulk MIC and, in the in-silico variant, predict that MIC from short-read WGS with the k-mer XGBoost pipeline of Developing an in silico minimum inhibitory concentration panel test for Klebsiella pneumoniae, which achieved 92% overall accuracy within one two-fold dilution on 1,668 isolates (98% for ciprofloxacin, 97% for amikacin) with a 3.1% overall very-major-error rate. That model is the strongest genomic competitor for quantitative MIC prediction, and it is precisely the system whose blind spot — low-frequency subclones — we target. It will be run on our benchmark's WGS data with its published hyperparameters and a tuning budget matched to ours, evaluated not only on MIC agreement but on heteroresistance detection, where we expect it to fail by construction because its training signal is a bulk phenotype.

The strongest single-cell competitor is FAST (Rapid susceptibility profiling of carbapenem-resistant Klebsiella pneumoniae), which on 48 carbapenem-resistant isolates reached a Matthew's correlation coefficient of 0.887 against broth microdilution with 100% sensitivity and 90% specificity, detecting persister cells that bulk testing missed, in 54 minutes (qualitative) to 158 minutes (quantitative MIC). We will run FAST in parallel on the exact same isolates, same inocula, same colistin lots, with thresholds tuned by the original protocol. Beating FAST means one of two things: strictly better detection of heteroresistant isolates against PAP ground truth (higher MCC), or matched detection plus mechanism attribution — correctly naming the resistance program (for example, an arnA-high, PhoPQ-activated cluster) where FAST can only say that cells survived. We pre-register both criteria and will report whichever holds.

The key ablation is our own model with the genotype conditioning removed: a chemCPA-style model trained on the same single-cell data but taking only antibiotic identity and concentration. If the conditioned and unconditioned models detect heteroresistant isolates equally well, the claim that genomes predict bifurcation is dead, and the honest conclusion would be that single-cell profiling works but not as a genomic predictor. A second ablation removes the PATRIC bulk pre-training; the chemCPA result (median $r^2 = 0.79$ with bulk pre-training versus $\approx 0.00$ without, from Predicting single-cell perturbation responses for unseen drugs) predicts this ablation should crater performance on held-out isolates, and if it does not, the pre-training is not earning its cost. One further reference point anchors the genome-feature choice: reference-based curated variant representations beat reference-free k-mers for polygenic polymyxin resistance (AUROC 0.894 versus 0.692, $P = 0.015$, from Predicting Phenotypic Polymyxin Resistance in Klebsiella pneumoniae through Machine Learning Analysis of Genomic Data), so $g$ will be built from reference-based variants plus accessory-element content, not k-mer counts alone.

Why not just…?

Why not just improve the existing WGS machine-learning models? Because their failure modes on this question are documented, not hypothetical. The same k-mer paradigm degrades sharply on polygenic polymyxin resistance (AUROC 0.692 for reference-free k-mers), and it cannot see non-mutational resistance at all: in the phage-therapy study, strains RM01, RM02, and RM12 were fully resistant with no coding mutations, having silenced the target locus transcriptionally. No amount of additional DNA features fixes a problem that lives in expression. Our benchmark deliberately includes such isolates to make this failure measurable.

Why not just run FAST, since it already catches persisters in under three hours? This is the most serious alternative, and we treat it as a real threat rather than a straw man. For a pure yes/no susceptibility answer at the bedside, FAST may win outright on speed, and we will say so if the head-to-head shows it. What FAST cannot do is attribute mechanism or predict across unmeasured isolate–antibiotic pairs: it produces no sequence data, so it cannot distinguish an arnA-induced heteroresistant subclone from a beta-lactamase-carrying one, tell you the subpopulation's MIC, or predict the response of a held-out isolate. The decisive experiment is the leave-one-sequence-type-out test: if our genotype-conditioned model predicts heteroresistance in unseen sequence types while FAST (which requires an assay per isolate) has no predictive analogue, the genomic claim stands even where FAST matches detection.

Why not just pair bulk RNA-seq with WGS, as the phage-resistance study did? Because bulk averaging destroys the signal we need. The resistant program lives in 1 cell per $10^4$–$10^6$; a bulk measurement of the same culture is overwhelmingly the transcriptome of the dying majority. Direct single-cell work has shown that bulk dilution smooths out extreme cell-specific expression entirely (Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation), and that within-type single-cell variation is conserved biological signal, not noise — which is exactly what makes it predictive rather than ignorable.

Why not just standardize PAP, the accepted gold standard? PAP works — it is our ground truth — but it requires serial plating, 24–48 hours of incubation, and manual colony counts, too slow for bloodstream infections where each hour of delay costs survival. PAP tells you the frequency of resistant cells after the fact; our model aims to predict that frequency from the genome before any subculture. PAP remains the referee, never the competitor.

Proposed Experiments and Decisive Checks

The decisive checks are, in order of escalating claim strength. First, assay validation: known operons must co-vary within single cells, and spiked-in genotype-phenotype pairs (a labelled colistin-resistant clone mixed into a susceptible population at known 1-in-$10^3$ to 1-in-$10^6$ ratios) must be recovered at the correct frequency — a direct, quantitative recovery test. Second, mechanism recovery: the model must surface an arnA-high, PhoPQ-pathway cluster in heteroresistant isolates matching the RT-qPCR findings of the ECC study, without that pathway being given to it. Third, prediction: on held-out isolates (leave-one-sequence-type-out), predicted $\hat\theta$ must correlate with PAP-measured resistant-subpopulation frequency, and detection MCC against PAP must beat both the bulk-WGS pipeline (expected near chance on the heteroresistant subset) and match or beat FAST's 0.887 on the shared isolates. Fourth, generalization across drugs: the model, trained on colistin, should transfer to a second antibiotic class with the same operon-aware machinery, the way chemCPA transferred across drug structures. If the wet-lab milestone fails, we fall back to the simulation-calibrated study and state plainly that the genomic-predictor claim is then untested in bacteria. No reported result exists on Lacuna for any genotype-to-single-cell-transcriptome predictor in bacteria; that absence, given how close each neighbouring line has come, is the reason this work is worth doing now.