Skip to content

minimax-ai/open-protein-binder-design

v0.1.0MIT

Replay, audit, and evaluate protein binder design campaigns with the openbinder CLI — source registry, protocol compiler, campaign replay, ranking audit, agent decision evaluation, external handoff, and evidence-graded final reports. No protein models, no GPU.

GitHubOpen Website

open-protein-binder-design

Replay, audit, and evaluate protein binder design campaigns in MiniMax Code — seven Skills on top of the openbinder CLI, covering source registration, protocol compilation, campaign replay, ranking audit, agent decision evaluation, external handoff, and evidence-graded final reports, without running any protein model.

The problem

Anthropic's public Claude binder design release (protocol, campaign records, candidates, precomputed scores, wet-lab labels) is a rare window into how an AI agent runs a real science campaign. But a raw data dump is not a workflow: from the files alone you cannot tell what was known and decided at each step, whether the published ranking survives recomputation, or how another model would decide given the exact same inputs. This Plugin turns the release into a replayable, auditable, comparable workflow — with strict gap discipline (missing is reported as missing, never interpolated) and byte-identical, wall-clock-free artifacts that diff cleanly under review.

Try it

回放一下内置的 synthetic-failure campaign，给我时间线要点和全部 gap

用 synthetic-ranking fixture 审计这个 campaign 的排序：recorded 和 recomputed 对得上吗

用 mock 和 replay 两个 provider 跑一遍 decision case 评测并对比结果

Expected result: the agent runs openbinder replay --fixture synthetic-failure and returns a four-section report (Summary / Timeline / Decisions / Gaps) with every skipped stage and unfinalized event honestly listed, never filled in; openbinder audit-ranking --fixture synthetic-ranking recomputes ensemble scores and ranking metrics (top-N hit rate, average precision, Spearman, random baseline) and presents recorded vs recomputed side by side — discrepancies are findings, not errors; openbinder evaluate-agent --compare mock,replay freezes each provider's structured decision (SHA-256) before any scoring and reports protocol adherence, evidence usage, fabrication signals, and uncertainty handling. All three run fully offline.

What it does

  • binder-source — data source registry: list / inspect / lock sources (name, Hugging Face repo, pinned revision, license, attribution), with whitelist-only downloads (≤ 20 MB per file, ≤ 50 MB per session, deny by default; blocked items are reported, never bypassed).
  • binder-protocol — protocol compiler: turns a prose protocol draft into a structured, stage-addressable protocol.yaml (ten canonical stages with explicit inputs, allowed decisions, required outputs, hard rules, stop conditions), validates it, and cuts per-stage slices for planning.
  • binder-replay — campaign replay: normalizes public campaign provenance into a unified event stream and rebuilds what was known and decided at each step, for successful and failed campaigns. Gaps are reported, never interpolated; artifacts are byte-identical across runs.
  • binder-agent-eval — agent decision evaluation: MiniMax, Anthropic, ReplayProvider, or MockProvider produce structured JSON decisions on identical stage inputs; a leakage guard scans the assembled prompt before any model call (exit 3 on hit), decisions are frozen before scoring, and reports carry a fixed method-and-limitations section.
  • binder-ranking-audit — ranking audit: recomputes ensemble scores and ranking metrics from published precomputed scores and wet-lab labels, side by side with the recorded ranking. Missing scores are skipped and named; missing labels are excluded — never treated as zero or as non-binder.
  • binder-report — final report bundle: collects replay / eval / audit artifacts into one deliverable, labels every substantive statement with one of six evidence grades (recorded / recomputed / model-generated / externally imported / experimentally measured / unavailable), and writes limitations.md.
  • binder-handoff — external compute handoff: exports a declarative seven-file task package for an external platform (human-gated by --confirm, fail-closed), validates returned result packages (extension whitelist, path-traversal checks, required-file contract), and imports them as data — nothing inside a package is ever executed.

Requirements

  • Python 3.10+ on PATH, and the openbinder CLI installed from the source repository:

    pip install "openbinder @ git+https://github.com/Hylouis233/open-protein-binder-design@v0.1.1"

    (or follow the install instructions in the source repository). Verify with openbinder --help (or python -m openbinder).

  • No API key is required for the offline flows: fixtures, campaign replay, ranking audit, and the mock/replay evaluation providers work with zero credentials and zero network. Without an API key the minimax evaluation provider is unavailable; to enable it, set the OPENBINDER_LLM_* environment variables described in the source repository and check the configuration with openbinder provider doctor --provider minimax first. MiniMax Code gateway users: the agent gateway serves the Anthropic-style messages API only, so also set OPENBINDER_LLM_API_STYLE=anthropic (the doctor detects this automatically and prints the same hint).

  • openbinder source inspect/lock/download and the --source modes of replay/audit need network access to Hugging Face; everything else is offline.

  • The stage-gate approvals and evidence capsules mentioned in the Skills come from the optional openscience Plugin. Without it the Skills degrade gracefully: the agent presents the artifacts and waits for your verbal approve / revise / reject, and the provenance journal (.openscience/provenance.jsonl) is still written by the openbinder CLI itself.

  • Works on Windows, macOS, and Linux.

Boundaries

  • No protein models. The Plugin never runs RFdiffusion, BindCraft, AlphaFold, ESMFold, or any other generative / structure-prediction inference, local or remote. It cannot and does not generate new binder sequences.
  • No GPU anything. CPU-only; no GPU scheduler, no container images, no model weight downloads.
  • No bulk data. Whitelist-only small files; PDB/PAE/sensorgram archives are blocked by policy, and a policy rejection is the discipline working, not a malfunction.
  • Missing is missing. Failed lookups are source gaps, not negative results; replay gaps are never interpolated; uncomputable evaluation cases are excluded and disclosed, never guessed.
  • The evaluation compares research decision-making on identical inputs — it is not a claim about any model's ability to design binders, and mock/replay comparisons are offline controls, not real-model capability conclusions.

Data and network

  • huggingface.co — contacted by openbinder source commands and the --source replay/audit modes to inspect and fetch whitelisted files from the pinned Anthropic/claude-protein-binder-design revision. The release is still uploading at the pinned revision (tables/docs/manifests not yet present), so the real-data ranking channel reports a structured unavailable until the tables land — that is the correct result, not a failure.
  • The minimax / anthropic evaluation providers contact their configured LLM endpoints only when you explicitly select them and provide credentials.
  • No telemetry, no analytics, no callbacks. The package ships zero credentials; all artifacts are written locally under output/binder-<skill>/<slug>/<timestamp>/, with run metadata appended to .openscience/provenance.jsonl by the CLI.

Notes

  • This package is the MiniMax Code port of github.com/Hylouis233/open-protein-binder-design (plugins/science-protein, v0.1.1).
  • The upstream agents (binder-planner, binder-reviewer) are not included — MiniMax Code Agent Plugins do not load them. Their definitions are available in the source repository under plugins/science-protein/agents/; the discipline they encode (stage-slice-only planning, the review-fence output contract) is described inside the Skills themselves.
  • Some Skills reference files that live in the source repository (e.g. docs/scope.md, evals/decision-cases.json, fixtures/); clone the source repository when a workflow needs them.

License

MIT. See LICENSE. Third-party data referenced by the workflow — the Anthropic Claude protein binder design release (protocol text, campaign records, candidate sequences, experimental results) — remains under its original license (CC-BY-4.0 as declared on the dataset page) and is not covered by this package's MIT license. The Plugin registers and references that data by source ID and pinned revision, never by vendoring it. See THIRD_PARTY_NOTICES.md in the source repository for the full notices and attribution requirements.