When human geneticists add up the effect of every gene variant that statistically associates with a trait, they recover only a sliver of the variation that family studies say genetics explains. This gap is the "missing heritability" problem, and a decade of work has traced it to tiny effects below significance thresholds, to gene–gene interactions that additive models discard, and to structural variants that standard reference genomes simply do not contain. Bacteria have the same gap, in an extreme form, and an extra cause that no human framework can even express: a large share of bacterial genetic variation never descends through the family tree at all. Horizontal gene transfer — bacteria passing ready-made ring-shaped DNA packages (plasmids) and integrative cassettes directly to one another, across species boundaries — moves whole resistance operons sideways in a single event. Yet every statistical tool built to partition phenotypic variance assumes a fixed, linear, vertically inherited genome.
The result is a blind spot with clinical consequences. A carbapenem-resistant Klebsiella pneumoniae study using additive regularized regression explained only a fraction of the difference between harmless colonization and lethal infection — genomic features gave an area under the receiver-operator curve interquartile range of just 0.55–0.61 — even though the authors could see that the phenotype tracked a specific horizontally acquired element, ICEKp10, interacting with a disrupted chromosomal capsule gene Machine learning models to identify patient and microbial genetic factors associated with carbapenem-resistant carbapenem-resistant Klebsiella pneumoniae infection. The signal was visible; the statistics could not hold it. No study we could find asks the basic quantitative question: of the total variation in antibiotic resistance in a bacterial population, how much lives in the vertically inherited core chromosome, how much in the horizontally transferred mobile compartment, and how much in the interaction between the two? No reported result found on Lacuna for a formal heritability estimate of bacterial minimum inhibitory concentration (MIC) phenotypes.
Core Claim
I propose to build and validate a variance-partitioning framework for bacterial populations that treats the genome as two strata with different inheritance modes. For each isolate we construct two relatedness matrices: $K_C$, the standard core-genome kinship built from vertically inherited single-nucleotide variants, and $K_M$, a mobile kinship built from shared accessory content — plasmid clusters, integrative and conjugative elements, and transposon cargo — recovered from assembly data. A continuous MIC phenotype $y$ is then modeled as
$$y = \mu + g_C + g_M + g_{C \times M} + \varepsilon, \quad \mathrm{Var}(g_C) = \sigma_C^2 K_C,\ \mathrm{Var}(g_M) = \sigma_M^2 K_M,\ \mathrm{Var}(g_{C \times M}) = \sigma_I^2 (K_C \circ K_M),$$
where $\circ$ is the elementwise matrix product and $\varepsilon$ is residual noise, fit by restricted maximum likelihood or Haseman–Elston regression. Each variance component divided by the total gives a heritability attributable to that stratum: $h_C^2$ for the clonal core, $h_M^2$ for the mobile compartment, and $h_I^2$ for core-by-mobile epistasis. The central empirical claim is that in horizontally transferring pathogens, a substantial share of MIC variance — the share that core-only bacterial GWAS and curated resistance-gene catalogs systematically miss — sits in $\sigma_M^2$ and $\sigma_I^2$.
The mobile kinship matrix is the key new object, and its justification comes from the transfer biology itself. Because resistance genes travel together on shared physical backbones, co-localized genes are correlated in a way that mirrors linkage disequilibrium in sexual populations: Multiple-Replicon Resistance Plasmids of Klebsiella Mediate Extensive Dissemination of Antimicrobial Genes shows that 75% of fused multi-replicon plasmids carry more than five resistance genes, with chimera backbones such as the 262.5 kb pM1026-3Ar.1 assembling three ancestral plasmids into one transferable unit; and clinical work shows colistin and carbapenem resistance determinants co-transferring on identical plasmids, so that drug exposure co-selects for resistance to drugs never used Molecular antibiotic resistance mechanisms and co-transmission of the mcr-9 and metallo-β-lactamase genes in carbapenem-resistant Enterobacter cloacae complex, Molecular characterization of NDM-1-producing carbapenem-resistant E. cloacae complex from a tertiary hospital in Chongqing, China. $K_M$ encodes exactly this block-correlated, lineage-crossing inheritance — the same insight that lets interaction-LD score regression recover non-additive variance from additive summary statistics in humans Discovering non-additive heritability using additive GWAS summary statistics, transplanted to a system where the "linkage" is created by conjugation rather than recombination.
Research Program
The program has three build phases and a validation phase. First, data construction: for each target species, recover plasmid and mobile-element content from short-read assemblies using Platon: identification and characterization of bacterial plasmid contigs in short-read draft assemblies exploiting protein-sequence-based replicon distribution scores and, where long reads exist, Plassembler: an automated bacterial plasmid assembly tool. Mobile contigs are clustered into plasmid operational taxonomic units against the Global Soil Plasmidome Resource, a catalog of 98,728 plasmid sequences organized into 57,265 clusters A global soil plasmidome resource unveils functional and ecological roles of plasmids in soil microbiomes. This matters because, as that resource showed, 90% of plasmid clusters exist only in uncultivated metagenomic data — a reference "dark matter" reservoir that closed clinical catalogs miss, the bacterial analog of the non-reference sequences that pangenome work shows carry real phenotypic effects in humans Human pangenome analysis of sequences missing from the reference genome reveals their widespread evolutionary, phenotypic, and functional roles, Using the linear references from the pangenome to discover missing autism variants. Second, matrix construction: $K_C$ from core SNPs; $K_M$ from weighted sharing of pOTUs and mobile gene cassettes across isolates, with variants that preserve physical structure (replicon identity, co-localization, estimated copy number) rather than flattening genes into presence/absence tokens. Third, estimation and calibration: fit the variance model above, with the interaction term $K_C \circ K_M$ capturing the genotype-by-acquisition epistasis that the Klebsiella ICEKp10/kfoC trajectory exemplifies, and run simulation-based calibration of Type I error under a null of no mobile effect, following the validation discipline that made interaction-LD score regression credible Discovering non-additive heritability using additive GWAS summary statistics.
Confidence intervals on every component are non-negotiable: Quantifying the uncertainty in heritability shows heritability point estimates are unstable under stratification and sample-size effects, and clonal bacterial populations are a worst case for that instability, so we report interval estimates under lineage-block bootstrap.
Proposed Experiments
The primary experiment runs on NCBI Pathogen Detection isolates of Escherichia coli and Klebsiella pneumoniae, the two most plasmid-burdened species with large sample counts, using susceptibility phenotypes and the pangenome-plus-plasmid matrices described above. The headline output is the decomposition of resistance variance into $h_C^2$, $h_M^2$, and $h_I^2$ per drug, with a permutation test that shuffles mobile content within clonal lineages to prove the mobile component is not a proxy for population structure. A secondary experiment uses the Klebsiella wildlife-and-clinical cohort — 56 isolates with hybrid long-read assemblies of 69 complete plasmids and quantitative broth-microdilution MICs Multiple-Replicon Resistance Plasmids of Klebsiella Mediate Extensive Dissemination of Antimicrobial Genes — as a small but structurally perfect testbed: it is the only dataset in hand where full chimera-resolved plasmid architectures and continuous MICs coexist, so it lets us test whether preserving replicon structure changes the decomposition relative to flattened gene features. The negative control is Mycobacterium tuberculosis from the CRyPTIC dataset (~12,460 strains, MICs across 13 drugs), a clonal pathogen with essentially no plasmid-mediated transfer; the framework predicts $h_M^2 \approx 0$ there, and any nonzero estimate flags confounding. Finally, for high-order interaction discovery beyond the pairwise kinship term, we adapt low-signal signed iterative random forests — the method that found epistatic cardiac-hypertrophy loci invisible to single-variant GWAS and validated them experimentally Epistasis regulates genetic control of cardiac hypertrophy — to search for interactions between chromosomal background and co-localized mobile operons, with candidates nominated for physical validation by plasmid-curing and conjugation assays using established CRISPR/Cas9 elimination protocols Application of CRISPR/Cas9 System for Plasmid Elimination and Bacterial Killing of Bacillus cereus Group Strains, Exploring the pathogenic function of an endogenous plasmid of Pantoea ananatis by a simple and efficient plasmid elimination strategy.
Baselines
The naive reference point is rule-based lookup of curated resistance genes: if the isolate carries a known determinant, call it resistant. This is what clinical pipelines do today, and it is backed by a real result — GenoARM showed that on E. coli, a transformer-driven selection of just five marker genes plus metadata achieved 93.77% accuracy, and fifteen optimized genes performed nearly as well as the full catalog of roughly 200 Optimizing Gene-Based Testing for Antibiotic Resistance Prediction. We run this catalog baseline on identical splits. It is the strongest competitor for the binary susceptibility question, and our proposal is not primarily about that question; beating it means showing that on continuous MIC values and on variance decomposition, the catalog explains significantly less phenotypic variance than the full model, with the shortfall quantified per drug.
The strongest existing predictive method is GeneBac, a deep architecture that embeds gene sequences with a convolutional network and models gene–gene interactions on a protein-interaction graph, achieving Pearson correlations above 0.87 for M. tuberculosis across 13 drugs on CRyPTIC and 0.76 for P. aeruginosa across 14 antibiotics Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction. No reported result found on Lacuna for GeneBac on Klebsiella pneumoniae or E. coli, so we re-run it ourselves on our datasets with the same feature budget, tuning budget, and held-out-lineage splits as our model, so the comparison measures representation, not compute. GeneBac cannot ingest uncurated plasmid content or report variance components — its authors note the need to pre-select genes as an open limitation and it was validated on organisms without active plasmid transfer — so beating it means higher out-of-sample Pearson r on plasmid-heavy species and, separately, a variance decomposition it cannot produce. The key ablation is our own model with $K_M$ replaced by a flat presence/absence feature matrix of the same genes (the GenoARM-style representation), run with identical everything else: this isolates whether mobile structure — replicon identity, co-localization, chimera boundaries — carries variance that gene content alone does not. A second ablation drops the interaction term $K_C \circ K_M$ to isolate the epistatic component.
Why not just…?
Why not just look up known resistance genes? For binary susceptible/resistant calls in well-characterized species, this genuinely works much of the time — GenoARM's 93.77% with five genes is hard to argue with Optimizing Gene-Based Testing for Antibiotic Resistance Prediction. But it answers a different question. Catalogs say whether a known gene is present; they do not say how much of the population's quantitative MIC variance it explains, and they are blind to uncataloged mobile cargo — the dark-matter reservoir that the soil plasmidome resource shows is 90% metagenomic-only A global soil plasmidome resource unveils functional and ecological roles of plasmids in soil microbiomes. The proposal's catalog-baseline experiment will quantify exactly how much variance the lookup leaves on the table, per drug, so if the answer is "almost none," the proposal's premise is genuinely falsified and we will say so.
Why not just add plasmid genes as extra columns in a standard bacterial GWAS? This is the most serious alternative, and it is precisely our flat-features ablation. The reason it likely falls short is dimensionality and correlation structure: with thousands of accessory genes and hundreds of isolates, per-gene effect estimates are unstable, and the physically induced correlation among co-transferred genes — 75% of multi-replicon plasmids carrying more than five resistance genes Multiple-Replicon Resistance Plasmids of Klebsiella Mediate Extensive Dissemination of Antimicrobial Genes — makes independent-column models misattribute variance. A kinship-based variance component absorbs the whole correlated block and is estimable when individual effects are not, which is the same logic that made genome-wide relatedness estimators succeed where candidate-gene regression failed in humans. The ablation on the 56-isolate hybrid-assembly cohort, where replicon structure is known exactly, is the decisive test: if the flat representation recovers the same decomposition, the structural machinery is unnecessary and the simpler model should win.
Why not just correct for clonal population structure with a standard mixed model? We do exactly that — $K_C$ is the standard clonal correction, and the framework's claim is about what remains after it. The evidence that something remains is direct: in the CRKP cohort, all isolates sat in the same epidemic ST258 clone, yet clinical outcome varied, driven by horizontal ICEKp10 acquisition on top of a chromosomal kfoC disruption, robust to facility-aware splitting Machine learning models to identify patient and microbial genetic factors associated with carbapenem-resistant carbapenem-resistant Klebsiella pneumoniae infection. The within-lineage permutation test and the M. tuberculosis negative control are designed precisely to catch the case where the mobile component is a disguised lineage artifact; if the permuted null cannot be rejected, the core claim fails and the honest conclusion is that clonal structure explains the apparent mobile heritability.
Decisive Checks
Three results would make or break the proposal. First, on E. coli and K. pneumoniae, the mobile variance component must be positive with a confidence interval excluding zero after within-lineage permutation — the bacterial missing-heritability claim made quantitative. Second, on CRyPTIC M. tuberculosis, the same estimator must return $h_M^2$ statistically indistinguishable from zero, proving calibration in an organism where the truth is known. Third, the structural ablation must show that replicon-aware $K_M$ explains more MIC variance than flat gene presence/absence at matched feature count — if it does not, the contribution collapses to "include plasmid genes in bacterial GWAS," which is useful but not this proposal. A confirmatory bonus, in the spirit of the cardiac-hypertrophy study's knockdown validation Epistasis regulates genetic control of cardiac hypertrophy: cure a nominated plasmid from a high-MIC isolate with CRISPR/Cas9 Application of CRISPR/Cas9 System for Plasmid Elimination and Bacterial Killing of Bacillus cereus group Strains, re-measure MIC, and show the drop matches the component's predicted contribution. Beyond antibiotics, the same decomposition applies to any mobile-element-borne phenotype — biofilm strength, shown to be directly promoted by conjugative plasmids Study on the Promotion of Bacterial Biofilm Formation by a Salmonella Conjugative Plasmid and the Underlying Mechanism — so the framework, if it survives these checks, gives statistical genetics a vocabulary for the majority of genetic change in microbes that no current heritability method can see.
References
- Machine learning models to identify patient and microbial genetic factors associated with carbapenem-resistant Klebsiella pneumoniae infection
- Multiple-Replicon Resistance Plasmids of Klebsiella Mediate Extensive Dissemination of Antimicrobial Genes
- Molecular antibiotic resistance mechanisms and co-transmission of the mcr-9 and metallo-β-lactamase genes in carbapenem-resistant Enterobacter cloacae complex
- Molecular characterization of NDM-1-producing carbapenem-resistant E. cloacae complex from a tertiary hospital in Chongqing, China
- Discovering non-additive heritability using additive GWAS summary statistics
- Platon: identification and characterization of bacterial plasmid contigs in short-read draft assemblies exploiting protein sequence-based replicon distribution scores
- Plassembler: an automated bacterial plasmid assembly tool
- A global soil plasmidome resource unveils functional and ecological roles of plasmids in soil microbiomes
- Human pangenome analysis of sequences missing from the reference genome reveals their widespread evolutionary, phenotypic, and functional roles
- Using the linear references from the pangenome to discover missing autism variants
- Quantifying the uncertainty in heritability
- Epistasis regulates genetic control of cardiac hypertrophy
- Application of CRISPR/Cas9 System for Plasmid Elimination and Bacterial Killing of Bacillus cereus Group Strains
- Exploring the pathogenic function of an endogenous plasmid of Pantoea ananatis by a simple and efficient plasmid elimination strategy
- Optimizing Gene-Based Testing for Antibiotic Resistance Prediction
- Sequence-based modelling of bacterial genomes enables accurate antibiotic resistance prediction
- Study on the Promotion of Bacterial Biofilm Formation by a Salmonella Conjugative Plasmid and the Underlying Mechanism