When bacteria settle on a surface and encase themselves in a sticky extracellular matrix, their tolerance to antibiotics jumps by a factor of 100 to 1,000 — without any genetic mutation. Two competing explanations dominate the literature. The first is purely physical: the matrix electrostatically traps positively charged drugs and slows their diffusion inward, so the cells never see a lethal dose. The second is metabolic: cells in oxygen- and nutrient-starved interior layers deliberately slow their growth through stress-signaling programs, and most antibiotics simply cannot kill cells that are not growing. Both stories are well supported, but no study has measured how much of the tolerance each one actually carries in a living biofilm, because bulk assays mix the two effects into a single number. This proposal builds a pipeline that splits them apart: it maps spatially resolved transcriptomes of a two-species biofilm into weighted co-expression modules, uses those modules to predict local antibiotic survival, and then knocks out the hub genes of the best-predicting modules to measure, causally, how much tolerance each metabolic program contributes. Functional validation is not an afterthought here; it is the measuring instrument.
The clinical stakes are concrete. Biofilm tolerance is phenotypic and reversible, which makes it a tempting drug target — but only if the metabolic component is large. If tolerance is mostly a passive physical barrier, metabolic-targeting therapies are misdirected effort, and anti-matrix or dispersal strategies are the right investment. The literature gives evidence on both sides without settling the split: peptide 1018 clears mature biofilms by degrading the stress alarmone (p)ppGpp, and (p)ppGpp-null mutants fail to form biofilms at all, which argues the metabolic side is load-bearing Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response. On the other side, cells wrapped in an exogenously induced poly-N-acetylglucosamine matrix show a 10-fold tolerance boost to polymyxin B within one hour — faster than any transcriptional program could rewire metabolism — and deleting the LPS genes that anchor the matrix abolishes both aggregation and that tolerance Genetic Dissection of an Exogenously Induced Biofilm in Laboratory and Clinical Isolates of E. coli. Filamentous Pf phages secreted by P. aeruginosa act as pure decoys that soak up cationic antibiotics Pf Bacteriophage and Their Impact on Pseudomonas Virulence, Mammalian Immunity, and Chronic Infections, and curli fibers measurably retard oxidant penetration Elucidating the Role of Virulence Traits in the Survival of Pathogenic E. coli PI-7 Following Disinfection. The split is genuinely unresolved, and it determines therapeutic strategy.
Core Claim
The central hypothesis is that the metabolic share of biofilm tolerance is concentrated in a small number of co-expression modules — coordinated gene teams whose activity predicts antibiotic survival better than any single gene — and that deleting the hub genes of those modules removes a quantifiable, additive fraction of the tolerance gap between biofilm and planktonic cells. Weighted Gene Co-expression Network Analysis (WGCNA — a method that clusters genes by continuous, softly thresholded co-expression rather than hard cutoffs, then summarizes each cluster by its first principal component, the "module eigengene") is the right instrument because it reduces thousands of stress-responsive genes to a handful of continuous scores that can be correlated directly with local survival measurements, avoiding the multiple-testing swamp of single-gene statistics. Metabolic pathway inference — imputing community metabolic potential from 16S marker-gene surveys via a reference genome-function lookup matrix — supplies the functional annotation that makes the modules interpretable and gives us a cheap, independent read of each stratum's metabolic capacity. The functional validation requirement is the epistemic bar: a module counts as a cause only if its hub genes pass loss-of-function and gain-of-function tests, a standard that biomarker papers across oncology and cardiovascular research repeatedly show is skipped at the field's peril.
Research Program
The experimental system is deliberately chosen to match the strongest existing biofilm model so the comparison is apples-to-apples: P. aeruginosa PAO1 and S. aureus N315, mono-species and mixed consortia, grown in flow-cell biofilms under controlled nutrient supply. This is exactly the organism pair of the dynamic flux balance analysis model built on the genome-scale reconstructions iMO1056 and iMH551, which predicts that oxygen and glucose gradients partition the two species vertically, with S. aureus dominating the anaerobic glucose-rich base and P. aeruginosa occupying the oxygenated surface, reaching a mixed maximum thickness of 80 µm Metabolic modeling of a chronic wound biofilm consortium predicts spatial partitioning of bacterial species. We grow the real counterpart of that simulation and add what it cannot do: empirical transcriptomics and perturbation.
Three data layers are collected at 6, 24, 48, and 72 hours of development, in untreated biofilms and under a three-agent antibiotic panel designed to stress the two competing stories differently — polymyxin B (a cationic peptide that the matrix should trap), ciprofloxacin (a fluoroquinolone whose efficacy should depend on growth state), and peptide 1018 (the stringent-response disruptor). First, spatially resolved RNA-seq: biofilms are cryosectioned into upper and lower halves, giving us the transcriptome of each stratum separately so that co-expression structure can differ across microenvironments instead of being averaged away. Second, 16S amplicon sequencing of parallel communities over time, feeding the standard taxon-abundance-times-reference-matrix imputation of Kyoto Encyclopedia of Genes and Genomes (KEGG) Ortholog abundances per stratum — the same predictive-profiling logic used to link gut taxa to shifted pathway content in metformin-treated mice Metformin Alters Gut Microbiota of Healthy Mice: Implication for Its Potential Role in Gut Microbiota Homeostasis and in chronic hepatitis B patients Gut Microbial Dysbiosis Is Associated with Altered Hepatic Functions and Serum Metabolites in Chronic Hepatitis B Patients. Third, the phenotype itself: log-reduction in colony-forming units per stratum per treatment, plus matrix density and antibiotic penetration profiles measured with fluorescent drug analogues.
The computational core runs WGCNA on the pooled gene-by-sample expression matrix across strata, conditions, and time points, using soft thresholding to a scale-free topology and the topological overlap measure to guard modules against spurious pairwise correlations. For the mixed consortium, each species is treated as a "view" and a multi-view decomposition in the spirit of MVTLASSO — a heavy-tailed multi-cohort co-expression method shown to separate shared networks from view-specific noise on Bacillus subtilis compendia Robust Multi-view Co-expression Network Inference — splits modules into species-specific programs and genuinely interspecies modules, the latter being the untested territory where lactate cross-feeding is predicted but never measured transcriptionally Metabolic modeling of a chronic wound biofilm consortium predicts spatial partitioning of bacterial species. Module eigengenes are then regressed against local survival to rank modules by predictive power, and the imputed pathway profiles serve as an independent functional cross-check on what each module is doing.
The validation phase is where bystanders are separated from drivers. For the top-ranked metabolic modules, we knock out hub genes (CRISPR or clean deletions) and re-run the full assay, measuring the causal effect on survival as
$$\Delta_g = \log_{10}\mathrm{CFU}_{g^{-}} - \log_{10}\mathrm{CFU}_{g^{+}},$$
the difference in viable-cell log-reduction between the knockout and wild type under identical antibiotic exposure, normalized so that $\Delta_g = 0$ means the gene contributed nothing to tolerance and $\Delta_g$ equal to the full biofilm-versus-planktonic gap means the gene carried all of it. Two positive controls anchor the scale: $\Delta relA/\Delta spoT$, which should impair biofilm maturation per the (p)ppGpp evidence Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response, and $\Delta rfaY/\Delta rfaP$, which should abolish matrix anchoring and its 10-fold polymyxin tolerance Genetic Dissection of an Exogenously Induced Biofilm in Laboratory and Clinical Isolates of E. coli. Candidate metabolic hubs include S. aureus lactate dehydrogenase (the modeled cross-feeding entry point) and P. aeruginosa pyocyanin biosynthesis. Gain-of-function runs the test in reverse: overexpressing a hub in planktonic, matrix-free cells should confer part of the tolerance if the gene is a driver, and should do nothing if it is a passenger of the stringent response — the same bidirectional logic that biomarker studies in breast cancer and T-cell markers explicitly demand Evaluating the role of sirtuins in breast cancer, and whose absence leaves mechanisms unconfirmed, as the FAM83A pancreatic-cancer study concedes directly ncRNA-mediated upregulation of FAM83A is associated with poor prognosis and immune infiltration in pancreatic cancer. This bidirectional bar is exactly what the WGCNA literature most often fails to meet: the RNF181 coronary-artery-disease study validated in mouse models and GWAS but never perturbed the gene in isolated cells Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease, and the acute-myocardial-infarction signature stopped at immunohistochemistry Identification of clinical diagnostic and immune cell infiltration characteristics of acute myocardial infarction with machine learning approach. We build the perturbation in from the start.
Baselines
The naive reference point is planktonic minimum inhibitory concentration testing — the standard clinical susceptibility assay — run on the same strains in the same medium. It anchors the tolerance gap we are decomposing (the well-documented 100- to 1,000-fold biofilm boost) and defines "doing nothing" for prediction: no model, tolerance simply unmeasured in the state that matters.
The strongest existing competitor for the in silico part is the dynamic flux balance analysis spatial model of the same P. aeruginosa–S. aureus consortium Metabolic modeling of a chronic wound biofilm consortium predicts spatial partitioning of bacterial species, which couples the iMO1056 and iMH551 reconstructions to diffusion equations and reports a mixed-species maximum biofilm thickness of 80 µm, versus 30 µm for P. aeruginosa alone and 90 µm for S. aureus alone. It is the strongest because it is the only existing model that makes spatially resolved, species-level metabolic predictions in this exact system. We will run it on our measured biofilm geometries and uptake parameters and score it on our data; the paper reports no survival-prediction accuracy (no reported result found on Lacuna for antibiotic-survival prediction accuracy from that model), so the comparison metric is set by us: cross-validated error in predicting per-stratum log-CFU reduction. Beating it means our empirically learned co-expression modules predict local survival with materially lower error than the literature-parameterized simulation, which never observed a single transcript.
For the prediction subtask on community composition, the strongest published method is the graph neural network that encodes species as nodes with KEGG Orthology presence/absence features and predicts steady-state abundances, reporting an $R^2$ of 0.81 on the FRIEDMAN2017 synthetic community benchmark and 0.77 on BARANWALWCLARK2022 Modelling Microbial Communities with Graph Neural Networks. We adapt it as a baseline for predicting stratum-level species proportions from imputed functional profiles, retrained on our data with the same split structure, compute, and tuning budget as our eigengene regressors. Beating it means higher $R^2$ on our biofilm benchmark than its 0.81 reference on well-mixed communities — or, if it transfers poorly, a quantified demonstration that spatial structure breaks functional-feature-only prediction, which is itself a finding.
The single-gene baseline is differential expression analysis followed by per-gene survival prediction, in the style of the myocardial-infarction pipeline where individual genes from a 159-gene differential set achieved diagnostic AUCs between only 0.56 and 0.83 while the module-based signature reached 0.932 and 0.953 Identification of clinical diagnostic and immune cell infiltration characteristics of acute myocardial infarction with machine learning approach. We run it with matched feature budgets; beating it means module eigengenes predict survival with cross-validated error below the best single-gene and best greedy-gene-set predictors at equal size.
Finally, the key ablation isolates the proposal's central ingredient: the same WGCNA pipeline run with hard correlation thresholding ($r > 0.8$) instead of soft thresholding, and a second variant with taxonomy-only features instead of imputed pathway features. If the soft-thresholded, pathway-informed version does not beat both ablations on survival prediction, the specific machinery of this proposal — not just "more transcriptomics" — is doing no work, and we will say so.
Why not just…?
Why not just run differential expression instead of building co-expression networks? Because single-gene statistics cannot carry the causal decomposition this proposal is about. The myocardial-infarction study is the cleanest internal evidence: individual differentially expressed genes reached AUCs of 0.56–0.83, while the module-derived signature reached 0.932 and 0.953 Identification of clinical diagnostic and immune cell infiltration characteristics of acute myocardial infarction with machine learning approach, and the coronary-artery-disease work needed module topology to isolate one hub from 8,525 candidates Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease. More fundamentally, biofilm stress shifts thousands of genes at once — a sublethal antimicrobial alone induces the master matrix regulator csgD and whole curli programs Biofilm-mediated resistance to berberine in Escherichia coli — so a gene list cannot tell us which coordinated program matters. The single-gene baseline is included anyway, with matched budgets, precisely so this claim is tested rather than asserted.
Why not just simulate with dynamic flux balance analysis? Because the existing model, strong as it is, is a deterministic extrapolation from literature uptake kinetics and diffusion coefficients — its own acknowledged limitations — with no transcriptomic input, no stress-response machinery, and no perturbation experiment to check a single prediction Metabolic modeling of a chronic wound biofilm consortium predicts spatial partitioning of bacterial species. Its steady-state growth-maximization objective is actively wrong for biofilms, where the stringent response pushes cells into non-proliferative persistence that flux balance cannot represent Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response. It also has no vocabulary for physical barriers like Pf-phage antibiotic sequestration Pf Bacteriophage and Their Impact on Pseudomonas Virulence, Mammalian Immunity, and Chronic Infections. That said, this is a real threat rather than a strawman: if the dFBA model, refit on our measured parameters, predicts stratum survival as well as our learned modules, then the empirical machinery is unnecessary and the field should invest in better simulations. The head-to-head in the baselines section settles it either way.
Why not just do the computation and skip the knockouts? Because computational signatures in stressed biological systems systematically mix drivers with bystanders, and the biofilm literature already shows the confound: a broad transcriptional program (the stringent response) and a chemical-exposure response (csgD induction) both shift hundreds of genes, and only perturbation — the $\Delta relA/\Delta spoT$ mutants, the (p)ppGpp-overproduction rescue, the $^{31}$P-NMR binding assay — established which molecule the peptide actually targeted Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response. In the biomarker world the same lesson repeats: the GGT5 immunotherapy study validated spatially with multiplex staining but never inhibited the enzyme, so causation remains open GGT5: a potential immunotherapy response inhibitor in gastric cancer by modulating GSH metabolism and sustaining memory CD8+ T cell infiltration. Our headline quantity — the causal fraction $\Delta_g$ — is simply not defined without the wet-lab phase; the computation ranks candidates, the knockouts measure them.
Why not just measure penetration and argue from physics? This is the honest danger. If tolerance were fully explained by matrix trapping, drug-penetration profiles plus the $\Delta rfaY$ control would tell the whole story and the transcriptomic apparatus is decoration. We therefore build the decisive check around it: gain-of-function overexpression of top hub genes in planktonic, matrix-free cells. If overexpression confers measurable tolerance with no matrix present, the metabolic program has an effect physics cannot explain. If it confers nothing and only the matrix deletions move survival, the physical story wins, the metabolic modules are bystanders, and the paper reports that as its result — a genuinely useful negative for anyone planning anti-metabolite biofilm therapies.
Decisive Checks
Three outcomes decide the proposal. First, prediction: module eigengenes must beat the single-gene baseline and both ablations (hard thresholding, taxonomy-only features) on cross-validated survival prediction, and must match or beat the refit dFBA model; the GNN's 0.81 $R^2$ on well-mixed communities Modelling Microbial Communities with Graph Neural Networks is the external reference bar for the composition subtask, with no reported result found on Lacuna for co-expression-based prediction of antibiotic survival in biofilms. Second, causation: at least one metabolic module's hub knockouts must remove a reproducible fraction of the biofilm–planktonic tolerance gap without impairing planktonic growth, and its gain-of-function must confer tolerance matrix-free — the bidirectional standard that separates drivers from passengers. Third, attribution: the sum of module-attributable fractions plus the matrix-control fraction must account for most of the measured tolerance gap, giving the field its first quantitative answer to how biofilm resistance is actually divided between physics and metabolism — whichever way the answer falls.
References
- Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response
- Genetic Dissection of an Exogenously Induced Biofilm in Laboratory and Clinical Isolates of E. coli
- Pf Bacteriophage and Their Impact on Pseudomonas Virulence, Mammalian Immunity, and Chronic Infections
- Elucidating the Role of Virulence Traits in the Survival of Pathogenic E. coli PI-7 Following Disinfection
- Metabolic modeling of a chronic wound biofilm consortium predicts spatial partitioning of bacterial species
- Metformin Alters Gut Microbiota of Healthy Mice: Implication for Its Potential Role in Gut Microbiota Homeostasis
- Gut Microbial Dysbiosis Is Associated with Altered Hepatic Functions and Serum Metabolites in Chronic Hepatitis B Patients
- Robust Multi-view Co-expression Network Inference
- Altered sirtuin expression is associated with node-positive breast cancer
- ncRNA-mediated upregulation of FAM83A is associated with poor prognosis and immune infiltration in pancreatic cancer
- Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease
- Identification of clinical diagnostic and immune cell infiltration characteristics of acute myocardial infarction with machine learning approach
- Modelling Microbial Communities with Graph Neural Networks
- Biofilm-mediated resistance to berberine in Escherichia coli
- GGT5: a potential immunotherapy response inhibitor in gastric cancer by modulating GSH metabolism and sustaining memory CD8+ T cell infiltration