BD Decision Intelligence

A multi-agent pharma BD valuation system that evaluates drug assets and produces dual-view deal pricing with buyer mapping.

What it does

Type any drug asset and get:

• Science profile — positioning, de-risking signals, adjusted PTRS
• Comparator analysis — SOC discovery, metric comparison, differentiation verdict
• Three-scenario valuation — standalone NPV, platform displacement, strategic deal price
• Buyer mapping — likely acquirers ranked by urgency (patent cliff + flush capital + deal velocity)
• Dual-view output — risk-adjusted (PTRS applied) and if-succeed (assumes approval) side by side

Architecture

Sequential pipeline — market agent reads PTRS from science agent, synthesizer reads everything.

Three agents do the analytical work:

• Science agent — searches PubMed/ClinicalTrials.gov, builds positioning profile, identifies de-risking signals, computes adjusted PTRS
• Market agent — discovers comparators, runs metric comparison, determines differentiation verdict (best/better/me-too/worse/new-class-creation), sizes dual peak sales
• Synthesizer — maps buyers (patent cliff + flush capital tables), scores bidding tension, computes deal multiple, produces three scenarios with derivation math

The research planner is a lightweight classifier that routes queries (specific asset vs. discovery scan).

Tech stack: FastAPI + LangGraph + Claude Sonnet (backend), React + Vite + Tailwind (frontend)

Project structure

bd-intelligence/
├── backend/
│   ├── agents/           # science_agent, market_agent, synthesizer, research_planner, discovery_agent
│   ├── utils/            # ptrs_lookup, buyer_context, deal_benchmarks
│   ├── tools/            # Tavily search wrappers
│   ├── graph.py          # LangGraph sequential pipeline
│   ├── main.py           # FastAPI: /analyze, /analyze/stream, /recalculate
│   └── state.py          # IndicationAnalysis TypedDict + BDState
├── frontend/
│   └── src/components/   # ChatWindow, ResultCard, ValuationWaterfall, AssumptionsPanel, FilterSidebar
├── docs/
│   ├── agent_logic.md    # Detailed scoring rubric, formulas, tuning guide
│   └── stategraph.png    # Auto-generated graph visualization
└── tests/                # Real deal validation suite

Setup

# Backend
cd bd-intelligence
cp .env.example .env     # add ANTHROPIC_API_KEY and TAVILY_API_KEY
uv sync
uv run uvicorn backend.main:app --reload

# Frontend (separate terminal)
cd frontend
npm install
npm run dev

Backend runs on http://localhost:8000, frontend on http://localhost:5173.

API

POST /analyze/stream (primary)

SSE endpoint — streams progress events then final result.

{"message": "TERN-701 allosteric BCR::ABL1 TKI CML Phase 1/2", "filters": {}}

POST /recalculate

Re-runs from market agent onward with user overrides (e.g. edited PTRS, changed comparator).

{
  "asset_name": "TERN-701",
  "indications": [...],
  "overrides": {"indication_name": "CML", "field": "ptrs_adjusted", "value": 0.35}
}

Valuation methodology

Each scenario shows two values:

Scenario	Risk-adjusted	If-succeed
Standalone NPV	peak x PTRS x rev_mult x NPV	peak x rev_mult x NPV
Platform displacement	+ SOC capture, PTRS applied	+ SOC capture, no PTRS
Strategic deal price	deal_multiple x PTRS	deal_multiple (what buyer pays)

Deal multiple = base(phase) + urgency_adj + bidding_adj + differentiation_adj. See docs/agent_logic.md for the full formula.

Key files

File	What it does
utils/ptrs_lookup.py	Base PTRS table + get_adjusted_ptrs() with 9 de-risking signals
utils/buyer_context.py	Patent cliff (12 buyers) + flush capital (3 buyers) tables
utils/deal_benchmarks.json	20 M&A + 7 licensing deals from 2025-2026
agents/synthesizer.py	Buyer mapping, bidding tension, three-scenario output
docs/agent_logic.md	Complete scoring rubric, formulas, tuning guide