When a doctor tests whether a bacterial infection is susceptible to an antibiotic, the test grows millions of cells together in a tube and measures the average. That average hides the problem: in roughly a quarter of Enterobacter cloacae complex isolates that standard testing calls fully sensitive to the last-resort drug colistin, a tiny subpopulation — about one cell in a million — survives at drug concentrations sixteen times the clinical breakpoint (Clinical characteristics, molecular epidemiology and mechanisms of colistin heteroresistance in Enterobacter cloacae complex). Today the only reliable way to find these hidden cells is a manual plating assay that takes two to three days, which is often longer than the patient has. This proposal builds a same-day alternative: trap individual bacteria in tiny water droplets, watch how fast each one acidifies its droplet as a live readout of its metabolism, do this while the antibiotic is present and across a panel of external pH conditions, and detect the rare cells that keep metabolizing when everything around them has stopped. The pH of each bacterium's surroundings is not an arbitrary measurement channel — the resistance mechanism itself is a pH-coupled catalytic process, and the acidic environments where colistin is most used clinically can shift whether that mechanism even shows up.
Core Claim
Colistin heteroresistance in Gram-negative pathogens can be detected phenotypically, at single-cell resolution, within a single working shift, by measuring per-cell extracellular acidification under antibiotic challenge. Two mechanisms make this possible where current testing fails. First, acidification is a metabolic output: a susceptible cell's proton production collapses under colistin, while a resistant cell — whose PhoPQ two-component system upregulates the arnA operon, catalytically decorating lipid A with aminoarabinose to neutralize the surface charge colistin attacks — keeps pumping out protons (Clinical characteristics, molecular epidemiology and mechanisms of colistin heteroresistance in Enterobacter cloacae complex). Second, that catalytic charge modification and the PhoPQ sensor that drives it are themselves pH-dependent, so a panel of assay pH values (5.5, 6.5, 7.4) should not merely read the phenotype but modulate its expression — meaning some isolates may be heteroresistant only in the acidic regime typical of urine, where colistin is a frequent last resort (Drug Resistance Patterns and Pathogen Distribution in Patients with Urinary Tract Infection in Suzhou, China: A Five-Year Retrospective Study (2020–2024)).
Why Proton Flux, and Why Per Cell
The signal and the instrument both have precedent, but never together and never on bacteria. SCO‐ph: Microfluidic Dynamic Phenotyping Platform for High‐Throughput Screening of Single Cell Acidification showed that single mammalian cells trapped in 140-picoliter droplets with a ratiometric fluorescent pH dye produce measurable, single-cell-specific acidification trajectories: hyperglycolytic cancer cells dropped droplet pH by 0.17 units in three hours, 2.6 times more than untreated cells. That paper's own open problems are our starting conditions: it tested only large eukaryotic cells, required unphysiologically weak buffering, and never touched antimicrobials. Bacteria are roughly tenfold smaller and generate far fewer protons per unit time, so adapting the assay is a genuine engineering problem, not a port. On the masking side, the phenomenon we chase is well documented but poorly instrumented: bulk resistance testing misses rare resistant subpopulations just as bulk tumor profiling misses rare stem-like cells that actually drive drug failure (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties), and single-cell sequencing can find resistance genes in individual gut bacteria — even showing that two morphologically identical Klebsiella pneumoniae strains carry divergent resistomes — but measures no enzyme activity, no kinetics, and no pH context (Elucidation of population-based bacterial adaptation to antimicrobial treatment by single-cell sequencing analysis of the gut microbiome of a hospital patient). The gap is a phenotypic, dynamic, pH-resolved single-cell assay for bacteria under drug pressure.
Research Program
Three linked components. The first is the instrument: a droplet microfluidic platform adapted from the SCO-pH architecture to bacterial scale, trading droplet count for signal per droplet (larger volumes where needed, a brighter ratiometric dye, and longer optical pathlength), with a Bayesian error-correction layer for droplet occupancy and identity errors — the SCO-pH platform itself reports a 16.5% identification error rate that had to be corrected computationally, and we expect worse at bacterial densities. A short antibiotic-conditioned enrichment step (four to six hours in sub-inhibitory colistin) precedes encapsulation, because 40,000 imaged droplets cannot enumerate a one-in-a-million subpopulation directly; enrichment amplifies the resistant fraction to the droplet assay's dynamic range, and this amplification is modeled explicitly so the assay back-calculates the original fraction.
The second component is a pH-panel phenotyping protocol: every isolate is profiled at external pH 5.5, 6.5, and 7.4, with and without colistin, yielding a per-cell four-condition fingerprint of acidification trajectories. The differential between drug-free and drug-exposed trajectories at each pH is the resistance signal, which controls for isolate-to-isolate variation in baseline metabolism.
The third component is computational, and it closes the loop back to bulk clinical practice. A bulk broth under colistin has a pH trajectory that is an arithmetic mixture of single-cell proton-flux phenotypes — precisely the averaging problem that the deconvolution literature attacks for gene expression. Count Bridges enable Modeling and Deconvolving Transcriptomic Data models bulk mixtures as discrete counting processes and recovers integer single-cell profiles; we borrow its commitment to discrete, mechanistically grounded mixture modeling, but for proton fluxes rather than transcript counts. Writing the per-cell acidification trajectory of phenotype $k$ as $\phi_k(t)$, its fraction as $\pi_k$, and the buffer's proton-to-pH mapping as $g$, the bulk signal is
$$\text{pH}_{\text{bulk}}(t) = g^{-1}\!\left(\sum_k \pi_k\, \phi_k(t)\right),$$
and the model infers $\pi_k$ — including the resistant fraction — from ordinary broth kinetics, using the droplet-derived $\phi_k$ as its reference dictionary. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level shows how far a good single-cell reference can push bulk attribution (up to 73% improvement in cross-cohort silhouette score on cancer atlases), and its acknowledged weakness — dependence on reference-atlas completeness — is why we build the bacterial reference dictionary ourselves rather than importing one.
Proposed Experiments
Phase one is feasibility on reference strains: a colistin-susceptible E. cloacae and a confirmed heteroresistant clinical isolate, measured per-droplet over six hours. The decisive number is signal-to-noise of the per-bacterium acidification rate against the dye's noise floor, benchmarked against the 0.17 pH units per three hours that SCO-pH resolved in single mammalian cells (SCO‐ph: Microfluidic Dynamic Phenotyping Platform for High‐Throughput Screening of Single Cell Acidification). If bacteria in the redesigned droplets produce less than three standard deviations of distinguishable pH change in six hours, the phenotypic arm fails and the project retreats to the bulk-deconvolution arm alone.
Phase two is detection on a clinical cohort of roughly two hundred non-repetitive E. cloacae complex isolates, mirroring the 212-strain Chongqing collection (Clinical characteristics, molecular epidemiology and mechanisms of colistin heteroresistance in Enterobacter cloacae complex). Ground truth comes from Population Analysis Profiling, the diagnostic gold standard; metrics are sensitivity, specificity, and area under the receiver operating curve against PAP, plus time-to-call. The clinical base rate is known: 24.4% of isolates that bulk automated testing calls sensitive are actually heteroresistant, subpopulation minimum inhibitory concentrations reach at least 128 mg/L, and high-risk lineages ST116 and ST56 carry heteroresistance in 75% of strains — all from that same paper, and all numbers our assay must reproduce. Success is sensitivity above 90% at specificity above 95% within eight hours.
Phase three runs the pH panel on the same cohort to ask whether any isolate shows pH-conditional heteroresistance — resistant only at acidic pH — which would mean standard neutral-pH testing is blind to exactly the infections (urinary tract, abscess) where colistin is deployed. As an exploratory arm, we apply EpHod, the sequence-based enzyme-optimum-pH predictor (Machine learning prediction of enzyme optimum pH), to the ArnA and β-lactamase sequences of each isolate and test whether predicted pH optima correlate with the observed pH-conditional phenotypes. EpHod reports a root mean squared error of 1.27 pH units with $R^2$ of 0.68 on 9,855 BRENDA enzymes, and it models purified enzymes in homogeneous bulk assays — so agreement here would be an early, coarse signal that sequence predicts in-situ pH-dependence; disagreement would be informative too, since periplasmic microenvironments are exactly what bulk predictors miss.
Phase four evaluates the bulk deconvolution model: droplet trajectories are mixed in silico at known resistant fractions from $10^{-3}$ down to $10^{-6}$ to generate ground-truth bulk curves, and the model's recovered fraction is compared against truth, reporting the limit of detection.
Baselines
The naive reference point is doing what clinics do now: standard automated bulk antimicrobial susceptibility testing. It is the strongest competitor on throughput and cost, and the ECC cohort study quantifies exactly what it misses — 24.4% of colistin-sensitive-classified isolates harbored resistant subpopulations with MICs at or above 128 mg/L (Clinical characteristics, molecular epidemiology and mechanisms of colistin heteroresistance in Enterobacter cloacae complex). Every isolate in our cohort gets a bulk AST call, and beating this baseline means flagging a strict superset of the isolates bulk AST misclassifies, with sensitivity above 90% at specificity above 95%, at comparable per-sample cost. The strongest existing method is Population Analysis Profiling, the manual plating gold standard used in that same paper, which detects the subpopulations reliably but needs 48 to 72 hours of incubation and hands-on work. We run PAP on every isolate ourselves — it supplies the ground-truth labels — so the comparison is on exactly the same strain set with no difference in isolate handling; beating PAP means matching its sensitivity and specificity while cutting time-to-call from two to three days to under eight hours. The key ablation isolates this proposal's central ingredient, the single-cell reference dictionary: a generic two-population mixture model fitted to the same bulk pH and optical-density curves without any droplet-derived $\phi_k$ kinetics, tuned with the same compute and hyperparameter budget. If the dictionary-free mixture matches the dictionary-guided model's limit of detection, the entire microfluidic program is unjustified overhead and the paper should say so. A second ablation drops the pH panel, running the assay only at pH 7.4, to isolate the value of pH conditioning. For the exploratory sequence arm, the natural competitor is ARGGNN, which classifies resistance genes from structure and sequence with macro F1 of 63.78% versus 59.7% for the best prior method on COALA (Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks); it predicts categorical resistance labels, not pH-dependent kinetics, so the comparison is informative rather than direct. No reported result found on Lacuna for any rapid phenotypic assay that detects heteroresistance within a single working shift.
Why not just…?
Why not just do Population Analysis Profiling? Because its own users flag the latency as the open problem: the ECC paper notes the 48–72 hour manual workflow and no rapid alternative exists in the reviewed literature. In septic patients, colistin dosing decisions are made in the first day, exactly when the hidden subpopulation is invisible. PAP also measures a colony count at a fixed time, discarding all kinetic information, so it cannot see pH-conditional phenotypes at all.
Why not just sequence single cells? Single-cell sequencing of the gut microbiome already resolves resistance genes per cell and found divergent resistomes in indistinguishable K. pneumoniae strains (Elucidation of population-based bacterial adaptation to antimicrobial treatment by single-cell sequencing analysis of the gut microbiome of a hospital patient), but it measures DNA, not function: gene presence does not tell you the enzyme's turnover rate under drug, its pH dependence, or whether the cell expressing it actually survives colistin. That study also measured no environmental pH and drew from one patient. Our assay is the phenotypic complement — and phase two includes whole-genome sequencing of every isolate so we can directly test how often genotype and pH-phenotype disagree, which is itself a result.
Why not just run the existing SCO-pH platform on bacteria unchanged? This is the most serious objection. Bacteria are an order of magnitude smaller than the EL4 lymphoma cells SCO-pH was built for, produce far fewer protons, and the platform already needed weakly buffered medium and still carried a 16.5% identification error rate (SCO‐ph: Microfluidic Dynamic Phenotyping Platform for High‐Throughput Screening of Single Cell Acidification). A naive port plausibly fails at phase one. That is why feasibility is the first gate, why the droplet geometry and dye are redesigned rather than reused, why the enrichment step exists, and why the bulk-deconvolution arm is designed to stand alone if single-bacterium pH resolution proves physically out of reach.
Why not just predict everything from sequence? EpHod predicts enzyme optimum pH from sequence with useful accuracy (RMSE 1.27 pH units, $R^2$ 0.68) but only for purified enzymes in bulk assays (Machine learning prediction of enzyme optimum pH), and ARGGNN assigns static resistance categories without kinetics or protonation states (Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks). Neither models per-cell expression heterogeneity — the very thing this proposal targets. We keep the sequence arm exploratory and explicitly secondary for this reason.
Decisive Checks
Three results decide the proposal's fate. The per-bacterium acidification signal-to-noise ratio in phase one must exceed three, or the single-cell arm is abandoned. Phase two must reproduce the known epidemiology — the 24.4% hidden-heteroresistance rate among bulk-susceptible isolates, and the MIC-at-or-above-128 mg/L magnitude of the subpopulations — within confidence intervals, or the assay's calls cannot be trusted against PAP ground truth. And the dictionary-guided deconvolution must beat the dictionary-free ablation on limit of detection by a margin that justifies the microfluidics; if it does not, the honest paper is the mixture model on bulk curves alone, and that paper can still be written.
References
- Clinical characteristics, molecular epidemiology and mechanisms of colistin heteroresistance in Enterobacter cloacae complex
- Drug Resistance Patterns and Pathogen Distribution in Patients with Urinary Tract Infection in Suzhou, China: A Five-Year Retrospective Study (2020–2024)
- SCO‐ph: Microfluidic Dynamic Phenotyping Platform for High‐Throughput Screening of Single Cell Acidification
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- Elucidation of population-based bacterial adaptation to antimicrobial treatment by single-cell sequencing analysis of the gut microbiome of a hospital patient
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Machine learning prediction of enzyme optimum pH
- Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks