Why AI Drug Discovery Models Need Non-Public Genomic Training Data

The core problem with public genomic databases for AI training

The world's major genomic databases — GenBank, Ensembl, UniProt, the NCBI — contain billions of sequence records. They are extraordinary scientific achievements. And for training AI drug discovery models, they have a structural limitation that is becoming the field's primary bottleneck.

Public databases are not a representative sample of biological diversity. They are a heavily biased sample skewed toward model organisms (humans, mice, Arabidopsis, yeast, E. coli), commercially prioritised crops (maize, rice, soybean, wheat), and organisms easy to culture in laboratory conditions. Vast swathes of the evolutionary tree — including most of the planet's 400,000+ plant species — are represented by partial sequences, low-quality assemblies, or nothing at all.

An AI model trained on this data learns the chemical and genetic patterns that occur in the organisms that have been prioritised for study. It extrapolates poorly to the chemical spaces occupied by unstudied organisms, because those spaces are not in its training data.

What AI drug discovery platforms actually need

The leading AI drug discovery companies — Recursion Pharmaceuticals, Isomorphic Labs, BenevolentAI, Insilico Medicine, Schrödinger — are not primarily limited by their model architectures. The field has capable graph neural networks, transformer models, and generative molecular AI. What they are limited by is the quality and diversity of the training data those architectures are fed.

What high-performing AI drug discovery models specifically require from their training data:

• Chemical novelty: Compounds and scaffolds not already in ChEMBL, PubChem, or other public compound databases. Novel chemistry forces the model to explore unfamiliar regions of molecular space — which is where the next generation of therapeutics likely sits.
• Linked data across every layer, tied to a single specimen: Data where genomic, transcriptomic, and metabolomic layers are linked at individual specimen level, so the model can learn the causal relationships between genes and compounds — not just correlations across populations.
• Annotated biosynthetic context: Not just "here is a compound" but "here is the biosynthetic pathway that produces it, here are the genes involved, here is the tissue and environmental context." That is the structure that enables a model to learn mechanistic chemistry.
• Non-public exclusivity: Training on the same public data as every competitor produces models with similar capabilities and similar blind spots. Non-public, proprietary training data is a genuine competitive differentiator.

The Madagascar proposition for AI platforms

Madagascar's endemic flora represents a training data opportunity that is structurally unlike anything currently available at scale in public databases:

• 165 million years of evolutionary isolation producing chemistry with minimal overlap to anything in ChEMBL, PubChem, or any published natural product database.
• 12,000+ endemic species — a training corpus of biological diversity that dwarfs the characterised flora of any other single region.
• 8-layer specimen-level multi-omics (genome, transcriptome, NIR, LC-MS/MS, phenomics, soil chemistry, germplasm, provenance metadata) providing the linked data across every layer, tied to a single specimen that AI models require but cannot assemble from fragmented public sources.
• Non-public, non-replicable data — this dataset cannot be assembled by a competitor scraping databases, because the databases do not contain it.
• FAIR-compliant, blockchain-verified provenance — every record is legally defensible for commercial use, with full Nagoya Protocol compliance documentation.

What the market is signalling

The acquisition landscape in 2025–2026 is providing clear signals about where the AI life sciences market is heading. Anthropic's acquisition of Coefficient Bio — a biological AI company building models for drug discovery — at approximately $400 million, just eight months after its founding, signals that the major AI labs are moving aggressively into biological data territory. Google DeepMind's AlphaFold programme demonstrated that structural biology data at scale could produce models of considerable commercial value. The same logic is now being applied to genomics and metabolomics.

Pharmaceutical companies including Sanofi, Bristol-Myers Squibb, Takeda, and AbbVie have formed or joined AI training data consortia specifically to access novel biological data that public databases do not contain. The market is explicitly acknowledging that the public data ceiling has been reached — the next phase of AI drug discovery requires novel, proprietary, high-quality data at scale.

The data platform versus the model: a distinction that matters

IsoGentiX is not building an AI drug discovery model. It is building the training data infrastructure that AI drug discovery models need. This is a deliberate and important strategic distinction.

The AI model layer is capital-intensive, rapidly evolving, and inhabited by well-funded competitors with large engineering teams. The biological data layer — generating genuinely novel, Nagoya-compliant, specimen-level multi-omics from an undercharacterised biodiversity hotspot — requires a specific combination of field capability, genomics infrastructure, regulatory access, and community relationships that cannot be replicated quickly, cheaply, or from a desk in San Francisco.

The scarcity and non-replicability of the data asset is the moat. The AI platforms that want to access the chemical and genetic space of Madagascar's endemic flora have exactly one compliant route to that data. That structure — one compliant provider of a non-substitutable, non-public data asset — is the commercial architecture that underlies IsoGentiX's licensing model.