Statistical Genomics Foundry
Design docs prior-art research

Survey — Bioconductor / R

State of the art through the Bioconductor / R lens.

Research lens 1 of 4. Source: research subagent web survey, 2026-06-26. Evidence tags: [verified] = fetched primary source; [secondary] = mentioned but not verified at source; [inference]; [speculative].

1. Lens & scope

What the R/Bioconductor ecosystem already offers for doing statistical-genomics analysis well (Family A) and refereeing its statistical validity (Family B), plus the emerging LLM/agent layer over R/Bioconductor — assessed against our goal of a portable, inspectable "skills" KB that keeps an LLM both productive and statistically honest.

2. Landscape inventory

Name URL What it does Family Form Maturity Relevance to "rails for an LLM"
OSCA (Orchestrating Single-Cell Analysis) bioconductor.org/books/release/OSCA Book-style end-to-end scRNA-seq workflows (QC → norm → feature sel → clustering → DE/annotation) A book/vignette mature, peer-reviewed (Nat Methods 2020) High — the "do-the-analysis, established method, reproducible" curriculum; canonical Family-A skill-content source [verified]
Bioconductor workflow packages (biocViews: Workflow) + vignettes contributions.bioconductor.org Per-domain validated analysis recipes shipped as packages with executable vignettes A book/vignette + R package mature High — vignettes are the closest existing analog to inspectable "skills" with runnable code [verified]
Bioconductor package review / standards contributions.bioconductor.org/review-expectation.html Open GitHub peer review gating every package: build/tests/vignettes/docs/code style/interoperability neither (B-adjacent) review-process mature Medium — a refereeing mechanism, but reviews software quality, not statistical validity ("ease-of-use, documentation, well-written code, interoperability"; reviewers "should not request reassignment for lack of domain expertise"). Key negative result: the field's main gate does NOT referee method validity [verified]
countsplit / data-thinning (Neufeld, Witten lab) anna-neufeld.github.io/countsplit ; github.com/anna-neufeld/countsplit Splits a count matrix into independent folds (Poisson/NB thinning) so you cluster on one fold and test DE on another — the principled fix for post-clustering double-dipping B R package mature (CRAN; PMC 2024/2025) Very high — a named invalidity (double-dipping) with a concrete empirical remedy; archetype of a Family-B "construct-the-check" skill [verified]
DHARMa (Hartig) cran.r-project.org/.../DHARMa.html Simulation-based scaled residuals + tests for over/under-dispersion, zero-inflation, autocorrelation in GLM/GLMM/GAM B R package mature Very high — turns "are model assumptions met?" into an empirical, simulation-based diagnostic; template for assumption-audit skills [verified]
simr (Green & MacLeod) cran.r-project.org/package=simr ; github.com/pitakakariki/simr Power analysis for GLMMs by Monte-Carlo simulation; power curves vs sample size from a (pilot or synthetic) fitted model B R package mature (Methods Ecol Evol 2016) High — power-by-simulation under known truth = a core Family-B empirical check [verified]
pwr CRAN Task View: ReproducibleResearch Analytical power for t-test/ANOVA/chi²/correlation/proportions B R package mature Medium — closed-form power for simple designs; complements simr [secondary]
sva (incl. ComBat, ComBat-seq, svaseq) bioconductor.org/packages/sva Surrogate-variable estimation + known-batch removal; ComBat-seq does NB-based batch adjustment for counts A→B boundary R package mature (sva 3.59 dev, Jan 2026) High — confound/batch handling is a validity precondition; surrogate variables address unknown confounders [verified]
RUVSeq Bioconductor Removes unwanted variation using control genes / replicate samples A→B boundary R package mature High — same confound-control role; control-gene-based design is itself a "negative control" idiom [secondary]
scDblFinder bioconductor.org/packages/scDblFinder ; PMC9204188 Doublet detection via simulated-doublet density + iterative classification; top-ranked in independent benchmarks A (QC) R package mature Medium — QC guardrail built on simulation under known truth (simulated doublets) [verified]
DE-tool built-in diagnostics (DESeq2 / edgeR / limma) bioconductor.org/.../DESeq2.html Dispersion-fit plots, MA plots, p-value histograms, independent filtering, mean-variance/voom trend B (embedded) R package mature High — the field's everyday validity checks (p-value-histogram shape, dispersion fit) are already standardized; ripe to encode as referee checks [verified]
targets (Landau) books.ropensci.org/targets Make-like dependency-tracked R pipelines; reruns only stale steps reproducibility reproducibility-tool mature Medium — deterministic re-execution substrate; supports "doing never self-certifies" [verified]
renv CRAN Task View: ReproducibleResearch Project-local package library + lockfile; restore on another machine reproducibility reproducibility-tool mature Medium — pins the environment so a referee's checks are reproducible [verified]
workflowr CRAN Task View Version-controlled, literate, web-published reproducible analysis projects reproducibility reproducibility-tool mature Low–Medium — provenance/transparency scaffolding [secondary]
BiocWorkflowTools Bioconductor Authoring helpers to publish Bioconductor workflow vignettes (e.g. to F1000) reproducibility/authoring reproducibility-tool mature, niche Low — relevant only to packaging "skills as vignettes" [inference]
Biomni (Snap/Stanford) github.com/snap-stanford/Biomni ; biorxiv 2025.05.30.656746 General-purpose biomedical agent; LLM + retrieval-augmented planning + code execution; 105 software tools, 150 bio tools, 59 DBs, 25 domains incl. scRNA-seq annotation A (LLM) LLM-agent active 2025 High (as foil) — broad "doing" agent, but no built-in statistical validation / self-check; runs LLM code with full privileges, relies on human oversight. Shows the exact gap we fill [verified]
CellAgent / CASSIA / scExtract / AnnoAgent LLM4Cell survey arxiv 2510.07793 LLM (often multi-agent) frameworks for automated scRNA-seq analysis & cell-type annotation A (LLM) LLM-agent active 2024–2025 Medium — automate "doing"; annotation-focused, validity checks largely absent/heuristic [secondary]
BioMANIA biorxiv 2023.10.29.564479 NL→bioinformatics: learns a Python tool's API (e.g. Scanpy) and executes workflows conversationally A (LLM) LLM-agent proto/active Medium — tool-wiring assistant; no statistical refereeing [verified]
MCPmed arxiv 2507.08055 ; PMC12927880 Position/initiative: MCP-wrap bioinformatics web services (GEO, STRING, UCSC Cell Browser) for agent discovery/invocation/verification enabling LLM-agent infra new 2025 (call-to-action) Medium — infra to make tools agent-actionable; "verification" framing aligns, but vision paper [verified]
POPPER (agentic sequential falsification) arxiv 2502.09858 Validates LLM-generated hypotheses by running sequential statistical tests with error-budget control to keep Type-I error below nominal B LLM-agent research 2025 Very high — a working "referee gate": empirical falsification + provable Type-I control over LLM claims; biology-applicable [verified]
"When Stability Fails" / BioAgentBench / agent-stats-selection evals arxiv 2603.15840 ; 2601.21800 ; PMC12627256 Evidence that LLM agents pick wrong tests, that stability ≠ correctness, and benchmarks for bioinformatics agents B (evidence) reporting/eval research 2025–2026 High — documents the precise failure (plausible-but-invalid, self-consistent) motivating the thesis [verified]

3. Closest prior art (top 3)

  1. POPPER (2502.09858) [verified] — closest to our structural bet: an agent that does not self-certify but hands hypotheses to an external sequential-falsification gate with provable Type-I error control. The Family-B "doing-hands-off-to-referee" pattern, already implemented and biology-applicable. Differs: validates claims/hypotheses, not method choices; isn't a portable inspectable skills KB.
  2. countsplit / data-thinning [verified] — canonical worked example of the whole thesis at the method level: a named invalidity (double-dipping / circular inference) paired with a concrete empirical remedy. Family B wants a "countsplit-style remedy library."
  3. DHARMa [verified] — the model-assumption analog: converts "are the assumptions met?" into simulation-based, interpretable empirical diagnostics with explicit tests. Pattern to generalize across model families.

4. White space / gaps

  • No statistical-validity referee in the standard gates. Bioconductor's review process — the field's flagship refereeing mechanism — explicitly checks software, not method validity [verified]. The "did you double-dip / confound / violate assumptions / inflate Type-I" review is left entirely to ad-hoc human peer review. This is the central gap Family B targets.
  • Diagnostics are scattered, not orchestrated. DHARMa (assumptions), countsplit (double-dipping), simr (power), sva/RUVSeq (confounders), p-value-histogram conventions all exist as individual tools; nobody packages them into a portable, inspectable checklist a referee agent runs ("for analysis of type X, here are the empirical checks the field trusts").
  • The "invented method" failure is essentially unguarded. Existing R tooling validates known methods' assumptions; nothing systematically detects an LLM fabricating a plausible-but-novel method. Closest defense is POPPER-style falsification of the output claim, not auditing the method's derivation — a genuine open niche.
  • "Doing" agents lack the referee handoff. Biomni, CellAgent, BioMANIA, scExtract etc. optimize productivity (wiring/scale/reproduction) and explicitly lean on human oversight for correctness [verified for Biomni/BioMANIA] — exactly the productive-but-not-honest profile our project flags.
  • Negative-control / null-calibration as a first-class skill is underdeveloped: control genes (RUVSeq), simulated doublets (scDblFinder), permutation nulls exist in pockets, but there's no unifying "construct the negative control / permutation null the field trusts here" abstraction.

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

  • Family A doers: Biomni [verified], CellAgent/CASSIA/scExtract/AnnoAgent (LLM4Cell survey) [secondary], BioMANIA [verified]. All productivity-first; correctness ≈ human-in-loop or heuristic.
  • Family B referees (rare): POPPER [verified] stands out — empirical-falsification gate with Type-I control. No R/Bioconductor-native referee agent found.
  • Enabling infra: MCPmed [verified] (agent-actionable tool layer with a "verification" aspiration).
  • Evidence base for the thesis: "When Stability Fails" (stability ≠ correctness) [verified], BioAgentBench [secondary], statistical-test-selection eval showing LLMs mis-select tests [secondary]. These corroborate the "plausible-but-invalid, self-certified" failure mode.
  • Honest caveat: found no project combining (a) an R/Bioconductor skills KB for doing with (b) a non-self-certifying statistical referee gate using the field's empirical checks. That combination appears to be white space [inference].

6. Follow-up URLs worth a deeper read

One-line takeaway: the R/Bioconductor world has a rich, mature catalog of empirical validity checks (countsplit, DHARMa, simr, sva/RUVSeq, p-value-histogram/dispersion diagnostics) and excellent doing curricula (OSCA), but its institutional referee gate (package review) and its new LLM agents (Biomni et al.) both decline to referee statistical validity — leaving our "non-self-certifying referee that runs the field's trusted empirical checks" as genuine white space, with POPPER the nearest agentic prior art.