A copilot for drug discovery —
not a black-box oracle.
Twelve specialized agents that hypothesize, design, dock, and critique — orchestrated by a shared reasoning graph that human scientists can inspect, veto, and steer at every step of the process.
Human-in-the-loop at every gate.
Most AI-drug-discovery platforms hand you a scored candidate list and ask you to trust it. drug.AI hands your medicinal chemists a research partner they can inspect, question, and steer — because in rare oncology, "trust me" is not an acceptable answer.
Wet-lab validated.
- Every candidate carries a full provenance chain — every score, every rejection, every human override preserved and queryable.
- Scientists inspect the reasoning graph and veto agents in real time. Overrides become training signal for the next cycle.
- Agents argue with each other. When they disagree, the debate is surfaced to a human — not resolved silently.
- Efficacy is measured against wet-lab ground truth on every program. Predictions never ship un-validated.
- Compliance-by-design: regulatory readiness baked into every model, dataset, and experimental record from day one.
- Modular agent architecture — when something fails, you know exactly which agent, which decision, and why.
Unclear efficacy.
- Ranked candidate list handed off — no way to inspect why a molecule was chosen or why others were dropped.
- No mechanism to override or steer without retraining the whole model. Chemist intuition is discarded.
- One monolithic model. When it's wrong, you can't tell which sub-component failed or how to fix it.
- Retrospective benchmarks over prospective validation — the paper looks great; the clinic often doesn't.
- Regulatory trail is an afterthought — bolted on before IND, not designed in from day zero.
- Trust the black box. Or don't. Either way, you're on your own.
Every prediction is a proposal. Every proposal has a debate. Every decision has a human sign-off. That's the difference.
The agents.
Each agent owns a stage of the discovery loop. They communicate through a shared knowledge graph — passing structured hypotheses, not natural language, so nothing gets lost in translation and every step is inspectable.
+ 4 more specialist agents (Assay Designer, Formulation Chemist, Patent Analyst, Clinical Translator).
Watch a discovery cycle unfold.
A real run trace from cycle #4472 — target nomination through candidate promotion, in ~17 elapsed days compressed here to a few seconds.
Atoms placed by diffusion,
not filtered from libraries.
Our composer grows molecules atom-by-atom inside the target pocket — steered by binding potential, ADMET priors, and synthesizability constraints all at once. This isn't virtual screening. It's de novo design.
Drag the molecule to rotate. Hover any atom to inspect.
Agents that argue,
not just answer.
Every decision traces back through a graph of competing hypotheses. Human scientists inspect, veto, and inject constraints — the system learns from every intervention.
This isn't a black-box oracle. It's a transparent partner that shows its work at every step.