One evaluation stack across four design modalities. Form factor follows target.

• 01Linear peptides. Fast, synthesis-friendly, broadest target class coverage.
• 02Head-to-tail cyclic peptides. Higher stability and selectivity; preferred for skin and cosmetic actives.
• 03Mini-proteins. For surfaces too large or shallow for short peptides — receptor surfaces, transcription factor interfaces.
• 04Ternary geometries. PROTAC and molecular-glue ternary modeling for E3 ligase recruitment.

We use the best published models. Our advantage is not in owning the models — it is in choosing them per target, calibrating them against comparators, and naming the conditions under which they lie.

Models in regular use include AlphaFold-family structure prediction, Chai-1, ProteinMPNN, RFdiffusion, AfCycDesign, ESM-family sequence models. Classical methods include PyRosetta energetics and OpenMM molecular dynamics. The stack is updated as the field moves; the discipline layer is constant.

Five practices we apply on every program. They are unfashionable. They are the reason our predictions survive contact with a wet lab.

Canonical sequence verification

Every screen begins with a diff of the target sequence against the canonical UniProt reference, with motif spot-checks asserted in the build script. Sequence chimeras in published AI-bio work are common; we catch them before they invalidate downstream results.

Calibration before screening

Before running predictions on a new target class, we calibrate against a published SAR ladder for a comparator. On our most recent calibration — a 13-peptide published affinity ladder — the pipeline reproduced 90% of the rank-order. We report rank-ordering accuracy, not absolute affinity: a pipeline that cannot rank known binders against known non-binders does not deserve to be trusted on novel candidates.

Independent cross-check

Every candidate is scored by a primary structure-prediction model, then re-scored by a second, independent model. Agreement is a hard filter — candidates the second model does not reproduce do not advance. In a parallel senescence-clearance program, the primary model scored a set of de novo candidates as strong, selective binders; the independent cross-check recovered the published benchmark cleanly but scored those candidates far lower. They were dropped before any synthesis spend. The leads we carry forward are specifically the ones that survive this filter.

Pre-declared failure modes

We publish — under NDA, in scoping discussions — the conditions under which our predictions are known to mislead. Example: structure-prediction confidence gaps without direct atomic contact to a unique residue are noise, not selectivity. We have proved this five times so our clients don't have to.

Kill-or-continue gates

Every program has explicit decision points. We tell clients when to stop spending. The most informative dollar in an AI-driven program is the one that funds the first wet-lab read; we structure the program around that.

The pipeline is a sequential filter. A candidate must clear every gate to advance; a failure at any stage drops it.

• 01Binding pose. Cyclic-peptide complex structure prediction.
• 02Independent cross-check. A second structure-prediction model must agree.
• 03Interface energy. Physics-based interface scoring with explicit relaxation.
• 04Pocket contact. Atomistic contact with the canonical functional pocket residues.
• 05Selectivity counter-screen. The same protocol run against paralogs or isoforms.
• 06Property gate. Permeability and stability indices benchmarked against marketed comparators.
• 07Synthesis feasibility. Aspartimide, proline-junction, and aromatic-burden screens.
• 08Aggregation & liability. Aggregation, PAINS, and oxidation / deamidation motif checks.
• 09Calibration. Reproduce a published affinity rank-order before novel scores are trusted.

Each program runs through the same loop. Cycles complete in weeks, not months.

• —We don't sell tools or seats. We run programs.
• —We don't promise wet-lab-confirmed leads from computation alone. Across the field, roughly 30–50% of modern in-silico cyclic-peptide designs bind when tested — which is exactly why every program ends at a wet lab.
• —We don't take credit for binding affinity our models predict. We hand off to wet-lab validation and report what comes back, positive or negative.
• —We don't engage without an NDA in place before substantive disclosure.

• —Candidate set. De novo cyclic-peptide leads with predicted target engagement, rank-ordered against a published comparator. Standard all-L-amino-acid, head-to-tail cyclic — synthesizable at any peptide CRO, no unnatural building blocks.
• —Selectivity assessment. In silico counter-screen against paralogs or isoforms.
• —Cosmetic-property profile. Skin-permeability index, aggregation risk, and synthesis feasibility — gated against marketed cosmetic peptides.
• —Wet-lab handoff package. RFP-ready CRO submission: synthesis spec, assay panel, controls, budget.
• —Optional. Wet-lab validation managed with our CRO partners; INCI registration support.

• DoneIn silico discovery — candidates designed, cross-checked, scored, property-gated.
• 8–10 wkCRO Phase 1 — binding affinity, functional displacement, selectivity, cellular activity, full controls.
• 12–16 wkCRO Phase 2 (on Phase 1 pass) — mechanism confirmation, skin permeability, stability, irritation.
• 6–12 moCommercial path — INCI registration, EU CPNP / US OTC dossier, formulation or B2B partnership.

End-to-end to a validated, IP-protected cosmetic active is realistically $90–160K per program over 10–14 months — against $1–3M and 18–24 months for a comparable pharma-tier program. The compression comes from replacing the synthesize-screen-iterate loop with computation; it does not remove the need for proper biology.

Cosmetics programs run as co-funded wet-lab validation partnerships. Engagement scales from material-only evaluation, through single-program validation, to running both programs in parallel with shared CRO overhead.

Compare engagement models →

• 01Oncogene paralog discrimination. RAS-family isoform selectivity at the binder level, with selectivity decomposed into computationally identifiable mechanisms.
• 02Undruggable transcription factor surfaces. Shallow protein-protein interfaces on nuclear regulators.
• 03E3 ligase substrate handles. Degron design and substrate-adaptor recognition for targeted protein degradation.
• 04PROTAC ternary geometries. Ternary complex modeling, linker geometry, ligase recruitment efficiency.
• 05Neurodegenerative aggregates. Disordered protein aggregate disruption and stabilization.

Specific program details under NDA. We are happy to share our calibration data on a published comparator for your target class as part of scoping.

• —Hit set. De novo peptide or mini-protein candidates with predicted target engagement and selectivity, rank-ordered against a published comparator.
• —Modality recommendation. Linear, cyclic, mini-protein, or ternary — chosen for your target's surface and constraints.
• —Calibration data. Pipeline performance on a published SAR ladder for a comparator in your target class. We tell you up front where AI prediction is and isn't trustworthy for your target.
• —Wet-lab handoff package. RFP-ready, with assay panel, controls, and explicit kill-or-continue criteria.

Our therapeutic programs to date are computational. Wet-lab validation is a separate workstream.

Our offering is to bring the AI/ML and structural work to a level where wet-lab spend is well-targeted — not to replace the wet lab. We are best when paired with a partner that has wet-lab capability and a strong target hypothesis.

We do not promise hits. We promise calibrated rank-ordering, honest failure modes, and a wet-lab plan that earns its budget.

Therapeutics programs typically start on the Milestone or Co-development tier. Discovery programs are rare for therapeutics — scope rarely fits a fixed-fee deliverable. We can scope a Discovery-tier engagement when the target is well-defined and the modality choice is constrained.

Compare engagement models →

• —First engagement with an AI peptide discovery group. Start on Discovery. The deliverable is concrete and the conversation continues based on what comes back.
• —Validated mechanism, hit-to-lead acceleration. Milestone. Lower upfront, gated commitment, mutual escape clauses.
• —Differentiated asset, willing to share upside. Co-development. Shared funding, equity or royalty position, longer-arc relationship.

• —NDA in place before any substantive disclosure of your target or our methods.
• —Mutual escape clauses at every gate. Either party can exit cleanly.
• —Honest reporting at every gate, including negative results. We will tell you when to stop.
• —No engagement without a clear target hypothesis. We do not run programs against undefined targets.