Introducing Gnosis: an integrated AI system for small-molecule drug design

Most drug-discovery software runs like a relay race. Potency goes first, then selectivity, toxicity, metabolism, and off-target liabilities get checked later, one handoff at a time. It is an arrangement almost engineered to fail late: a compound series can look brilliant for months and then fall over on a property that was knowable on day one. Today we are introducing Gnosis, Vecentra’s integrated AI system for small-molecule drug design, built to surface those trade-offs at the first molecule instead of the last.

One decision system, not a drawer full of tools

Gnosis couples molecular generation and evaluation inside a single architecture. Five predictive modules score a candidate on the axes that actually decide whether it becomes a drug, and a generative engine, Gnosis Forge, proposes molecules inside their combined constraints rather than perfecting one property and praying the rest survive. The output is candidate prioritisation against many requirements at once, which is roughly how a seasoned medicinal chemist reasons, only faster and without the coffee.

The modules

• Gnosis I, target engagement: does the molecule bind the target, and how strongly?
• Gnosis II, disease-cell selectivity: does it hit diseased cells while leaving healthy ones alone?
• Gnosis III, safety and toxicity: which liabilities (hERG, CYP interactions) deserve a red flag?
• Gnosis IV, downstream gene-expression response: what does the cell actually do afterwards?
• Gnosis V, ADMET: absorption, distribution, metabolism, and clearance, or will it behave like a drug once it meets a body?

Two more layers wrap the modules. Gnosis Forge generates and optimises molecules within those combined constraints, and Gnosis Lens supplies structural interpretability, surfacing the molecular features behind each prediction. A calibration layer sits over the top and flags anything low-confidence or outside the model’s applicability domain, which is our polite way of saying "do not trust this one yet."

Gnosis I, and a bet on the undruggable

Target engagement in Gnosis runs through more than one binding module. Gnosis I Fused is an adaptive ensemble for protein-ligand affinity ranking. Gnosis I RNA extends binding prediction to RNA, and that is a deliberate bet: RNA-targeting small molecules are one of the most promising routes to the large slice of biology that classical protein-binding drugs simply cannot touch. Reaching historically undruggable targets is central to our mission, and it happens to be the kind of problem that punishes sequential design and rewards integration. RNA-targeting and induced-proximity molecules (PROTAC, RIBOTAC, and AUTAC-style chimeras) have to satisfy multiple binding interactions, linker geometry, ternary-complex formation, cellular exposure, and pharmacokinetics all at once. Their design is multidimensional by nature, which is exactly the job Gnosis was built for.

On performance, Gnosis I Fused returns binding-affinity rankings in seconds per ligand. That matters because the molecular-dynamics free-energy methods it competes with, such as FEP+, can take hours to days per calculation. Across 16 public benchmark series (590 ligands), Fused reached a mean Spearman correlation of 0.75. Because reliability varies by target, Fused weighs its complementary signals using known actives. On a blinded WRN test that adaptive weighting dragged ranking correlation from an actively unhelpful −0.17 for the underlying co-folding model up to +0.60. We will also say the quiet part out loud: adaptive selection currently needs at least four known actives, and absolute-potency estimates require a separate target-specific calibration gate.

Safety and developability, in the loop from the start

Since failing late is the expensive way to fail, Gnosis III and Gnosis V pull safety and ADMET forward. In held-out evaluation, Gnosis reached AUROC 0.847 for hERG risk and 0.826 for CYP3A4 inhibition, with applicability-domain gating on shakier endpoints like P-gp so the system abstains instead of bluffing. Gnosis II and IV add the biology that potency quietly ignores: whether a molecule is selective for diseased cells, and what transcriptional response it actually triggers.

Interpretability, and knowing when to keep your hands off

A prediction is only useful if a scientist knows when to act on it. Gnosis Lens points to the molecular features driving each score, so chemists and biologists can interrogate why a compound was prioritised or flagged rather than accepting a number on faith. The calibration layer marks where predictions should not be trusted without experimental evidence. None of this is meant to replace experimental science. It is meant to make computational predictions transparent, testable, and useful at the exact moment you decide what to synthesise.

Built for the scientist making the next decision

Gnosis grew out of a very practical frustration at the border between computational modelling and experimental pharmacology: predictions are too often produced in splendid isolation from the constraints that decide whether a molecule survives. So we built the system around the person who has to choose which compounds justify synthesis, screening, and validation, and we hand back its output as Candidate Intelligence Codex reports: prioritised structures, documented scientific rationale, and concrete experimental recommendations.

We are running Gnosis cancer-first, where selectivity, safety, and pathway effects can be learned together, with the same intelligence transferring next into neurodegenerative disease. Early programs with academic research partners span RNA-targeted approaches in ALS and frontotemporal dementia, synthetic-lethal targets in glioblastoma, LRRK2 inhibition for Parkinson’s, and TRPM2 in oxidative-stress-driven neuronal injury.

Generative chemistry can spin up molecules in seconds. That was never the bottleneck. The hard part is optimising across competing biological and pharmacological requirements from the outset, and that is what Gnosis is built to do.

Dr. Mahan Azad, Co-founder & Chief Innovation Officer

This is the first of several deeper looks at how Gnosis works. In upcoming notes we will climb inside disease-cell selectivity and the downstream-omics modules, and share more of the validation roadmap as experimental results land.