← Proof
Mechanism proof
In-silico lab — assessment matrix and rejection gates.
The in-silico assessment matrix and the rejection gates that prune candidates before any wet-lab cost.
In-silico lab
Before wet lab: automatic computational and mechanism triage.
A standardized assessment matrix runs across every candidate. Failing scores, unverifiable mechanism claims, unsupported assumptions, or over-broad claim language constrain or reject; passing scores produce a claim-bounded priority list.
| Assessment | What it checks | Why it reduces search space |
|---|---|---|
| Scientific claim ledger | Classifies every capability and result as formal verification, physics-bearing reduced model, heuristic surrogate, calibrated prediction, wet-lab observation, or orchestration contract | Prevents heuristic scores from being presented as calibrated probabilities or biological validation |
| Public calibration staging | Imports ChEMBL, BindingDB, PDBbind, and wet-lab rows into typed calibration manifests with splits and checksums | Turns ranking signals into calibrated probabilities only after a held-out calibration artifact exists |
| Wet-lab export package | Generates candidate selection reports, synthesis specs, target-specific assay protocols, result templates, controls, readiness issues, and checksums | The current wet-lab prep package selects 10 candidates across 10 targets and includes 13 synthesis specs plus 10 protocols |
| Protein-LM task heads | ESM-2-width 2560-dimensional embeddings feed six committed task heads retrained as 1024-wide MLPs on normalized public peptide data and exported as precompiled Core ML bundles | Moves descriptor proxies toward local Mac model-backed ranking evidence while keeping uncalibrated neural scores below assay truth |
| BioFoundation context | Cell atlas, protein-sequence, regulatory-variant, perturbation, and external foundation-model plans are converted into scorecards with provider provenance and calibration state | Lets candidates carry biological representation context without allowing rank-only model output to become efficacy, safety, or target-engagement evidence |
| Pathway assertion registry | Release-pinned Reactome, GO-CAM, and Open Targets records carry license policy, evidence namespaces, organism/cell/anatomy context, assertion layer, knowledge status, and confidence | Prevents a pathway prior from being treated as universal biology or redistributed outside its source terms |
| Federated body-twin model cards | Process twins and executable islands declare context of use, calibration state, update cadence, uncertainty method, validation endpoints, privacy boundary, allowed claims, and blocked claims | Keeps reduced glucose, beta-cell, and blood-pressure axes useful for triage without implying clinical whole-body prediction |
| Boltz-1 contact maps | Previews receptor-peptide complex requests, HealthOmics run descriptors, and contact-map evidence for receptor fit and docking pose | Replaces pocket fiction with explicit predicted contacts before wet-lab advancement |
| Receptor / pathway relevance | Alignment of candidate to selected receptor and pathway evidence | Discards candidates with no plausible biological target |
| Pathway mechanism verification | Builds verifier-ready mechanisms from governed assertions, then checks reachability, intervention blockade, protected-node safety, and conservation claims | Rejects candidates whose proposed mechanism does not follow from the encoded and source-qualified pathway assumptions |
| Lean 4 audit artifact | Generates a formal module and external verification receipt for high-value mechanisms | Makes pathway claims reproducible and reviewable instead of hidden inside a score |
| Perturbation evidence support | Scores whether assays, omics, CRISPR, or chemical perturbations support the mechanism assumptions | Prioritizes candidates with testable and convergent biological evidence |
| Sequence motif plausibility | Presence and orientation of binding motifs and pharmacophores | Removes scaffolds that violate known structural priors |
| Stability (predicted) | Half-life proxies, protease cleavage liability, oxidation risk | Filters chemically fragile sequences early |
| Solubility | Predicted aqueous solubility under assay-relevant conditions | Avoids candidates that cannot be tested at meaningful concentrations |
| Aggregation risk | β-sheet propensity, hydrophobic patches, self-association cues | Reduces wet-lab failure from precipitation |
| Synthesis feasibility | SPPS difficulty, cyclization route, ncAA availability | Prevents prioritization of impractical sequences |
| Off-target concerns | Predicted cross-reactivity to related receptors and proteins | Surfaces candidates needing selectivity engineering |
| ncAA compatibility | Compatibility of non-natural residues with target chemistry | Keeps v2 ncAA expansion within feasible bounds |
| Assay readiness | Match between candidate and available wet-lab assays | Ensures handoffs are testable as designed |
| Wet-lab priority | Pareto-front ranking, active-learning acquisition, and assay-readiness checks across all assessments above | Selects the next testable LabSpace batch instead of simply sorting by one MPO score |
More on proof
Related on this track
Mechanism verification
Pathway hypotheses, Swift checks, Lean receipts.
Combination controllers
Lean-verified cross-domain combinations.
PeptCheck v8.6 benchmark
56 scored peptide-drug entries; 100.0% full agreement.
Pathway Walk
visionOS proof topology.
Why rail
Inspectable evidence for candidate decisions.
Evidence tutor
Retrieval-grounded peptide reasoning.
TensFormer
Typed world-model spine.
← Back to Proof index
All sections on the proof track.