Mechanism proof

Mechanism verification — pathway hypotheses, Swift checks, Lean receipts.

How pathway hypotheses are encoded, Swift-checked, and discharged with Lean receipts.

New: pathway intelligence + Lean verification

DiscoveryLab now verifies why a peptide should work — not just whether it scores well.

The discovery engine has been extended with pathway-mechanism hypotheses, graph-level Swift verification, Lean 4 audit artifacts, a reusable package plan, an external verifier CLI, perturbation evidence scoring, and a claim-boundary ledger. The result is an auditable bridge from AI-generated biological rationale to wet-lab testing.

implemented · 01

Pathway mechanism hypotheses

Candidate peptides can now carry a typed mechanism: biological nodes, causal edges, intervention-blocked nodes, reachability goals, adverse-pathway blockade claims, safety claims, and conservation laws.

implemented · 02

Graph-level verification in Swift

The local verifier checks whether a therapeutic endpoint is reachable, whether adverse endpoints are blocked under intervention, whether protected pathways remain safe, and whether declared reactions are conserved.

implemented · 03

Lean 4 audit artifacts

Each mechanism report can emit a Lean module with node types, reachability checks, theorem names, and a checksum binding the formal artifact to the candidate report. The Evidence workspace now exposes these receipts in an audit drawer.

implemented · 04

PathwayLean package plan

The app now surfaces a reusable Lean package plan with package name, version, generated file list, checksum, and caveats so formal artifacts can move toward a public pathway-lean library.

implemented · 05

External Lean verifier CLI

A new peptiter-lean-verify executable writes Lean source, runs a configured Lean binary or dry-run path, validates checksums, captures diagnostics, and emits a JSON verification receipt for CI or regulated review.

implemented · 06

Perturbation evidence loop

Wet-lab, transcriptomic, CRISPR, chemical perturbation, and partner assay records can now be attached to pathway assumptions and scored for coverage, agreement, disagreement, and missing evidence.

implemented · 07

In-silico lab assay integration

The lab now exposes mechanismVerification and perturbationEvidence assays so pathway logic and evidence support can participate in candidate ranking before LabSpace handoff.

implemented · 08

Claim-boundary ledger

Mechanism, structure, heuristic, calibration, and wet-lab claims now share one ledger so the product can say what each result is allowed to support and what language is blocked.

Verification claims

Reachability

Can this peptide plausibly move the selected pathway toward the desired endpoint?

Blockade

Does the proposed intervention block a causal route into an adverse endpoint?

Safety

Are protected nodes spared, still reachable, or explicitly marked as unresolved?

Conservation

Do reaction-level assumptions preserve declared molecular quantities?

Evidence

Which assumptions are supported, contradicted, or still untested?

Audit

Can the mechanism artifact be reproduced and verified by Lean in CI?

mechanism audit · DL-MECH-0421Lean 4 · receipt JSON

source

LLM + graph-AI proposal anchored to curated pathway evidence

baseline

GLP1R → cAMP → insulin-secretion endpoint reachable

intervention

adverse appetite-pathway branch blocked under selected nodes

conservation

declared reaction balance checks pass for encoded assumptions

perturbation

5 supported · 1 unresolved · 0 contradicted assumptions

receipt

checksum match · theorem names captured · CI-verifiable

peptiter-lean-verify \
  --input DL-MECH-0421.artifact.json \
  --output DL-MECH-0421.receipt.json \
  --work-dir /tmp/peptiter-lean

status: verified
checksumMatches: true
theorems: desiredReachability, adverseBlockade, protectedSafety

What comes next

Import curated pathway sources

Add SBML, BioPAX, Reactome, and Open Targets importers so mechanism graphs are generated from curated biological knowledge instead of only hand-authored candidate hypotheses.

LLM / graph-AI proposal service

Connect retrieval, biomedical knowledge graphs, and structured prompting so the model proposes mechanisms with citations, then hands them to the verifier as explicit assumptions.

Lean proof hardening

Move from bounded executable checks to richer proof templates, signed receipts, CI retention, and optional reviewer-facing proof bundles for high-value candidates.