When a patient's infecting bacterium tests susceptible to an antibiotic in the lab, doctors expect the drug to work. Often it does not. In a Singapore primary-care cohort of 299 culture-positive urinary tract infections, patients whose uropathogen was non-susceptible to the prescribed drug still cleared their symptoms within three days 45% of the time, and for ciprofloxacin the early-resolution rate was actually higher among non-susceptible than susceptible infections (54% vs 48%, P = 0.762) (Interaction between Antibiotic Resistance, Resistance Genes, and Treatment Response for Urinary Tract Infections in Primary Care). This genotype-to-outcome disconnect is one of the core open problems of the antibiotic resistance crisis: current prediction models read the pathogen's DNA (GenoARM, ARG-BERT, ARGGNN) or the patient's chart (Affordable and real-time antimicrobial resistance prediction from multimodal electronic health records) but never look at what the host is actually doing at the tissue level, so they cannot see the mucosal immunity or inflammatory programs that override a resistance gene.
Functional enrichment analysis — the statistical practice of asking which curated biological pathways are over-represented among genes whose expression changed, rather than ranking single genes — has already shown that the host response to bacterial infection is a coherent pathway-level phenomenon: in a 180-patient ICU metatranscriptomic cohort, bacterial infection was marked by TNF and NF-κB signaling enrichment, viral infection by MHC assembly and respiration shifts, and a 21-gene pathway-grounded signature reached an independent-test AUC of 0.865 for diagnosing infection (Stacking ensemble learning models diagnose pulmonary infections using host transcriptome data from metatranscriptomics). But that study, and the sepsis multi-omics work before it (Integrative analysis of metabolomics and transcriptomics to uncover biomarkers in sepsis), lumped all bacterial infections together regardless of resistance, and never excluded patients with chronic inflammatory or pulmonary disease whose baseline biology contaminates the very pathways being measured. This proposal argues that the disconnect between in vitro susceptibility and clinical outcome is a pathway-level, host-side phenomenon, and that it can only be measured cleanly if we borrow two disciplines that have never been combined: exclusionary cohort filtering from observational clinical research, and resistance-stratified functional enrichment from host transcriptomics.
Core Claim
If we stratify infection cohorts by the pathogen's resistance status and the patient's clinical outcome, apply strict exclusionary filtering before analysis (removing chronic inflammatory disease, prior immunosuppression, and recent antibiotic exposure), and then run functional pathway enrichment on host transcriptomes, we will isolate a set of host pathways that (a) explain why a substantial fraction of patients infected with in vitro non-susceptible pathogens nonetheless recover, and (b) predict clinical treatment failure more accurately than pathogen genotype alone. The methodological claim is equally important: without the exclusionary filtering step, those same pathways are confounded by chronic baseline disease, as the biomarker literature has repeatedly demonstrated for acute-versus-chronic signals (Clinical significance of cytokeratin 19 fragment in COVID-19 patients: a retrospective study).
The role each input direction plays is deliberate. Confounder-controlled cohort selection is the methodological lens — the soundproof booth that makes host pathway signals interpretable. The antibiotic resistance crisis is the application regime — the clinically urgent setting where genotype and outcome diverge. Functional enrichment analysis is the core mechanism — the analytical instrument that turns noisy host transcriptomes into pathway-level readouts.
Research Program
The program has three linked components, each building directly on methods already validated in the literature rather than inventing new architectures.
First, we assemble a linked cohort that joins host metatranscriptomics with pathogen susceptibility phenotypes and clinical outcomes. The seed is the 180-patient ICU metatranscriptomic cohort from Stacking ensemble learning models diagnose pulmonary infections using host transcriptome data from metatranscriptomics, whose 91 bacterial infections we will stratify retrospectively by available culture and susceptibility records into susceptible and non-susceptible groups, and by early treatment response. We will supplement this with public sepsis transcriptomic datasets already used for validation in Integrative analysis of metabolomics and transcriptomics to uncover biomarkers in sepsis (GSE6535, GSE12624, GSE28750, GSE63042, GSE74224) and with the 299-patient primary-care UTI cohort (Interaction between Antibiotic Resistance, Resistance Genes, and Treatment Response for Urinary Tract Infections in Primary Care), which contributes the richest available resistance-versus-outcome discordance data even though it lacks transcriptomics — it anchors the phenotype definition and provides the calibration target for outcome modeling. We acknowledge up front that no single existing dataset carries all three modalities; the assembly of that linked cohort, under a pre-registered filtering and analysis protocol, is itself a contribution of this work.
Second, we build the exclusionary filtering protocol as a first-class, reproducible artifact. Following the design logic of Clinical significance of cytokeratin 19 fragment in COVID-19 patients: a retrospective study — where CYFRA 21-1 was only interpretable as an acute lung-injury marker after malignancies and pre-existing structural organ disease were removed — we define a staged exclusion cascade: baseline chronic inflammatory and pulmonary disease (COPD, asthma, autoimmune disease), chronic organ failure, immunosuppressive and long-term immunomodulatory therapy, and antibiotic exposure in a defined pre-enrollment window. Crucially, we will run every analysis twice, once on the filtered cohort and once on the full cohort with only regression adjustment, so the cost and the benefit of the filtering decision are measured rather than assumed. This directly addresses the known open problem that exclusionary filtering trades internal validity for sample size and generalizability.
Third, we run the resistance-stratified functional enrichment itself. Host DEGs within each (resistance × outcome) stratum are mapped to Gene Ontology Biological Process and KEGG pathway terms via overrepresentation analysis, implemented on the distributed Spark-based pipeline of funkea: Functional Enrichment Analysis in Python, which already supports multiple enrichment formulations and scales to large gene sets. Because pathway terms overlap heavily, we will apply leading-edge and term-clustering approaches to collapse redundant annotations and identify primary driver programs. The key analytic question is whether host pathways separate the four strata: the biologically interesting prediction is that patients who fail treatment despite susceptible pathogens — the blind spot of every existing AMR model — will show enrichment in pathways distinct from canonical TNF/NF-κB bacterial-response programs, plausibly involving tissue-barrier and mucosal immune programs. Finally, we convert the enriched pathways into a pathway-level predictor of treatment failure and test whether it adds predictive value over genotype-only models.
Baselines
The naive reference point is pathogen genotype alone: a lookup-style classifier that predicts clinical failure whenever a resistance determinant is present, calibrated on the primary-care UTI cohort, where susceptible infections resolved early in 67% of cases versus 45% for non-susceptible ones (P = 0.001) (Interaction between Antibiotic Resistance, Resistance Genes, and Treatment Response for Urinary Tract Infections in Primary Care). This baseline makes no host assumptions and directly exposes how much of the outcome the genotype explains; any host-pathway model must beat its resolution prediction by a margin that survives bootstrap confidence intervals on the same held-out patients.
The strongest existing method on this question is the stacking-ensemble host-transcriptome diagnostic of Stacking ensemble learning models diagnose pulmonary infections using host transcriptome data from metatranscriptomics, which reports an independent-test AUC of 0.865 overall and 0.871 for bacterial infection specifically using a 21-gene signature. It is the strongest competitor because it is the only published model that already uses host transcriptomes from an infected ICU cohort at scale. We will retrain it from its published signature genes under identical data splits, identical tuning budget, and identical compute, but with our stratified outcome label (treatment failure) instead of its infection-type label. Beating it means achieving a higher held-out AUC for failure prediction than the same architecture achieves when aimed at the easier diagnostic task, while also showing the pathway-level enrichment that the ensemble's black-box gene signature cannot provide.
A third baseline anchors the EHR-modality claim: the multimodal MMIM model of Affordable and real-time antimicrobial resistance prediction from multimodal electronic health records, which achieved an AUROC of 0.7565 but an AUPR of only 0.1825 for Gentamicin resistance on MIMIC-IV. Its precision collapse under class imbalance sets the bar a pathway model must clear on the precision-recall side, where resistance prediction actually hurts clinically.
The decisive ablation isolates the proposal's key ingredient by running the identical enrichment pipeline on the full unfiltered cohort versus the exclusion-filtered cohort. If the enriched pathways shift substantially between the two — for example, if chronic pulmonary patients drive spurious enrichment in remodeling or m6A-related regulatory pathways, which the COPD literature suggests they will (m6A RNA methylation regulators could contribute to the occurrence of chronic obstructive pulmonary disease) — then filtering is load-bearing. If the pathways are stable under filtering, the filtering step is unnecessary and we will report that honestly, because the simpler analysis then suffices.
Why not just…?
Why not just adjust for confounders with regression or propensity scores on the full cohort? This is the strongest objection, and the literature supplies concrete evidence against it in the biomarker setting: in Clinical significance of cytokeratin 19 fragment in COVID-19 patients: a retrospective study, post-hoc adjustment could not rescue CYFRA 21-1 because pre-existing malignancy changes the biomarker's baseline production entirely, not additively — only upfront exclusion made the signal interpretable. On the other side, the AUPR of 0.1825 achieved by a sophisticated deep model on an unfiltered ICU cohort (Affordable and real-time antimicrobial resistance prediction from multimodal electronic health records) shows statistical adjustment does not automatically rescue precision. Still, we treat this as an empirical question rather than a settled one: the full-versus-filtered ablation described above is designed precisely so that, if adjustment performs as well as exclusion, the finding is that the simpler standard approach suffices.
Why not just use top differentially expressed genes or a supervised gene classifier instead of pathway enrichment? Individual-gene classifiers exist and work: the 21-gene stacking ensemble is our strongest baseline precisely for that reason. But the evidence from the same literature shows why pathways are needed on top of genes. In Integrative analysis of metabolomics and transcriptomics to uncover biomarkers in sepsis, 485 individual DEGs and 1083 metabolites became interpretable only when the 40 cross-omic concordant molecules were mapped to amino-acid-metabolism and macrophage pathways, which is what localized ITGAM and C3AR1 to the right cell type. And in Optimizing Gene-Based Testing for Antibiotic Resistance Prediction, testing 15 optimized genes performed nearly the same as ~200 genes because resistance determinants fall into redundant functional classes — a warning that gene-level selection overfits to redundant markers. Our design concedes the predictive contest to the gene-signature baseline on its own terms and asks the additional question the signature cannot answer: which biological program distinguishes treatment failure, which is what transfers across cohorts and mechanistic hypotheses.
Why not just look up resistance genes in curated databases (BLAST, CARD) and be done? Database lookup fails on two fronts documented in the file. Under sequence divergence, BLAST degraded to 0.917 accuracy at 40% sequence identity while ARG-BERT held 0.927 (Prediction of antibiotic resistance mechanisms using a protein language model), and ARGGNN improved macro F1 from 59.7% to 63.78% on the COALA benchmark (Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks) — but the more fundamental failure is that these models predict the wrong target. Even a perfect genotype classifier cannot explain the 45% of non-susceptible UTI patients who cleared their infections anyway, or the ciprofloxacin inversion (48% vs 54%, P = 0.762) (Interaction between Antibiotic Resistance, Resistance Genes, and Treatment Response for Urinary Tract Infections in Primary Care). Our first baseline is the lookup-style classifier, so if genotype alone turns out to predict outcomes as well as host pathways do, the experiments will demonstrate that and the proposal's motivation will have been refuted.
Proposed Experiments and Decisive Checks
The experimental sequence follows the cohort program. Experiment one runs the exclusionary filtering protocol on the 180-patient ICU metatranscriptomic cohort, then performs enrichment separately for the filtered and unfiltered versions within each resistance-outcome stratum. The pre-registered check is pathway-set overlap: if the Jaccard overlap between filtered and unfiltered enriched-pathway sets is high, filtering is unnecessary; if it is low, and if the unfiltered set shows chronic-disease pathway contamination consistent with COPD-driven remodeling (m6A RNA methylation regulators could contribute to the occurrence of chronic obstructive pulmonary disease), the filtering claim is supported. Experiment two tests discordance directly: enrichment run on patients with non-susceptible pathogens who recovered versus those who did not. The decisive result would be a host pathway program enriched specifically among the recovering discordant patients, absent from canonically defined bacterial-response pathways (TNF, NF-κB) as characterized in Stacking ensemble learning models diagnose pulmonary infections using host transcriptome data from metatranscriptomics. Experiment three turns the enriched pathways into a pathway-level failure predictor, evaluated against the genotype-only baseline and the retrained 21-gene stacking ensemble (whose independent-test AUC of 0.865 sets the bar), with precision-recall reported alongside AUC so the comparison to the MMIM AUPR of 0.1825 is on the metric that matters clinically.
A final check addresses generalizability head-on, since filtering cohorts shrinks them. The pathway-level predictor will be evaluated on the unfiltered held-out remnant — the patients the protocol excluded — to measure how much internal validity costs external validity. If the pathway signature transfers to multimorbid patients, we get the best of both; if it does not, we will quantify the boundary of validity explicitly, which is itself an unanswered question in the exclusionary-filtering literature. No reported result exists on Lacuna for host-pathway enrichment performed on resistance-stratified, exclusion-filtered infection cohorts, which is the gap this proposal fills.
References
- Interaction between Antibiotic Resistance, Resistance Genes, and Treatment Response for Urinary Tract Infections in Primary Care
- Optimizing Gene-Based Testing for Antibiotic Resistance Prediction
- Prediction of antibiotic resistance mechanisms using a protein language model
- Structure-Aware Antibiotic Resistance Classification using Graph Neural Networks
- Affordable and real-time antimicrobial resistance prediction from multimodal electronic health records
- Stacking ensemble learning models diagnose pulmonary infections using host transcriptome data from metatranscriptomics
- Integrative analysis of metabolomics and transcriptomics to uncover biomarkers in sepsis
- Clinical significance of cytokeratin 19 fragment in COVID-19 patients: a retrospective study
- funkea: Functional Enrichment Analysis in Python
- m6A RNA methylation regulators could contribute to the occurrence of chronic obstructive pulmonary disease