vaxtract 0.1.0

Schema-validated extraction of neoantigen cancer-vaccine immunogenicity data from primary papers, on the Claude Agent SDK (bring-your-own-key).

vaxtract

Schema-validated extraction of neoantigen cancer-vaccine immunogenicity data from primary papers, built on the Claude Agent SDK. Point it at a folder of paper files (PDF / XLSX / DOCX) and it returns a schema-validated, provenance-tracked JSON extraction (per-peptide / per-epitope immunogenicity, HLA restriction, evidence, survival outcomes, …) for human sign-off.

Bring your own key (BYOK). You run the agent and pay for your own Anthropic usage (~$3/paper, varies). Your files never leave your machine — there is no hosted service.

Output is silver, not gold. Every record carries provenance and is meant for a curator to review before use, not to be treated as ground truth.

Install

pip install vaxtract                    # core: the schema/vocab data contract only
pip install "vaxtract[agent]"           # + the extraction agent (Claude Agent SDK + readers)
pip install "vaxtract[agent,figures]"   # + figure/image reading (PyMuPDF + Pillow)

pip install vaxtract pulls only pydantic, so you can import vaxtract.schema to validate records without the Claude Agent SDK. Running the agent (the vaxtract console script or vaxtract.extract_paper) needs the [agent] extra.

Running the agent also requires Python ≥ 3.10 and the Claude Code CLI on your PATH — the Claude Agent SDK shells out to the claude binary:

npm install -g @anthropic-ai/claude-code

(The Docker image below bundles this for you.)

Authenticate (pick one)

# A) API key — pay-per-token
export ANTHROPIC_API_KEY=sk-ant-...

# B) Claude subscription — use a logged-in plan via the `claude` CLI
#    (pass --subscription; the key is ignored)

Run

vaxtract ./my_paper_dir out.json
vaxtract --subscription ./my_paper_dir out.json   # use plan quota

my_paper_dir is a folder containing the paper's .pdf and any supplementary .xlsx / .docx. The agent reads the tables/text/figures, builds the record, self-validates against the schema, and writes out.json.

As a library

import asyncio
from vaxtract import extract_paper

asyncio.run(extract_paper("./my_paper_dir", "out.json"))

The data contract is importable without the SDK:

from vaxtract.schema import ExtractedPaper, SCHEMA_VERSION

What it extracts

Per paper: studies, patients, immunizing peptides, minimal epitopes, pools, immunogenicity evidence (assay/outcome/magnitude), neoantigen mutations, survival outcomes, clinical-benefit signals, safety, and vaccine-delivery covariates — all validated against a versioned Pydantic schema (SCHEMA_VERSION).

Notes

• The agent is restricted to a curated toolset and is headless-safe (no host shell access; it cannot read or write outside the files you give it and the output path).
• Cost/turn backstops (max_turns, max_budget_usd) guard against runaway runs.
• Figure reading is optional; install the [figures] extra and ensure a working PyMuPDF (it bundles its own libraries — no system Poppler needed).

License

MIT © 2026 Samuel Ahuno