Trust documentationDisclosed before asked

Dependency licensing

PetriDish runs on a curated stack. Every dependency below has its commercial-use status verified and recorded. Non-commercial models are firewalled to the research lane — they do not appear in any customer-facing endpoint. Pending licenses (most importantly the IndiGen commercial conversation with CSIR-IGIB) are disclosed openly rather than buried.

Source: Build & Launch Reference §3 (Licensing Decision Matrix) + §7 (Indian Data Layer).
DISCLOSED — INDIGEN COMMERCIAL LICENSING IN PROGRESS
The IndiGen 1,029-genome panel (CSIR-IGIB) is the foundation of India-specific allele frequencies in the platform. Research-tier use is free. Commercial licensing with CSIR-IGIB Business Development is in active conversation and will be resolved before paid product launch. We disclose this because — per Build & Launch Reference §7 — discovered late it reads as a gap; disclosed early it reads as rigor.
Production · Commercial OKPending — CSIR-IGIBVerify before usePaid license requiredResearch lane only

Production · Commercial OK

15
AssetCategoryPurposeLicenseIn shipped product
Llama 3.3 70B (via Groq)
Used under threshold for commercial deployment
Reasoning LLMCypher generation + grounded synthesisLlama Community License
Neo4j AuraDB
Knowledge graph4.4M+ edges India-weightedCommercial
PrimeKG (Harvard)
Knowledge graphPrecision-medicine KG seedMIT
OpenTargets
Knowledge graphTarget-disease evidence + PharmGKB widgetOpen / permissive
ESM-2 (Meta)
Protein LMProtein variant function (planned integration)MIT
BioMedLM (Stanford)
Clinical LMBiomedical reasoning (planned)Apache 2.0
IGVDB
Variant DBIndian variant frequency contextOpen
FAERS
Drug safetyAdverse event reporting referencePublic domain (FDA)
IMPPAT 2.0 (IIT-M curated subset)
100 of 17,967 compounds ingested; ACTREC expansion in progress
PhytochemistryAyurvedic compound knowledge layerAcademic / CC-BY (curated subset)
MHCflurry 2.0
Planned production integration. 91% of alleles ≥0.90 AUC. Pan-allele works on Indian HLA without Indian-specific training data.
Peptide-MHC presentation (Class I)Open-source pan-allele HLA Class I binding prediction for Protein AtlasApache 2.0 — open
BepiPred / DiscoTope
Planned integration for Protein Atlas Layer 3 (autoantigens) when real-time inference layer ships.
B-cell epitope predictionLinear / conformational B-cell epitope candidates for autoantigen layerOpen
ACMG/AMP Variant Interpretation Standards
Public clinical-genomics standard. IndoVar applies as the decision-support framework.
Classification frameworkACMG-2015 BS1/BA1/PM2/PP3/BP4 rule logic for IndoVar reclassificationPublic standard
gnomAD v4
The Western baseline IndoVar improves on. Critical reference for BS1 reclassification logic.
Western baseline referenceReference allele frequencies that IndoVar compares againstOpen · MIT
ClinVar (NCBI)
ClinVar is the existing baseline IndoVar concords with or reclassifies when Indian frequency justifies.
Variant databaseExisting clinical-significance calls that IndoVar concords with or reclassifiesOpen · public domain (NIH)
Evo 2 (Arc Institute + NVIDIA)
Planned integration as the production variant-effect model (§5).
DNA foundation modelVariant-effect backbone (planned production)Open · commercially usable

Pending — CSIR-IGIB

1
AssetCategoryPurposeLicenseIn shipped product
IndiGen (CSIR-IGIB)
Currently used under research framing. Commercial licensing conversation with CSIR-IGIB Business Development required before paid product launch (Build & Launch Ref §7).
Population genomics1,029 Indian whole genomes; India allele frequenciesResearch free · Commercial license required

Verify before use

6
AssetCategoryPurposeLicenseIn shipped product
AlphaMissense (DeepMind)
Not currently in shipped product; verify before adoption
Variant pathogenicityMissense pathogenicity scoresWeight/artifact license requires verification
PharmGKB redistribution
Annotations referenced only; not bulk-redistributed
PGx annotationsVariant-drug-response referencesFree access · commercial redistribution requires verification
ESM-3 / ESM-C (EvolutionaryScale)
Default to ESM-2 unless commercial terms confirmed
Protein LMNext-gen protein functionTiered — commercial terms vary
SIDER drug-side-effect DB
Drug safetySide-effect knowledgeCommercial terms require verification
NetMHCpan-4.1 / NetMHCIIpan-4.0
Class II auROC ~0.91. Default to MHCflurry + ESMCBA (open) for production unless DTU commercial terms confirmed.
Peptide-MHC presentation (Class I + II)Field-standard pan-allele binding prediction (DTU)Academic license — commercial terms require verification
ESMCBA (2025)
Peptide-MHC presentationESM-C-based Class I prediction — new SOTA direction (median Spearman 0.62 vs NetMHCpan 0.56)Academic — verify

Paid license required

1
AssetCategoryPurposeLicenseIn shipped product
DrugBank (full)
Open subset only is shipped; full DrugBank requires paid commercial license at scale
Drug compendiumComprehensive drug dataPaid commercial license required

Research lane only

3
AssetCategoryPurposeLicenseIn shipped product
AlphaFold3 (DeepMind)
Per autoimmune_protein_deep_research.md §9: best structural tool but firewalled to research lane. Used for paper-grade structural confirmation; never inside any customer-facing endpoint.
Structure / TCR-pMHCBest HLA-peptide-TCR structural model — research papers and validation onlyNon-commercial server + weights
MsgaBpred / AlphaEpi
Inherits AF3 non-commercial restriction. Research lane only.
B-cell epitope (research)AF3-based epitope prediction for research-paper validationResearch (uses AF3)
AlphaGenome (DeepMind)
Firewalled to validation paper analysis. Will not appear in any customer-facing endpoint.
Regulatory variantsRegulatory variant analysis (research papers)Non-commercial API only
Per Build & Launch Reference §3: every dependency's commercial-use status is verified at the moment of adoption and recorded in this matrix. Non-commercial models (AlphaGenome) are firewalled to the research lane. Paid-license dependencies (DrugBank full) are not shipped until the commercial agreement is in place.