Statistical Genomics Foundry
Design docs prior-art research

Survey — Beyond R

Statistical genomics prior art outside the R / Bioconductor world.

Research lens 2 of 4 for the "statistical rails for LLM-driven statistical genomics" project. Source: research subagent web survey, 2026-06-26. Evidence tags: [verified] = fetched primary source; [secondary] = mentioned but not verified at source; [inference] = subagent reasoning; [speculative].

1. Lens & scope

Validity/QC/calibration machinery and any LLM-agent efforts in non-R statistical genomics — GWAS/PRS CLI tools, Python statistical-genetics stacks, genomics-scale calibration conventions, and reporting/protocol standards — read against our goal of "do vs. referee" statistical rails for an LLM.

2. Landscape inventory

# Name · URL What it does Family Form Maturity Relevance to "LLM rails"
1 PLINK / PLINK2 (https://www.cog-genomics.org/plink/) Core GWAS engine: assoc tests, --maf/--hwe/--mind/--geno, --king-cutoff relatedness, --indep-pairwise LD pruning, PCA A (+B inputs) CLI-tool Mature, dominant The QC/calibration verbs an LLM must call correctly; flags encode validity gates. [secondary]
2 regenie (https://rgcgithub.github.io/regenie/) Whole-genome regression (step-1 ridge / step-2 LMM); Firth + SPA for case-control imbalance; burden/SKAT/ACAT A CLI-tool Mature (RGC) Calibration baked into method (Firth/SPA → λ≈1); a "correct method choice" target. [verified]
3 SAIGE / SAIGE-GENE+ (https://github.com/saigegit/SAIGE) GLMM + SPA for binary traits, controls case-control imbalance & relatedness A CLI-tool Mature Canonical example of a method chosen because naive logistic mis-calibrates rare/imbalanced data. [verified — Zhou 2018 PMC30104761]
4 GCTA (GREML / fastGWA / fastGWA-GLMM) (https://yanglab.westlake.edu.cn/software/gcta/) Heritability (GREML); fastGWA MLM with sparse GRM + PCs for stratification/relatedness A/B CLI-tool Mature Sparse-GRM thresholding is an explicit calibration knob (genomic inflation vs power tradeoff). [verified — Nat Genet 2019]
5 BOLT-LMM (https://alkesgroup.broadinstitute.org/BOLT-LMM/) Bayesian mixed-model assoc at biobank scale A CLI-tool Mature The high-memory reference fastGWA benchmarks against; mixed-model = stratification control. [secondary]
6 Hail (https://hail.is/) Scala/Python distributed genomics; sample_qc(), variant_qc(), GWAS, λGC A/B python-lib (dist.) Mature (Broad; gnomAD, UKB, GTEx) QC as first-class library methods → an LLM can call structured QC rather than free-write it. [verified]
7 LDSC (ldsc) (https://github.com/bulik/ldsc) LD-score regression; intercept distinguishes confounding from polygenicity; h², rg B python-lib/CLI (py2) Mature, canonical The calibration referee: intercept≈1 ⇒ inflation is real signal not confounding; ratio=(int−1)/(mean χ²−1). Exactly our "external check, not model reasoning." [verified — repo + Nat Genet ng.3211]
8 PGS Catalog (https://www.pgscatalog.org/) Open DB of published PRS (variants/alleles/weights) + curated metadata for reproducibility A/B reporting-standard / DB Mature (EBI) Machine-readable scoring files + provenance an agent can reuse instead of re-deriving. [verified — PMC11165303]
9 PRS-RS / ClinGen Complex-Disease reporting standards (https://clinicalgenome.org/docs/clingen-complex-disease-prs-reporting-standards/ ; Wand et al. Nature 2021) 33-item checklist across PRS derivation/testing/validation incl. generalizability B reporting-standard Mature A literal referee checklist — directly castable into audit "skills." [verified — clinicalgenome PDF + Nature s41586-021-03243-6]
10 GWASTutorial (https://cloufield.github.io/GWASTutorial/) End-to-end CLI+Python curriculum: QC, PLINK/regenie/LDSC/GCTA/SuSIE, Manhattan/QQ/λ A+B protocol/tutorial Active (UTokyo) Closest to a structured "do-it-right" knowledge graph in non-R world; sequences method + diagnostic. [verified — fetched site]
11 Marees et al. 2018 GWAS tutorial (https://github.com/MareesAT/GWA_tutorial) 4-part PLINK/PRSice QC→stratification→assoc→PRS scripts A+B protocol/tutorial Widely cited Codifies QC ordering + thresholds; a "correct workflow" exemplar. [verified — repo + PMC6001694]
12 Anderson et al. 2010 Nature Protocols (https://www.nature.com/articles/nprot.2010.116) Case-control data-QC protocol; PLINK + SMARTPCA ancestral-outlier removal B protocol Foundational The canonical QC-checklist literature an audit skill would cite. [verified — PubMed 21085122]
13 gwaslab (https://github.com/Cloufield/gwaslab) Python sumstats QC: harmonization, allele/chr standardization, λGC, Manhattan/QQ/Miami A/B python-lib Active Post-hoc calibration diagnostics as a Python object — agent-friendly referee toolkit. [verified — repo/docs]
14 qmplot (https://github.com/ShujiaHuang/qmplot) Python Manhattan + Q-Q plots from GWAS results B python-lib Modest Narrow: the QQ/calibration visual check in Python. [secondary]
15 scikit-allel (https://scikit-allel.readthedocs.io/) Pop-gen variation analysis (allele freq, diversity, PCA, selection scans) A/B python-lib Maintenance-only (→ sgkit) Diagnostics building blocks; flag the deprecation. [verified — docs note maintenance mode]
16 sgkit (https://github.com/sgkit-dev/sgkit) Xarray/Dask/Zarr statistical-genetics toolkit; scikit-allel successor A/B python-lib Active, maturing Structured stat-genetics methods over typed arrays — the modern Python substrate. [verified — repo + NumFOCUS]
17 GenoTools (https://github.com/dvitale199/GenoTools ; biorxiv 2024.03.26.586362) Python package wrapping genotype QC + ancestry pipeline (GP2) A python-lib/CLI Active Packages the QC checklist as callable Python — close to "QC-as-skill." [secondary — biorxiv]
18 nf-gwas / nf-core-style pipelines (PMC10849172) Nextflow regenie GWAS w/ pre/post-processing, reproducible A pipeline Active Reproducibility rail (not statistical-validity referee) — the "wiring" LLMs already do well. [secondary]
19 statsmodels (https://www.statsmodels.org/) General Python stats: GLM, multiple-testing (FDR/BH), diagnostics A/B python-lib Mature Generic inference engine + multipletests — the error-rate-control primitive under genomics code. [inference]
20 scanpy / scverse (https://scanpy.readthedocs.io/) Single-cell Python: clustering, DE testing A python-lib Mature Where "double-dipping" (cluster then test on same data) invalidity classically appears — a referee target. [inference]

Agent/LLM systems (detailed in §5)

# Name · URL What Family Maturity
21 GenoTEX / GenoAgent (https://arxiv.org/abs/2406.15341) Benchmark + multi-agent system for gene-expression → gene-trait analysis A (weak B) Research benchmark
22 BixBench (https://www.researchgate.net/publication/389548449) Benchmark: LLM agents interpret bioinformatics analytical results in research context A/B Research benchmark
23 BioAgent Bench / BioMaster / AutoBA (https://arxiv.org/pdf/2601.21800 ; https://arxiv.org/pdf/2512.09964) Agentic bioinformatics workflow execution / NGS downstream for non-experts A Research/early
24 TusoAI (https://arxiv.org/pdf/2509.23986) Agentic optimization of scientific methods neither/A Research

3. Closest prior art (top 3)

  1. GWASTutorial (#10) — the strongest non-R analog to what we're building: a structured, tool-spanning body of "do-it-right" knowledge that pairs each method with its diagnostic (QC → assoc → λ/QQ → LDSC intercept → fine-mapping). It's a knowledge base, not an agent or a gate, but it's the corpus a Family-A/B skill set would be distilled from. [verified]
  2. LDSC intercept (#7) — the cleanest existing instance of our structural bet: an external, empirical check (regress χ² on LD score) that adjudicates whether an inflated result is real or confounded, independent of the analyst's narrative. This is exactly "doing hands off to a referee using empirical checks, not the model's own reasoning." [verified]
  3. PRS-RS / ClinGen (#9) + PGS Catalog (#8) — a 33-item referee checklist plus a machine-readable provenance DB. Directly castable into Family-B audit skills and Family-A reuse (pull an established score instead of inventing one). [verified]

4. White space / gaps

  • No referee gate exists as software. All "validity" lives in human-readable checklists (PRS-RS, Anderson, Marees) or scattered diagnostics (λGC, QQ, LDSC intercept). Nothing forces a doing pipeline to hand off to an external check before certifying — the core thing we'd add. [inference]
  • "Invented method with a cool name" is unguarded. Existing tooling assumes you picked an established method; nothing detects an agent confabulating a novel, statistically-invalid procedure, or maps "proposed method → invalidity pattern (uncontrolled error rate / circular inference / unmet assumption)." [inference]
  • Calibration is per-tool, not portable. λGC, intercept, permutation, negative controls each live in a different tool/lang; no inspectable cross-tool "calibration skill" library. [inference]
  • Negative-control / permutation / simulation-under-known-truth scaffolding is DIY. Field trusts these but provides no reusable harness — exactly the Family-B "construct the empirical checks" deliverable. [inference]
  • Double-dipping in single-cell (scanpy) is widely committed and unguarded by the tooling itself. [speculative]
  • Agent benchmarks measure productivity/interpretation, not statistical honesty. None gate on calibration/error-rate control. [verified for fetched sources]

5. Agentic / LLM-specific efforts (called out)

  • GenoTEX / GenoAgent — multi-agent gene-expression analysis (dataset selection → preprocess → stat analysis) with "flexible self-correction." Validation = expert bioinformatician annotations baked into the benchmark, plus self-correction. No separate statistical-validity/referee layer distinct from the analysis agent — self-certification, the failure mode we're targeting. [verified — fetched abstract]
  • BixBench — evaluates whether agents can interpret analytical results in the context of a research question. Closest to a "referee" framing but it's a benchmark of judgment, not an enforced empirical gate. [secondary]
  • BioMaster / AutoBA / BioAgent Bench / NGS-downstream agent — workflow execution automation (tool selection → code → run), RAG over domain docs. Squarely the "wiring/reproducing" strength; statistical validity out of scope. [secondary]
  • TusoAI — agentic method optimization; adjacent (could even generate the invented-method risk). [secondary]
  • Honest read: the agent space is thin on exactly our axis. Everything either measures productivity or interpretation quality; none implements "doing never self-certifies; it hands to an empirical referee gate." That's our white space. [inference]

6. Follow-up URLs

Evidence caveats: exact PLINK QC thresholds (#1) and Hail/GenoTools/BixBench/BioMaster details are [secondary] (search snippets, source not fully fetched). regenie, LDSC, GWASTutorial, gwaslab, sgkit, GenoTEX abstract, PRS-RS/PGS are [verified] (docs/repo/paper fetched or read). Family-A/B mappings for statsmodels, scanpy, and the gap analysis are [inference]. The single-cell double-dipping prevalence claim is [speculative].