Asclepius

Live: https://asclepius-bio.vercel.app · methodology index · how the tool thinks — a worked case · worked example

A biotech research / analysis / advising tool. It treats two structural facts as first-class value drivers that generic valuation tools ignore — who holds the capital and who controls the supply — and it produces a defensible view on a single asset: a call, a conviction, and the one readout that would flip it. Every number ships its citation.

Asclepius is used by its author and a small trusted circle to analyze real pre-approval biotech assets. It is not a public product, a SaaS, or a forecasting service. Its value is the creativity, architecture, and methodology it embodies — a framework that forms a view, shows its work, and names what would change its mind.

The differentiators

1. Two named structural path-dependencies. Most tools price a drug as a probability-weighted cash-flow stream and stop. This one adds:

   • Reflexivity — a Spence-style signaling multiplier on probability-of-success keyed to sponsor capital state. Capital-constrained sponsors run worse trials; well-capitalized sponsors enable adaptive designs and stronger regulatory engagement. Grounded in Spence (1973) with modern empirical support (Kao 2024, Lo & Thakor 2022, Ma 2025).
   • Supply constraint — a manufacturability multiplier on PoS and a peak-sales ceiling in rNPV. A binding isotope/raw-material/capacity constraint is execution risk on the way in and a revenue cap at the top. Generic DCF models put neither on the revenue line.

   Both are reasoned structural hypotheses, grounded in named sources and honestly labeled as estimates, not calibrated corrections.

2. A recommendation that earns itself. The memo synthesis doesn't print a number — it forms a view: an explicit conviction level (orthogonal to buy/hold), a one-line verdict on how the path-dependencies net-moved the call, a priced close, and a falsifiable kill criterion (a named event with a threshold and a direction). The schema forces all four — a memo cannot come back without them.

3. An architecture that generalizes. A single DiligenceRecord pydantic model is the source of truth; three auto-discovered registries (data_sources/, modules/, agents/) plug into FastAPI on startup; the Next.js frontend renders one panel per module from the /api/modules manifest. Adding the entire supply-constraint dimension (a new path-dependency affecting two engines) required no core edits — it slotted into the same waterfall position reflexivity occupies. See docs/architecture.md.

A worked case

The showcase narrates one hard asset end to end — an Actinium-225 alpha-emitter from a capital-constrained junior — the one asset that lights up every differentiator at once (both path-dependencies bite, and it routes to the curated radiopharmaceutical comp cohort). Every number on that page is real engine output; the reading is editorial.

The canonical retrospective calibration example is adagrasib, run on only public information available before the KRYSTAL-12 Phase 3 readout (June 2022 cutoff):

Scenario	Framework value	Reality
Rule-based PoS (BIO base → reflexivity → final LOA)	~13%	—
ML PoS Prior (second opinion; PubMedBERT + LightGBM on HINT)	~24%	Real-outcome-supervised
Pre-readout (Phase 2) base-case rNPV	~$570M	Mirati's market cap at cutoff
Post-readout (at NDA) base case	~$4.9B	—
Actual BMS acquisition (Oct 2022)	$4.8B	Brackets within ~2%

BMS paid a small premium to the success-weighted central case — defensible as strategic KRAS-franchise value, no margin of safety for a financial buyer. Full write-up: 05-worked-example-adagrasib.md. This is calibration against a known outcome, not a prediction.

Try it

# Backend (Python 3.11+)
cd api
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
uvicorn app.main:app --reload --port 8000

# Frontend (Node 20+, pnpm 10+)
cd web
pnpm install
pnpm dev

Then open http://localhost:3000/diligence/adagrasib.

Two hero interactions:

1. Reflexivity slider (top of the Thesis section). Drag from "Adequate" to "Well capitalized" and watch the LOA microsplit, PoS waterfall, ML PoS band, rNPV base case, Monte Carlo distribution, and tornado all recompute live — the headline differentiator visibly moving rNPV.
2. Persona toggle (global header). Cycle VC Associate / IC Voter / Scientific Reviewer / Quant and watch the whole page reorganize around each reader's first question. Same data, four reads — "methodology applied to its own presentation" (product thesis).

What's in the build

The deterministic spine:

• PoS engine — a base × Π(multipliers) chain (base rate → modality → mechanism/biomarker → target validation → regulatory designations → reflexivity → supply constraint), every step logged in the audit trail with a primary-source citation. Frontier modalities (degrader, radiopharmaceutical, epigenetic editing) carry honest "Asclepius estimate, no BIO/Informa cohort" provenance rather than borrowing a citation they aren't entitled to.
• rNPV engine — phase-gated cash flows, mid-year discount convention, 10,000-path Monte Carlo (log-normal peak / beta LOA / normal WACC / triangular COGS), tornado sensitivity, and the supply-constraint peak-sales ceiling.
• 8-pillar diligence scorecard — including a computational-infrastructure pillar that prices a sponsor's AI/ML/data assets.
• Cited deal comparables — cohort routing by (therapeutic area, modality family) with a curated radiopharma cohort and an honest same-TA-approximation fallback; provenance on every row.

Three runtime agents (each one UI button; coordination via the deterministic core):

• Auto-Diligence — asset name → structured AssetInput with citations from a high-signal domain allowlist (CT.gov, FDA, EDGAR, top journals); output coerced to the domain enums at the boundary.
• Memo Writer — diligence record → a structured investment memo with the defensible-recommendation discipline (12-memo-synthesis.md).
• Game-Theory Adversary — a devil's-advocate pass surfacing signaling-equilibrium violations, winner's-curse adjustments, and Bayesian-persuasion flags on selective disclosure.

A second-opinion ML PoS prior — PubMedBERT-embedded eligibility text + structured features → LightGBM (phase-stratified), supervised on real HINT trial→approval outcomes, with a bootstrap-percentile band and split-conformal coverage exposed at /api/modules/ml_pos_prior/model_info. It is a check on the rule-based estimate, not a forecast: agreement raises confidence, divergence flags an asset for closer review.

Product surfaces:

• Persona-aware diligence page — the same record through four audience-specific reads (VC Associate / IC Voter / Scientific Reviewer / Quant).
• Narrated showcase — the "how the tool thinks" case study.
• Saved analyses — persist / list / reload a completed analysis (gated).
• Portfolio sizing — Kelly + conviction + barbell allocation for the fund-manager view (/portfolio).

Access

The live app's AI features (the three agents, saved analyses) are gated behind a shared passphrase so they don't spend the owner's API key — it is a tool for a small trusted circle, not a public service. The deterministic modules and the methodology are open.

Stack

• API: Python 3.11+, FastAPI, pydantic v2, numpy/scipy. Full app test suite passing.
• ML inference: torch (cpu) + transformers + lightgbm; PubMedBERT loaded in-container, the LightGBM artifact shipped with the Docker image.
• Web: Next.js 14 (App Router), Tailwind, recharts, react-markdown, TypeScript.
• Deploy: Vercel (web) + Fly.io (api). See docs/deployment.md.
• CI: GitHub Actions — ruff + pytest + pnpm typecheck + pnpm build on every push.

Reading the methodology folder

A first pass:

1. 00-product-thesis.md — what the tool is and what it's a proof of.
2. 02-reflexivity-thesis.md — the load-bearing claim (capital → PoS).
3. 22-supply-constraint-thesis.md — the second path-dependency (supply → PoS + a revenue ceiling).
4. 12-memo-synthesis.md — what makes a recommendation defensible.
5. 06-signaling-equilibrium.md — the game-theoretic derivation of reflexivity from Spence (1973).
6. 05-worked-example-adagrasib.md — the framework run end to end.

Or browse the methodology index live.

Sources and citations

Every empirical claim cites peer-reviewed work or a primary industry source. Canonical references:

• Spence, M. (1973). Job Market Signaling. Quarterly Journal of Economics.
• Akerlof, G.A. (1970). The Market for "Lemons". Quarterly Journal of Economics.
• Wong, C.H., Siah, K.W., Lo, A.W. (2019). Estimation of clinical trial success rates. Biostatistics 20(2).
• Lo, A.W., Thakor, R.T. (2022). Financing Biomedical Innovation. Annual Review of Financial Economics 14.
• Kao, J. (2024). Information Disclosure and Competitive Dynamics. Management Science.
• Ma, S., et al. (2025). Predicting accrual success for better clinical trial resource allocation. Scientific Reports 15.
• Fu, T., et al. (2022). HINT: Hierarchical Interaction Network for Trial Outcome Prediction. Patterns.
• Gu, Y., et al. (2021). Domain-Specific Language Model Pretraining for Biomedical NLP (PubMedBERT). ACM TOCH.
• BIO / Informa / QLS Advisors (2021). Clinical Development Success Rates 2011–2020.
• Damodaran, A. (2025). Cost of Capital by Industry (US). NYU Stern.

Full per-writeup citations are in methodology/.

License

MIT — see LICENSE. The methodology writeups are released under the same license.