Agent workspace for life science
A computational lab where
agents and biologists make more discoveries, faster.
An AI agent that works with the scientist through the four early stages of drug discovery. It reads the data, runs biology models on cloud GPUs and lays each step as a pipeline on one canvas, so every result traces to its source.
Kaimei's core
01 · Canvas & Nodes
Every experiment,
on one canvas.
A Kaimei canvas lays every experiment out as connected nodes, so you can follow inputs, model runs, structures, results and next steps without losing track.
- Boltz-2
- ESMFold
- OpenFold3
- OpenFold2
- AlphaFold DB
The problem
From a decade-long funnel to one team's next compound.
Most drug candidates fail late, in clinical trials, for lack of efficacy or for toxicity. Those failures are set early: in the first four stages, from target to preclinical, teams choose the target and the molecules on evidence still joined by hand. A sounder choice there can mean fewer failures later and safer medicines for patients.
- 01
Over 10,000 compounds go in.One approved drug comes out.
Developing one approved drug takes over 10-15 years and, on average, over 1-2 billion US dollars. About 90% of the candidates that enter clinical trials are never approved, most for lack of efficacy or for toxicity.
- years from discovery to approval
- 10-15+
- years from discovery to approval
- for each approved drug
- $1-2B+
- for each approved drug
- of drug candidates fail in clinical trials
- 90%
- of drug candidates fail in clinical trials
Where data and models helpPreclinical and clinicalOne medicine
Stage Compounds Cycle time % cost per NME Target - ~1.5 yr ~3% Screening >10,000 ~1.5 yr ~6% Optimization ~250 ~1.5 yr ~17% Preclinical 10-20 ~1 yr ~7% Phase I ~6 ~1.5 yr ~15% Phase II ~4 ~2.5 yr ~21% Phase III ~2 ~2.5 yr ~26% Approval 1 ~1.5 yr ~5% Figures: Sun D. et al., "Why 90% of clinical drug development fails and how to improve it?", Acta Pharmaceutica Sinica B, 2022. Stage values from its Figure 1, approximate. NME: new molecular entity, a drug whose active ingredient has never been approved before. - 02
Evidence for one target is scattered across many tools,and joined by hand.
One research question sends a scientist through many tools: databases such as UniProt, ChEMBL and PubChem, analysis and docking software, scripts and spreadsheets. Each tool keeps its own format, and often its own name for the same molecule, so the data is searched, copied, converted and matched by hand. The heavy models, such as structure prediction and docking, also need GPUs and bioinformaticians that small labs often lack.
Scattered across tools
UniProt FASTARCSB PDB mmCIFChEMBL TSVPubChem JSONDocking SDFADMET CSVScript PythonExcel by handScientistScientistScattered across tools
- UniProt FASTA
- RCSB PDB mmCIF
- ChEMBL TSV
- PubChem JSON
- Docking SDF
- ADMET CSV
- Script Python
- Excel by hand
- 03
Agentic AI is spreading in drug discovery,yet its risks get little attention.
Discovery data is a lab's intellectual property and part of its regulatory record. An agent with access can expose that data, or obey an instruction planted in a document. The agent can report a value no run produced, cite a trial that never ran or propose a compound no one can make, and send months of lab work the wrong way. The agent can also act unasked and leave no record of why, while a person still answers for every step.
What goes in
What can go wrong
Lab dataDocumentsModel runsMolecule designsTools and budgetYour requestData leakedPlanted instructionInvented valueUnmakeable compoundActed unaskedNo record of whyAI agentWhat goes in
- Lab data
- Documents
- Model runs
- Molecule designs
- Tools and budget
- Your request
AI agentWhat can go wrong
- Data leaked
- Planted instruction
- Invented value
- Unmakeable compound
- Acted unasked
- No record of why
How it works
One session, from the disease to the molecule.
An example session on EGFR in lung cancer. Its core is the screening run, from the target and its evidence to a ranked list of compounds with their measured activity, predicted heart and liver flags and a sourced report. A step the agent runs on one of the platform's tools is checked at the gates that tool's results ask for, and a separate review agent reads the finished report beside the records those tools wrote, so you can see what each result rests on. The same session then reads an approved antibody's interface, docks the hits the person takes onto the canvas, and reads the published work behind the one it puts first.
Early discoveryStart from the target: three agents of Kaimei's own take its entry, its structures and its papers at once.
Ask in plain language, next to your canvas. Before it plans, the agent asks for what it needs.
Step 1 of 13. The person asks what to know about EGFR in lung cancer before screening for it: the human entry, its experimental structures and what the recent papers say about resistance. The agent thinks it over, then asks which EGFR structure to work with, and 1M17, with erlotinib bound, is chosen.
An illustrative session. The target's evidence, the compounds and their measured activity, the published work, the structures, the interface score and the tools are real, read in September and November 2026; what the models predict for a compound is an example.
The stack
From your data to your canvas.
Six layers, one system. A scientist asks a question, approves the agent's plan, follows each step on the canvas, reads the sourced report and decides what goes to a lab test. From the ground up: your own secure space, public databases and your own files, biology and small-molecule models ready on cloud GPUs, the context that keeps the work about your project, the agentic AI that plans and runs the work by written methods, and your canvas on top, where every model step is a node you can read, change and rerun.
Why Kaimei
Quick to start, careful all the way.
The agent brings the evidence. You make the scientific call.
From data to results, one session
Papers, data, models on GPUs and the pipeline in one session, shared across the teams that work on it.
Agents explore, pipelines repeat
The analysis an agent ran becomes a workflow you reuse on other datasets and experiments.
Your lab's data, your control
A sandbox for each session, every step on record, and your data reaches an outside model only if your lab or you allow it.
Specialist agents
Agents for targets, proteins, chemistry and literature, each with its own specialist skills, so every part of the work is done by the one built for it.
Many tasks at once
Sub-agents run in parallel while the main agent stays ready for your next question and goal, so you wait less.
You approve, you check
You review each plan before the agents run it, then check and control what they deliver, so the work stays safe.
Kaimei means "to bring to light"
What will you bring to light?
Kaimei is in early access for biomedical scientists and drug discovery teams: university labs, institutes, startups and companies. Tell us your organization and team, and we review each request before opening an account.