When bacteria form a biofilm — a surface-attached community sealed in a self-made slime matrix — their tolerance to antibiotics jumps by 100- to 1,000-fold, often without any resistance mutation at all. Clinicians nonetheless decide therapy using MIC assays run on free-floating cells, and genomic resistance prediction models are trained on the same planktonic measurements. Meanwhile, human genetics spent a decade on the "missing heritability" problem: twin studies say a trait is largely genetic, yet the individual DNA variants found by standard scans explain almost none of it. This proposal imports that toolkit to the one bacterial phenotype where it has never been applied: the Minimum Biofilm Eradication Concentration (MBEC), the drug level needed to kill cells inside an established biofilm.
The empirical hook is a striking dissociation. In a cohort of 272 clinical Acinetobacter baumannii isolates, the most drug-resistant strains formed the thinnest biofilms (rank correlation $r_s = -0.720$ between planktonic resistance and biomass), yet every strain's biofilm raised the killing concentration by 16- to 2,048-fold over its planktonic MIC, and the size of that boost tracked the strain's genotype, not its biomass (Relationship between Antibiotic Resistance, Biofilm Formation, and Biofilm-Specific Resistance in Acinetobacter baumannii). That study had no genome sequences, so the question it begs is open: how much of the strain-to-strain variation in biofilm survival is heritable, and where in the genome does it hide? In planktonic Mycobacterium tuberculosis, heritability of resistance is high (64–88%) while known resistance loci explain only 0.01–0.24% of phenotypic variance — missing heritability in its purest bacterial form (GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions). Nobody knows whether biofilm tolerance shows the same gap, because nobody has measured it.
Core Claim
I claim that biofilm-specific survival is a heritable, polygenic, and substantially non-additive bacterial trait, and that its genetic signal is recoverable if — and only if — the analysis uses phenotypes and network priors that match biofilm biology. Concretely: the right phenotype is not MIC, not biomass, but the biofilm resistance amplification of each strain, defined as the log-ratio of MBEC to planktonic MIC. And the right model of gene-gene interaction is not a generic protein-interaction graph, but one curated to biofilm defense pathways: cyclic-di-GMP turnover, the stringent response, extracellular-matrix synthesis operons, and matrix-embedded decoys. If the claim is right, an epistatic model with biofilm-specific priors will predict held-out MBEC amplification where additive GWAS and planktonic-trained models do not, and will point to specific interacting regulatory loci that can be validated by paired knockouts.
Research Program
The program has three parts: a phenotyping campaign, a heritability partition, and an epistatic recovery model.
First, phenotyping. No existing dataset pairs whole-genome sequences with quantitative MBEC, so one must be built. The natural substrate is the global Pseudomonas aeruginosa collection — 9,829 genomes spanning cystic-fibrosis, hospital, and environmental sources, with 21 epidemic clones accounting for 51% of clinical infections (Evolution and host-specific adaptation of Pseudomonas aeruginosa). I would draw a stratified sample of roughly 1,500 isolates that over-samples clone-mates (to separate within-clone from between-clone variance) and measure, per isolate and per drug (starting with tobramycin and ciprofloxacin, the workhorses of CF therapy): planktonic MIC by broth microdilution; biofilm biomass by crystal violet staining at 550 nm; and MBEC using the peg-lid Calgary device, exactly the assay battery used in the A. baumannii cohort (Relationship between Antibiotic Resistance, Biofilm Formation, and Biofilm-Specific Resistance in Acinetobacter baumannii). Each phenotype is measured in triplicate with replicate correlation reported, since a noisy MBEC assay will cap any downstream heritability estimate.
Second, partitioning. For each of the three phenotypes I estimate chip heritability with a linear mixed model that models the kinship matrix from genome-wide SNPs and controls clonal structure, the same REML machinery used for M. tuberculosis MICs (GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions). Writing the phenotype as $y_i$ for isolate $i$, the model is $y = g + \varepsilon$ with $g \sim \mathcal{N}(0, \sigma_g^2 K)$ and $\varepsilon \sim \mathcal{N}(0, \sigma_e^2 I)$, and heritability is $h^2 = \sigma_g^2 / (\sigma_g^2 + \sigma_e^2)$. The key derived phenotype is the amplification residual $B_i = \log(\text{MBEC}_i) - \alpha \log(\text{MIC}_i)$, with $\alpha$ fit by regression, which strips out the component of biofilm survival that is just inherited planktonic resistance (the A. baumannii cohort found MBEC and MIC positively correlated) and isolates what the biofilm adds. The headline deliverable is a table: $h^2$ for MIC, biomass, and $B$, and the fraction of $B$'s variance explained by all additive genome-wide significant hits. My prediction, by analogy with the 64–88%-versus-0.01–0.24% gap in tuberculosis, is that $h^2(B)$ will be substantial while additive hits explain a few percent at most — a measured, bacterial instance of missing heritability on a trait the field treats as purely physical.
Third, recovery. Two complementary engines, both proven in other domains and never run on biofilm data. One is lo-siRF, the signed iterative random forest that finds high-order non-linear interactions invisible to additive GWAS and whose cardiac-hypertrophy hits were confirmed by dual knockdown (an 8.5% non-additive cell-size reduction, $p < 10^{-4}$) (Epistasis regulates genetic control of cardiac hypertrophy). The other is SuperDCA-style direct coupling analysis with phylogenetic filtering, which scales to $10^5$ bacterial SNPs (an 18-fold speedup; 94,028 SNPs in under 8 days) and separates genuine co-evolution from clonal linkage disequilibrium (SuperDCA for genome-wide epistasis analysis). To these I add the proposal's specific new ingredient: a biofilm-curated edge set for a graph model, replacing the generic STRING protein-interaction graph used by GeneBac (Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction) with edges drawn from the mechanisms the biofilm literature actually implicates — the $(p)\text{ppGpp}$ stringent response that maintains tolerance (Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response), the DksA1 regulator repeatedly converged on during CF airway adaptation (Evolution and host-specific adaptation of Pseudomonas aeruginosa), master transcriptional switches like csgD that turn matrix production on and off (Biofilm-mediated resistance to berberine in Escherichia coli), small-RNA control of anoxic biofilm formation (Cross-regulation by CrcZ RNA controls anoxic biofilm formation in Pseudomonas aeruginosa), and matrix decoys such as filamentous Pf phage that sequester cationic antibiotics (Pf Bacteriophage and Their Impact on Pseudomonas Virulence, Mammalian Immunity, and Chronic Infections). Where Farhat's study found that over 20% of resistance loci are intergenic promoters, I extend variant calling explicitly to promoter regions of these regulons (GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions).
Validation closes the loop, following two precedents for how statistical bacterial genomics earns biological credibility. Paired knockouts of the top interacting pairs in a lab P. aeruginosa strain test for non-additive MBEC shifts, mirroring the hiPSC-CM dual-silencing validation of lo-siRF (Epistasis regulates genetic control of cardiac hypertrophy). And the phenogenotyping platform of 158 barcoded M. tuberculosis strains across 14 stress conditions shows how intermediate phenotypes can anchor genotype to clinical outcome (a slow-growth subclade raised treatment failure from 6% to 22%) (High-throughput phenogenotyping of Mycobacteria tuberculosis clinical strains reveals bacterial determinants of treatment outcomes); analogously, I test whether predicted high-amplification genotypes are over-represented among chronic-infection isolates in the 9,829-genome collection.
Baselines
The naive reference point is the fixed-multiplier rule now implicit in practice: predict each strain's MBEC as its MIC scaled by a population-constant factor, using the 16- to 2,048-fold range reported for A. baumannii as the calibration interval (Relationship between Antibiotic Resistance, Biofilm Formation, and Biofilm-Specific Resistance in Acinetobacter baumannii). Its close cousin is biomass-as-proxy: predict MBEC from crystal-violet OD550. Both are cheap, need no genomes, and will be run on identical held-out isolates. Beating them means a genomic model with materially lower held-out error on $\log(\text{MBEC})$ and on $B$, and — decisively for the biomass proxy — positive predictive correlation with $B$ where biomass has none, which the A. baumannii data already suggest it cannot have.
The strongest existing competitor is the GeneBac architecture: a CNN embedding each gene's sequence, a graph neural network over STRING protein-interaction edges, and multi-task pre-training on paired DNA/RNA data (nearly 0.9 Pearson for predicting RNA abundance, worth about a 4% average gain) (Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction). It reported Pearson correlations over 0.87 on 12,460 M. tuberculosis strains and 0.76 on 1,109 P. aeruginosa strains for planktonic MIC — the number to beat on MBEC, though on a phenotype it was never trained for. To make the comparison fair I will fine-tune GeneBac itself on our MBEC dataset with the same train/validation/test splits, the same GPU budget, and the same hyperparameter search sweeps granted to my model, so the comparison isolates the input representation and the edge set, not tuning effort. Also run alongside it is the additive-locus baseline in the spirit of Farhat's linear mixed model with gene burden scores (GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions). Beating GeneBac means higher held-out Pearson on $B$ specifically, not just on raw MBEC, since raw MBEC is partly predictable from MIC.
The critical ablation swaps the edge set and nothing else: the same sequence encoder and the same graph architecture, once with STRING edges (the GeneBac prior) and once with the biofilm-curated edge set described above. If the biofilm-curated graph does not improve prediction of $B$ over the STRING graph under matched capacity, the proposal's central ingredient — that biofilm-specific pathway priors are what unlock the missing variance — is wrong, and I would report that plainly. A second ablation drops all pairwise-interaction terms, leaving a single-gene additive model on the same features, to quantify how much of the recovered variance is genuinely non-additive.
Why not just…?
Why not just call it phenotypic tolerance and skip the genetics? The strongest version of this objection says biofilm protection is physical shielding and dormancy — antibiotic-decoying phage in the matrix, curli fibers slowing oxidant diffusion, oxygen-starved cores — with no mutation, hence no heritability to find (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). But the machinery that builds the shield is genetically encoded — the stringent response, matrix regulators, cyclic-di-GMP switches — and the A. baumannii cohort shows stable, reproducible, strain-specific MBEC differences under standardized conditions (Relationship between Antibiotic Resistance, Biofilm Formation, and Biofilm-Specific Resistance in Acinetobacter baumannii). Still, the objection could be right in the limit: if the heritability measurement returns $h^2(B) \approx 0$, the proposal's premise fails and the honest conclusion is that amplification is plastic. That is itself a publishable, decisive answer, and the first experiment is designed to deliver it either way.
Why not just run a standard additive GWAS or Lasso on the MBEC data? Because the precedent says additive tools leave the variance on the table: in tuberculosis, heritability of 64–88% against canonical-locus explanations of 0.01–0.24% (GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions), and in sequence-based prediction, GeneBac's interaction-aware model substantially outperformed Lasso, ElasticNet, and a plain CNN (Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction). That said, the additive baseline is a real threat for the raw MBEC phenotype, since MBEC correlates with MIC and MIC is additively predictable. This is precisely why the analysis pivots on the amplification residual $B$, where the additive route through MIC has been subtracted out; if additive GWAS on $B$ performs as well as the epistatic model, the epistasis story collapses, and the ablation above is built to detect exactly that.
Why not just measure biomass and use it as the surrogate? Crystal-violet staining is far cheaper than peg-lid MBEC assays, and if it tracked survival it would be the sensible high-throughput choice. The A. baumannii data rule this out directly: resistance amplification was independent of biomass quantity, and the biomass-richest strains were the least drug-resistant ($r_s = -0.720$) (Relationship between Antibiotic Resistance, Biofilm Formation, and Biofilm-Specific Resistance in Acinetobacter baumannii). A biomass surrogate would systematically misclassify the most dangerous chronic-infection strains. I will still measure biomass and test the proxy empirically on our cohort rather than assert the negative from one species, but the burden of proof sits with the proxy.
Why not just worry that clonality fakes every interaction you find? This is the sharpest methodological objection: epidemic clones inherit whole haplotypes together, so apparent gene-gene coupling may be lineage structure in disguise. It is addressed head-on rather than argued away. The sampling design over-samples clone-mates so that within-clone contrasts identify effects; the mixed model absorbs the kinship structure as in Farhat's LMM-REML analysis (GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions); SuperDCA's phylogenetic filtering is the demonstrated tool for separating convergent co-evolution from clonal linkage in bacteria (SuperDCA for genome-wide epistasis analysis); and any surviving candidate pair faces the paired-knockout test, where a lineage artifact cannot produce a non-additive MBEC shift in a clean genetic background.
Decisive Checks
Three results would settle the proposal. First, $h^2(B)$ substantially above zero with additive hits explaining a small fraction of it — the existence proof for missing heritability in biofilm resistance; near-zero $h^2(B)$ refutes the premise. Second, the biofilm-curated graph beating both the STRING-graph ablation and fine-tuned GeneBac on held-out $B$, under matched tuning — the evidence that pathway-matched priors, not extra parameters, do the work; on raw MBEC the number to beat is the 0.76 Pearson GeneBac reported for P. aeruginosa (Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction), with the caveat that no reported result exists on Lacuna for any model trained on MBEC, which is the gap this proposal fills. Third, at least one predicted interacting pair showing a non-additive MBEC shift under paired knockout — the causal anchor. If the first two hold and the third fails, the finding stands as a predictive-genomics result with unresolved mechanism; if the first fails, biofilm tolerance joins the short list of bacterial phenotypes that are genuinely non-heritable, and clinical MBEC testing would need to stay phenotypic.
References
- Relationship between Antibiotic Resistance, Biofilm Formation, and Biofilm-Specific Resistance in Acinetobacter baumannii
- GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions
- Evolution and host-specific adaptation ofPseudomonas aeruginosa
- Epistasis regulates genetic control of cardiac hypertrophy
- SuperDCA for genome-wide epistasis analysis
- Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction
- Broad-Spectrum Anti-biofilm Peptide That Targets a Cellular Stress Response
- Biofilm-mediated resistance to berberine in Escherichia coli
- Cross-regulation by CrcZ RNA controls anoxic biofilm formation in Pseudomonas aeruginosa
- Pf Bacteriophage and Their Impact on Pseudomonas Virulence, Mammalian Immunity, and Chronic Infections
- High-throughput phenogenotyping of Mycobacteria tuberculosis clinical strains reveals bacterial determinants of treatment outcomes
- Elucidating the Role of Virulence Traits in the Survival of Pathogenic E. coli PI-7 Following Disinfection