Nine-Gate Discipline for Computational Drug Discovery
PeptAI’s published pipeline for autonomous peptide drug discovery. Each agent (one per receptor) runs all nine gates; only candidates clearing G1–G8 earn a wet-lab synthesis recommendation, and G9 cannot be computationally overridden.
| Gate | What it checks | Tools |
|---|---|---|
| G0 | Per-gate baseline calibration on known agonists and non-binders | ChEMBL binding data |
| G1 | Predicted structure quality | AlphaFold |
| G2 | Binding pose plausibility | Boltz2 |
| G3 | Contact conservation across known agonists | Boltz2 + ChEMBL agonist set |
| G4 | Computed binding affinity | PRODIGY |
| G5 | Conformational stability under MD | LiteFold MD |
| G6 | Proteolytic stability | PROSPERousPlus |
| G7 | Solubility & aggregation propensity | OpenSol, PlifePred |
| G8 | Off-target / toxicity safety | ToxinPred3 |
| G9 | Wet-lab synthesis & assay handoff | adaptyvbio (machine-to-machine via x402) |
Two design properties make this a discipline pattern, not just a tool list:
- Thresholds calibrated against ground truth. G0 sets per-gate baselines using ChEMBL agonists and non-binders, so each downstream gate inherits a defensible pass/fail line rather than an arbitrary cutoff.
- Open record by default. Every gate decision is published on Molecule Labs — including failures. The pipeline cannot be retroactively cleaned up, so the discipline is enforced by the audit trail.
Transferability — what generalises and what doesn’t
The literal gate set is calibrated for short peptide agonists at GPCRs in systemic circulation. Three structural assumptions break in other contexts:
- G0 / G3 assume a known agonist set against the same receptor. If the target is a structural protein (e.g., stereocilin) or a fold-rescue chaperone target, there are no agonist analogs to calibrate against; this gate has to be replaced with a different ground truth (e.g., known-pathogenic vs known-benign variants for a chaperone).
- G6–G8 are calibrated for blood-borne peptide pharmacokinetics. For tissues with bespoke barriers — cochlea, CNS, intracellular targets — the relevant ADMET-equivalents are different (e.g., round-window membrane permeability, perilymph clearance time).
- G9 + x402 + adaptyvbio assumes commodity peptide synthesis is the wet-lab step. Hypotheses requiring AAV cloning, mRNA-LNP formulation, or hydrogel self-assembly need a different G9 contract.
The transferable kernel is the structure: declare per-stage thresholds against a baseline, automate the pass/fail decision, log every result whether it advances or not.
Where this could apply in STRC
- h26 (engineered disulfide homodimer) — the closest fit. Mini-protein design with an explicit binding/folding objective. G1 (AF3 quality), G2 (Boltz binding to STRC monomer), G4 (interface energy threshold), G5 (MD stability of the disulfide pair) all map cleanly. G6–G8 can be skipped or replaced.
- h09 (peptide hydrogel HTC scaffold) — partial fit. G1–G2 apply for any peptide structure work, but the pipeline is designed around receptor binding, not self-assembly; G3 contact conservation has no direct analog for a hydrogel.
- h01 (E1659A pharmacochaperone) — different gate set but same discipline. The ground truth is fold-stabilisation magnitude (ΔΔG_fold), not receptor binding. Gates would be: hit rate from virtual screen → docking pose quality → MD stability of stabilised mutant → FEP ΔΔG threshold → ADMET for cochlear delivery (round window → perilymph PK).
Connections
[source]2026-04-25-bankless-peptai-desci-drug-discovery[applies]STRC Gene Therapy- Agent vs Tool Distinction — the conceptual basis: gates are agent decisions, not human ones
- Computational Confidence Scores as Epistemic Tools — gates are confidence scores tied to a specific advancement decision
[see-also]scientific-research-pipeline- External (STRC vault):