Microbes carry contiguous blocks of genes — biosynthetic gene clusters (BGCs) — that together build small molecules such as antibiotics. Most of these clusters are transcriptionally silent in the lab: the genes are present and intact, but never switched on. The scale of the waste is quantified: across 33,904 marine prokaryote genomes, 96.8% of biosynthetic space is unexpressed or unannotated (Global analysis of the biosynthetic chemical space of marine prokaryotes). Meanwhile, human and yeast genomics have a mature machinery for exactly this kind of question: expression quantitative trait locus (eQTL) analysis, which treats how strongly each gene is transcribed as a heritable trait and finds the DNA spelling variants that explain it, split into cis variants near the gene and distant trans regulators (The Genotype-Tissue Expression (GTEx) project).
The sharpest motivation for joining these fields is a single number pair from Salinispora: the ketomemicin cluster is present and syntenous in 79 of 118 sequenced genomes, yet only 25 of 118 strains actually produce the metabolite under identical culture (Structure Elucidation, Biosynthetic Gene Cluster Distribution, and Biological Activities of Ketomemicin Analogs in Salinispora). Fifty-four strains carry a complete, conserved pathway that stays asleep, and nobody knows which sequence differences — a broken promoter motif, a hyperactive repressor, a missing activator — are responsible. My proposal is to stop guessing with elicitor cocktails and instead run a population genetic association study on cluster dormancy itself: RNA-seq a natural isolate panel under standard and elicitor conditions, map the cis and trans variants that predict whether each cluster is on or off, and then wake silent clusters by editing the predicted regulators.
Core Claim
BGC dormancy across a microbial genus is not stochastic; it is a quantitative genetic trait with discoverable loci. If we profile transcriptomes and metabolomes across ~100 natural isolates that share a conserved cluster, a structured eQTL model will identify a small set of regulatory variants that separate producer from silent strains, and targeted edits to those variants will activate dormant clusters more reliably and more mechanistically informatively than unguided elicitor screening. Three design choices distinguish the model from off-the-shelf eQTL: the association graph comes from the clusters themselves (genes in one BGC must be co-expressed to make a product, so linkage between cluster-mates is a natural fusion prior); the estimator is robust to the heavy-tailed, outlier-prone expression and yield data typical of fermentations; and every strain is profiled under both standard and perturbation conditions so that genotype-by-environment effects — regulators that matter only under an elicitor — are mapped explicitly.
The closest prior work makes the gap concrete rather than hypothetical. Curtis et al. ran graph-guided structured eQTL mapping in 112 yeast segregants and found 22 regulatory hotspots, 10 of them novel, validating that knockouts of predicted regulators shifted 39 of 43 target genes in a ribosome-biogenesis module (Structured association analysis leads to insight into Saccharomyces cerevisiaegene regulation by finding multiple contributing eQTL hotspots associated with functional gene modules). That paper is the methodological seed for this proposal, but it analyzed primary eukaryotic metabolism and telomeric silencing, never secondary metabolite clusters, and left epistasis and nonlinear structure unmodeled. On the microbiology side, the Salinispora study measured the dormancy phenotype precisely and stopped there — no transcriptomics, no association mapping. Nobody has connected the two, which is the opportunity.
Research Program
The study has four components, built around Salinispora as the anchor system because it is the only taxon in the literature base with both a genus-wide genome set (118 strains) and matched public LC-MS/MS metabolomics defining who produces and who does not (Structure Elucidation... Salinispora).
First, cohort and variant preparation. Re-annotate BGCs across the 118 genomes using BGC-Master, a DNA foundation-model segmenter that works directly on raw sequence and reaches an F1-score of 0.521 on the 9-genome, 305-cluster gold-standard benchmark, versus 0.256 for antiSMASH 7 (BGC-Master: Detecting Novel Biosynthetic Gene Clusters with DNA Foundation Models); antiSMASH is run in parallel for comparability with prior annotations. Call the pan-genome: core-genome SNPs plus presence/absence variants of accessory regulatory genes. This matters because the plausible trans actors in bacteria — cluster-situated regulators, global pleiotropic repressors, two-component systems — are often accessory.
Second, the expression-by-condition matrix. Cultivate all viable strains under a standard condition and a small panel of chemical perturbations chosen from the literature: cobalt chloride and EDTA, which awakened silent pathways in Aquimarina muelleri (Global analysis...); halide supplementation, which triggered seven new metabolites in Talaromyces flavus (Bioactive polyketides and meroterpenoids from the mangrove-derived fungus Talaromyces flavus TGGP35); and humus acid, which induced the dormant anthrabenzoxocinone cluster in Streptomyces sp. FJS31-2 (Molecular Genetic Characterization of an Anthrabenzoxocinones Gene Cluster in Streptomyces Sp. FJS31-2 for the Biosynthesis of BE-24566B and Zunyimycin Ale). RNA-seq plus LC-MS/MS per strain per condition gives both the expression trait and the product phenotype. Curtis et al. showed that yeast regulators like PBF2 manifest only under heat shock (Structured association analysis...); the elicitor arms test exactly this condition-specific class, which single-condition eQTL misses by construction.
Third, the association model. For each BGC gene $g$ in strain $i$ and condition $c$, fit
$$\min_{\beta}\; \sum_{i,c} \rho\!\Big(y^{(g)}_{i,c} - x_i^{\top}\beta - \gamma\, x_i c\Big) \;+\; \lambda_1 \|\beta\|_1 \;+\; \lambda_2 \!\!\sum_{(g,h)\in E_{\text{BGC}}}\!\! w_{gh}\,\|\beta^{(g)} - \beta^{(h)}\|_2,$$
where $y$ is normalized log-expression, $x_i$ is the strain's genotype vector, $\rho$ is a bounded robust loss (the MM-estimator's bi-square function), $E_{\text{BGC}}$ is the graph connecting genes within the same cluster, and $\gamma$ captures the genotype-by-elicitor interaction. The three ingredients each come from a cited result: the graph-fusion penalty from GFlasso, which surfaced multi-locus hotspots that univariate scans missed in yeast (Structured association analysis...); the robust loss from the demonstration that ordinary least squares lost 112 real eQTLs to outliers in brain tissue, with OLS–robust concordance of only 50% in that dataset (Robust Linear Models for Cis-eQTL Analysis); and the interaction term from the condition-dependence observed for yeast regulators. Outlier-robustness is not decorative here: fermentation yields are extreme by nature — wild-type Diaporthe makes phomoxanthone A at 3 g/L while its minor terpenes are near-undetectable (Deletion of a polyketide synthase phoE induces monoterpenes production in ascidian-derived fungus Diaporthe sp. SYSU-MS4722) — precisely the heavy-tailed regime where OLS fails. Preprocessing will be treated as part of the inference problem, since normalization choice alone shifted heritability calls from 1,962 genes (RMA) to 44 (PLIER) in one human dataset, with cross-method overlap of only 10–43% (Impact of normalization and filtering on linkage analysis of gene expression data); the plan is to report the full sensitivity analysis across normalizations rather than pick one silently. Per-strain bioinformatic processing at this scale can be automated with a multi-agent workflow engine such as Liu-Hy/GenoMAS.
Fourth, validation by activation. The output that matters is not a p-value but a metabolite. Edit the top-predicted trans repressor (knockout or promoter attenuation) and the top-predicted cis variant (promoter swap) in a set of silent, cluster-harboring strains, and assay for the target metabolite by LC-MS/MS against the published ketomemicin spectra and analogs (Structure Elucidation... Salinispora).
A word on power, because a reviewer will raise it. GTEx power analysis gives 80% detection of cis-eQTLs at 2% minor allele frequency with roughly 750 donors (The Genotype-Tissue Expression (GTEx) project) — far beyond our cohort. But Curtis et al. recovered 22 hotspots from only 112 segregants (Structured association analysis...), and bacteria are more favorable than humans in two ways: linkage blocks are huge in clonal lineages, and the trans candidates are a bounded, enumerable set of regulatory genes. The design therefore restricts to common variants (allele count ≥ 10), pools signal across all BGC families a strain carries rather than testing one cluster at a time, and treats the study as discovery-then-validation rather than exhaustive genome-wide enumeration.
Baselines
The do-nothing reference point is standard-broth culture of the panel. On the existing Salinispora data this yields ketomemicin production in 25 of 118 strains, with 54 of the 79 cluster-harboring strains silent (Structure Elucidation... Salinispora); the do-nothing activation rate among silent strains is therefore zero, and any intervention must beat that on the same strains by LC-MS/MS. There is no reported result found on Lacuna for an elicitor-panel activation rate across a multi-strain Salinispora cohort, so this reference point is being measured as part of the study.
The strongest competitor for the biological goal is unguided elicitor screening (the OSMAC-style approach): apply the same elicitor panel — CoCl₂, EDTA, halides, humus acid — to every silent strain and count activations. This is the strongest because it is the field's default, it directly produces metabolites, and it requires no genomics. It will be run on exactly the same strains, same elicitors, same fermentation conditions, same LC-MS/MS instrumentation and same detection thresholds as the eQTL-guided edits, so the only difference is whether the intervention was chosen by genotype or by brute force. Beating it means either a higher activation rate per strain, or the same rate achieved with a mechanistic explanation and a portable predictor of which strains respond to which elicitor; the honest headline metric is silent strains converted to producers per strain assayed, with a secondary metric of cross-validated prediction accuracy (AUC) of producer/silent status from genotype alone, which elicitor screening cannot provide.
The strongest competitor for the statistical goal is the published GFlasso pipeline of Curtis et al., which found 22 hotspots (10 novel) in 112 yeast segregants and whose regulator knockouts shifted 39 of 43 target genes in a module (Structured association analysis...). It will be run as published on our data — same expression matrix, same variant set, same tuning via cross-validation — plus two nested variants: univariate OLS eQTL, and univariate robust MM-estimator eQTL in the style of Rantalainen et al., who recovered 112 eQTLs that OLS missed with 50% OLS–robust concordance in brain tissue (Robust Linear Models for Cis-eQTL Analysis). All models get identical inputs and compute; the comparison metric is held-out prediction of expression and of producer status, plus the downstream knockout-validation rate of predicted regulators.
Finally, the ablation that isolates the key ingredient: the same robust, interaction-aware model run with a degree-matched random graph in place of the BGC-structure graph, and with the OLS loss in place of the robust loss. If the operon-structure prior adds nothing over the random graph, the claimed ingredient is dead and the proposal collapses to robust univariate mapping; the ablation is designed so that this outcome is visible, not arguable.
Why not just…?
Why not just clone the cluster into a domesticated host? Heterologous expression under constitutive promoters is an established pillar of the design-build-test-learn cycle for natural products (Synthetic Biology Advanced Natural Product Discovery), and for the sole goal of obtaining a compound it may well win on speed for small clusters. But it sidesteps rather than answers the question of why the cluster is silent, and it scales badly: clusters like the 82.8 kb, 29-gene trans-acyltransferase PKS/NRPS hybrid found in Bacillus paralicheniformis Bac48 (In silico exploration of Red Sea Bacillus genomes for natural product biosynthetic gene clusters) demand large DNA assembly, tolerate poorly in E. coli or yeast, and often fail for lack of unusual precursor building blocks — no reported result found on Lacuna for a successful heterologous expression of that cluster class. This proposal produces knowledge that transfers (a regulator discovered in Salinispora points to its orthologs in the 28 non-Salinispora actinomycetes already reported to carry ktm homologs), and its validated activation edits are in the native host, sidestepping precursor incompatibility entirely.
Why not just screen lots of elicitors? This is the most serious alternative, and I will not argue around it: elicitation works — triclosan awakened huanglongmycins in a cave Streptomyces (Huanglongmycin A-C, Cytotoxic Polyketides Biosynthesized by a Putative Type II Polyketide Synthase From Streptomyces sp. CB09001), NaBr yielded seven new metabolites in Talaromyces (Bioactive polyketides...), and EDTA/CoCl₂ did so in Aquimarina (Global analysis...). The experiment settles the question rather than rhetoric: the elicitor panel is run on the same strains as the genotype-guided edits, under identical conditions. My claim is not that elicitation fails, but that it is genotype-blind — 54 of 79 ktm-harboring strains are silent under standard broth, and screening them all with every elicitor is combinatorially wasteful and gives no mechanism. If unguided elicitation matches the guided activation rate on this panel, that result will be reported prominently and the contribution narrows to the predictor; if the guided edits activate strains that no elicitor touched, the genetic-switches claim is directly supported. Either outcome is informative, which is what makes the comparison worth running.
Why not just delete the dominant metabolic sink? Deleting a high-flux core synthase redirected precursors to silent terpene clusters in Diaporthe, yielding eight compounds (Deletion of phoE induces monoterpenes...). But that worked because the wild type made phomoxanthone A at 3 g/L — a loud, dominant, already-characterized sink. The silent Salinispora strains have no such identifiable sink; their dormancy is regulatory, not flux-limited, and deleting arbitrary synthases gives no candidate and no mechanism. Sink deletion is also one specific intervention this framework could later rank against the others, not a replacement for it.
Why not just run standard OLS eQTL? Three cited results say the standard pipeline underperforms in exactly this data regime: OLS missed 112 real eQTLs to outliers in brain tissue with 50% concordance (Robust Linear Models for Cis-eQTL Analysis); univariate scans missed the coordinated multi-locus hotspots that graph-guided regression recovered in yeast (Structured association analysis...); and normalization choice alone scrambled which genes looked heritable, with 10–43% cross-method overlap (Impact of normalization and filtering...). Still, OLS is retained as a baseline with equal data and tuning budget — if it finds the same regulators, the robust and structured machinery was unnecessary and the simpler model should be preferred. That is what the ablation section is for.
Proposed Experiments and Decisive Checks
The program runs as four experiments. Experiment one builds the resource: 118 Salinispora genomes re-annotated with BGC-Master (F1 0.521 on the 9-genome benchmark, versus 0.256 for antiSMASH 7 and 0.117 for DeepBGC, per BGC-Master), a called pan-genome variant set, and an established producer/silent ground truth from the published LC-MS/MS comparison (Structure Elucidation... Salinispora). Experiment two generates the data matrix: RNA-seq and LC-MS/MS for every cultivable strain under the standard condition plus the four-elicitor panel, with all normalizations reported per the sensitivity problem documented by Impact of normalization and filtering.... Experiment three fits the model and the baselines — GFlasso as published, univariate OLS, univariate robust MM, and the random-graph and OLS-loss ablations — all on identical inputs, scored by cross-validated AUC for predicting producer/silent status from genotype and by the number of mapped hotspots. Experiment four is decisive: CRISPR-edit the top-predicted trans regulator and the top-predicted cis promoter variant in ten silent, cluster-harboring strains, and assay for ketomemicin production.
The pre-registered success criteria are concrete. The method succeeds if genotype alone predicts producer/silent status with cross-validated AUC above 0.8 (chance is 0.5); if regulator edits convert at least three of ten silent strains to detectable producers — recall the do-nothing rate is zero and the unguided-elicitor rate is being measured on the same strains; and if the knockout-validation rate of predicted regulators on cluster-gene expression approaches the 39-of-43 standard set in yeast (Structured association analysis...). The method fails informatively if the random-graph ablation matches the BGC-graph model (the structural prior is vacuous), if OLS matches the robust estimator (the tail-robustness is vacuous), or if unguided elicitation converts silent strains at the same or better rate than the guided edits — in which case the honest conclusion is that dormancy in this genus is environmental rather than genetic, itself a finding that would redirect the whole elicitation literature from single-strain anecdotes toward strain-panel design.
References
- Global analysis of the biosynthetic chemical space of marine prokaryotes
- The Genotype-Tissue Expression (GTEx) project
- Structure Elucidation, Biosynthetic Gene Cluster Distribution, and Biological Activities of Ketomemicin Analogs in Salinispora
- Structured association analysis leads to insight into Saccharomyces cerevisiaegene regulation by finding multiple contributing eQTL hotspots associated with functional gene modules
- BGC-Master: Detecting Novel Biosynthetic Gene Clusters with DNA Foundation Models
- Bioactive polyketides and meroterpenoids from the mangrove-derived fungus Talaromyces flavus TGGP35
- Molecular Genetic Characterization of an Anthrabenzoxocinones Gene Cluster in Streptomyces Sp. FJS31-2 for the Biosynthesis of BE-24566B and Zunyimycin Ale
- Robust Linear Models for Cis-eQTL Analysis
- Deletion of a polyketide synthase phoE induces monoterpenes production in ascidian-derived fungus Diaporthe sp. SYSU-MS4722
- Impact of normalization and filtering on linkage analysis of gene expression data
- Synthetic Biology Advanced Natural Product Discovery
- In silico exploration of Red Sea Bacillus genomes for natural product biosynthetic gene clusters
- Huanglongmycin A-C, Cytotoxic Polyketides Biosynthesized by a Putative Type II Polyketide Synthase From Streptomyces sp. CB09001