Bayesian Trials (bioSkills)

A Claude Code skill for designing Bayesian clinical trials — Phase I dose-finding, external-data borrowing with MAP priors, basket and platform designs — with the borrowing weights, stopping thresholds and regulatory anchors made explicit.

   
Type Claude Skill
Supplier GPTomics bioSkills (community OSS, MIT)
Availability GA — part of the bioSkills collection
Pricing Free / OSS (MIT) — RBesT, OncoBayes2, BOIN, dfcrm, escalation, trialr, bayesDP, psborrow2, bhmbasket and c212 are separately installed OSS R packages, several of which compile against Stan.
Capabilities Read/Write — Claude runs the skill’s workflow locally (R), not as an MCP tool
Verified works · 2026-08-10
Security cleared · 2026-08-10 — MIT, provenance matches, bundled CRAN packages open source

How to install

bioSkills is not an npm package — skills are plain markdown/code read directly by the agent. Clone the repo, then either run the installer for the whole category or copy the single skill directory.

  • Claude Code — clone and install via the bundled script: 
    git clone https://github.com/GPTomics/bioSkills
cd bioSkills
./install-claude.sh --categories "clinical-biostatistics"

    The installer copies matching skills into ~/.claude/skills/ (default target). Use ./install-claude.sh --list to preview the skills first.

  • Claude Code / other agents — copy just this one skill: 
    cp -r bioSkills/clinical-biostatistics/bayesian-trials ~/.claude/skills/

    (run from inside your clone — the previous step left you in bioSkills/; otherwise replace bioSkills/ with the absolute path of your clone). Install the R packages on first use, e.g. install.packages(c("RBesT", "BOIN", "dfcrm", "OncoBayes2")). RBesT and OncoBayes2 build on Stan, so expect a working C++ toolchain and a multi-minute first install.

What it does

Four design families, each with its package and its calibrated defaults:

  • Phase I dose-finding — BOIN (primary, get.boundary() / get.oc()), CRM and EWOC via dfcrm, mTPI-2 and Keyboard via escalation. Standard oncology target DLT rate 30%, with escalation below 0.6 × target and de-escalation above 1.4 × target; an early-stop futility floor at 12 patients on the lowest dose; CRM indifference-interval halfwidth 0.05 (Lee–Cheung 2009); EWOC’s overdose constraint P(dose > MTD) ≤ 0.25 (Babb–Rogatko–Zacks 1998).
  • External-data borrowing — meta-analytic-predictive priors with RBesT (gMAP(), robustify()). Target a MAP prior effective sample size of 20–80% of the new control arm (Schmidli 2014), and protect against prior-data conflict with a robust mixture weight of 0.1–0.3 on the informative component. Power-prior discounting for paediatric extrapolation runs γ = 0.3–0.6 as a working convention — the skill is explicit that the FDA January 2026 draft does not prescribe a range.
  • Basket trials — EXNEX (OncoBayes2, bhmbasket) with the Neuenschwander 2016 baseline 0.5 / 0.5 exchangeable / non-exchangeable weights, and a required sensitivity sweep over 0.1, 0.3, 0.5, 0.7, 0.9.
  • Platform trials and safety — the I-SPY 2, GBM AGILE and REMAP-CAP patterns, with I-SPY 2 graduation at a posterior predictive probability of Phase 3 success ≥ 0.85; Berry–Berry three-level hierarchical AE modelling via c212.

Diagnostics the skill requires before any posterior is reported: Stan R-hat < 1.01 (Vehtari 2021) and effective sample size > 1000 per chain.

Primary use cases: designing a Phase I dose-escalation study, borrowing a historical control arm, planning a basket or platform trial.

Notes

Borrowing weights are the design, not a detail. Every threshold above is a prior choice made before data are seen, and the skill treats the sensitivity sweep — over EXNEX weights, over robust-mixture weights, over the power-prior discount — as mandatory reporting rather than an optional extra. A single-weight analysis is not interpretable.

Regulatory anchors cited: the FDA Bayesian Devices Guidance (2010), the FDA draft guidance on Bayesian methodology in drug development (January 2026), the BOIN Fit-for-Purpose qualification (December 2021), and Project Optimus dose-optimisation. The January 2026 FDA document is a draft — treat it as one when writing an SAP.

An MTD chosen by BOIN or CRM answers a toxicity question, not a dose-optimisation question; Project Optimus expects randomised dose comparison for efficacy, which is a separate design.

Distributed as a SKILL.md (plus reference material) in the bioSkills collection — Claude executes the workflow locally rather than as an MCP server. The upstream skill front-matter name is bio-clinical-biostatistics-bayesian-trials; if invoked as a namespaced plugin command it resolves under the bioSkills plugin, not as a bare /bayesian-trials. Upstream directory: clinical-biostatistics/bayesian-trials.

Related: Adaptive Designs (frequentist group-sequential and SSR machinery, and the same BOIN/dfcrm/trialr dose-finding packages), Subgroup Analysis (Bayesian shrinkage across subgroups uses the same EXNEX/MAP toolkit), Power and Sample Size, PyMC for general-purpose Bayesian modelling outside the trial-design frame.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.