Genome to candidate compound.
Full stack. Any kingdom.
OmniRayn Foundry is a hosted, end-to-end discovery platform. Feed it a genome; it returns a ranked set of biosynthetic candidates, predicted pathways, predicted bioactivities, and wet-lab-ready cascade reconstitution protocols — with quantitative null-control validation on every candidate.
Nine stages, one hosted pipeline.
A genome assembly goes in. A ranked set of two-null-validated candidate BGCs comes out — with 3D structures, predicted enzymatic functions, full retrosynthesis-checked pathways, druggability scoring, and wet-lab handoff packages. Point it at any sequenced organism; the engine does not change.
- 0Ingestion
Genome FASTA in; normalization, chromosome enumeration, organism classification, state machine initialized.
- 1Discovery (OmniRayn)
Profile-HMM scan → foundation-model embedding at full hidden dimensionality → Quantized Pattern Geometry novelty score. Proprietary.
- 2Gene prediction
Augustus / GeneMark-ES / GlimmerHMM to resolve ORFs per candidate BGC.
- 3Structure (dual-model)
AlphaFold 2 vanilla + AlphaFold 2 abliterated + Boltz-1 co-folding. Disagreement is the signal.
- 4Function (dual-model)
CLEAN vanilla + CLEAN abliterated + ESP substrate prediction + CatPred k_cat. Row-2 disagreement flags novel catalytic mechanisms.
- 5Cascade assembly
Reaction DAG with MIBiG priors, cross-checked via retrosynthesis. Emits predicted final SMILES + coherence score.
- 6Compound novelty
NPAtlas / COCONUT / PubChem / ChEMBL nearest-neighbor in multiple chemical spaces.
- 7Bioactivity + druggability
Boltz-2 docking against disease-target panel + ADMET + QED + Lipinski + PAINS filters.
- 8Variant library
ProteinMPNN / RFdiffusion generate ~100 SAR-optimized enzyme variants. Filtered by AF2-abliterated and Boltz-1.
Six differences from every other BGC pipeline you've evaluated.
Dual-null validation — not "ML confidence".
Every candidate clears a compositional null and an in-distribution intergenic null. Published ablation shows 9 of 10 top-ranked shiitake candidates fail the intergenic null. We only ship the ones that pass.
Dual-model abliteration — disagreement is the signal.
Every foundation model runs vanilla and abliterated. Row-2 cases (vanilla-NO, abliterated-YES) are the novelty zone — and the competitive moat. Competitors using vanilla models alone dismiss exactly these candidates.
Full-stack, not a point tool.
Genome → structure → function → pathway → novelty → druggability → variant library → wet-lab handoff. Most competitors cover two or three of these and hand off the rest.
Organism-agnostic by design.
The engine does not know what kingdom it is operating on. A bacterial or plant or marine genome requires a configuration change, not a re-engineering sprint.
Wet-lab-ready output.
Codon-optimized sequences (paid tiers), cascade protocol with reaction order and conditions, partner-specific handoff packages for Ginkgo / Antheia / Protein Evolution / Arzeda.
Cross-domain provenance.
The same pattern-geometry engine was validated in financial signal and cryptographic R&D before the first genome ran. Robustness is a property of the engine, not a claim.
Two two-null-validated candidates are already on the workbench.
Both passed a dinucleotide-preserving compositional null and an in-distribution intergenic null at p ≥ 0.90 — the bar we ask every candidate to clear before it is dispatched to a wet-lab partner. Specific compound identity and therapeutic activity are to be determined by heterologous expression and characterization.
Cluster A
TWO-NULL VALIDATEDMycophenolic-acid pathway signature explicitly excluded. One flanking cytochrome P450 confirmed.
Cluster B
TWO-NULL VALIDATEDNon-overlapping with cordycepin (Cns) cluster. Strongest compositional-null signal in the 7-organism embodiment.
The engine is organism-agnostic. Every kingdom is a configuration change.
Today's marketing focus is medicinal Basidiomycete fungi. The engine does not change when we move to actinomycetes, medicinal plants, marine sponges, or extremophile archaea — only the Pfam panel, gene-prediction training model, and reference database swap.
| Kingdom | Primary customer | Launch |
|---|---|---|
| Medicinal Basidiomycete fungi | Nutraceutical & pharma NP divisions | Q2 2026 (current) |
| Actinomycetes | Big pharma antibiotic discovery | Q3 2026 |
| Medicinal plants | Phytopharma, traditional-medicine modernization | Q4 2026 |
| Marine organisms | Marine-derived drug startups | Q1 2027 |
| Bacterial pathogens (counter-discovery) | Biodefense, vaccine discovery | Q2 2027 |
| Extremophiles | Industrial-enzyme markets | Q3 2027 |
| Insect venoms / defensins | Peptide therapeutic startups | Q4 2027 |
Production-grade SDKs. OpenAPI. Free tier.
pip install omnirayn, npm install @omnirayn/client. Both are open-source (MIT); the pipeline they talk to is a hosted proprietary product. Full REST API with async job model, webhooks, and per-tier rate limits.
from omnirayn import Foundry
foundry = Foundry(api_key="or_live_...")
job = foundry.jobs.create(
genome_fasta="lentinula_edodes.fasta",
organism_hint="fungal",
stages=["discovery", "structure", "function", "cascade"],
dual_model=True,
)
results = job.wait_for_completion(timeout_minutes=120)
for c in results.candidates:
print(c.bgc_id, c.coordinates)
print(" QPG novelty :", c.discovery.qpg_novelty)
print(" p_comp :", c.discovery.p_compositional)
print(" p_intergenic :", c.discovery.p_intergenic)
if c.function.novelty_flags:
print(" → Row-2 novelty candidate")
foundry.wetlab.export_protocol(
job_id=job.id,
candidate_id=results.candidates[0].bgc_id,
partner="ginkgo",
output_path="cluster_A_wetlab.zip",
)The Foundry Exchange — candidate intelligence, by the listing.
Beyond bespoke pilots, the engine's output is packaged as a curated catalog of mechanism-anchored, tox-filtered, trial-whitespace-mapped candidates. Every listing carries a validation-state badge — Hypothesis → In-vitro → Lead — so the trust line is never blurred. Novel-composition structures stay NDA-gated. All listings are computational and pending wet-lab; no disease or efficacy claims.
Teaching packages
Free, public-safe L0–L3 academic packages — entity card, evidence, learning exercises, and a BD case study. No login, no paywall. “Not an efficacy grade.”
Single dossier → mechanism bundle
Commercial-use license, pay per dossier — a single candidate dossier ($2,450) or a full mechanism bundle ($3,950, its ranked candidate list). No subscription. Published-literature corroboration on every listing.
Novel-composition leads
Exclusive or non-exclusive rights to engine-generated novel-composition leads. Structures disclosed only under NDA; Foundry retains title. Enterprise scouting is contact-sales.
Book a pilot.
Pilot tier is $50k and runs against one organism genome end-to-end. Program tier ($250k–$1M) enables the Stage 8 variant library. Partnership tier is bespoke. All tiers include a direct channel to the core science team.