How PumpPepsLab works, end-to-end.

PumpPepsLab is an autonomous research platform for nonclinical peptide discovery. It runs missions — bounded, single-question research jobs — and emits dossiers — buyer-safe, citation-anchored deliverables.

Every mission goes through a 5-layer DAG of agent pools. Layer 1 ingests evidence; layer 2 annotates and scouts novelty; layer 3 grades evidence A → X; layer 4 reasons + critiques; layer 5 synthesizes the dossier. The dossier never assumes more than the evidence allows.

• 01Literature Miner

  PubMed E-utilities. Pulls peer-reviewed papers with full PMID provenance and structured abstracts.

  upstream
• 02Sequence & Structure

  UniProt + PDB + AlphaFold. Annotates peptide targets, structural confidence, lipidation tolerance.

  upstream
• 03Target & Pathway

  OpenTargets GraphQL + Reactome. Maps targets to diseases and mechanism-of-action pathways.

  upstream
• 04Variant Linker

  ChEMBL REST. Links targets to known small molecules and binding-assay endpoints.

  upstream
• 05ADMET Developability

  Heuristic ADMET scorecards: solubility, half-life, peptide developability, off-target risk.

  upstream
• 06Novelty Scout

  Whitespace detector. Scores each finding for novelty against historical mission corpus.

  reasoning
• 07Patent Competitive

  Patent-density and freedom-to-operate signals approximated from ChEMBL ligand counts and PubMed coverage per target.

  reasoning
• 08Thesis Generator

  Composes structured, falsifiable hypotheses with evidence-cited mechanism claims.

  reasoning
• 09Evidence Grader

  Grades each finding A/B/C/D/X using study type, replication and reporting strength.

  reasoning
• 10Red Team

  Three personas — Skeptic / Scientist / Senior Reviewer. Only Senior can issue a hard block.

  reasoning
• 11Synthesizer

  Consolidates cross-pool findings and theses into the synthesis document.

  output
• 12Dossier Assembler

  Buyer-safe markdown dossier with PMID citations, hedged claims and full evidence chain.

  output

Before any agent runs, PumpPepsLab computes:

message = JSON.stringify({
  query: "...",
  target_class: "...",
  schema: "pumppepslab.commit.v1",
  salt: <random 16 bytes hex>,
});

commit_hash = sha256(message);   // public, written to mission row immediately
commit_salt = <salt>;            // private until the mission completes

On completion, PumpPepsLab publishes the salt. Anyone can re-hash the original question + salt and verify the run was honest end-to-end. If a mission is aborted before completion, the salt remains sealed and the commit hash stays as a public commitment that no answer was ever delivered for that question.

Every finding is graded with an explicit rubric. The rubric is shipped with PumpPepsLab, not learned, not opaque.

• A

  Multiple independent peer-reviewed studies, replicated, with concordant readouts.

• B

  Single peer-reviewed study with rigorous methodology, or strong concordance from indirect sources.

• C

  Preprint, conference, or single-method evidence; plausible but not replicated.

• D

  Indirect inference, weak methodology, or fragile single-source claim.

• X

  Insufficient or contradictory evidence; cannot ground a thesis.

• PubMed (NCBI E-utilities)esearch + efetch, PMID-anchored citations
• UniProtProtein entries, taxonomy, function annotations
• AlphaFoldPredicted structures and pLDDT confidence per residue
• OpenTargets GraphQLTarget ↔ disease associations + therapeutic areas
• ChEMBL RESTTargets, ligands, bioactivity priors
• ReactomePathway membership and cross-references

Every dossier follows the same skeleton, by construction:

## Question
<the original mission query, verbatim>

## Cross-pool consensus
- Literature ...
- Sequence/structure ...
- Target/pathway ...
- ChEMBL ligand prior ...

## Open questions
## Risks
## Recommended next steps

The dossier is deterministic given the upstream evidence. It does not invent claims, does not embed PMIDs that weren’t retrieved, and never makes a human-use claim.

• POST/api/missions

  Start a mission. Body: { query, target_class?, depth?, budget_cents? }. Returns 202 + mission_id.

• GET/api/missions

  List missions, latest 50.

• GET/api/missions/:id

  Mission state: tasks, findings, theses, critiques, dossiers.

• GET/api/dashboard?mission_id=…

  Aggregated dashboard payload (the same one /app uses).

• GET/api/stream?mission_id=…

  Server-Sent Events feed of swarm events. Heartbeats every ~700ms.

PumpPepsLab’s reasoning agents — Thesis Generator, Red Team and Synthesizer — call POST https://api.kie.ai/codex/v1/responses with model gpt-5-4. Reasoning effort is configurable per call (low → xhigh).

POST https://api.kie.ai/codex/v1/responses
authorization: Bearer $KIE_API_KEY
content-type: application/json

{
  "model": "gpt-5-4",
  "stream": false,
  "input": [
    { "role": "user", "content": [{ "type": "input_text", "text": "..." }] }
  ],
  "reasoning": { "effort": "medium" }
}

KIE_API_KEY is required everywhere — there is no templated fallback in any environment. Every reasoning call is real model output. A missing key, a network error, or a non-2xx response from kie.ai aborts the task hard so a failure can never silently masquerade as a real result. The same rule applies to the Red Team and Synthesizer: if the model output cannot be parsed into the expected shape, the task fails rather than writing a placeholder critique or synthesis to the ledger.