Kaimei

Agent workspace for life science

Everything a question needs,
in one place.

A canvas for the pipeline, scientific tools to run in it, and an agent that does the work with you. All on one record, so no result loses the thread back to how it was made.

Request accessSee how it works

01 · Canvas

Your pipeline,
in plain sight.

Every step is a node: the target, its structure, each model run and what it produced, wired together. Pick a row in a result and the viewer and the charts follow it.

  • Boltz-2
  • ESMFold
  • OpenFold2
  • AlphaFold DB

02 · Tools

Heavy science,
made usable.

Structure prediction, docking, ADMET and public databases are set up and wired in, so you use them without installing software, managing GPUs or converting files.

03 · Agent

An agent that
works with you.

It asks what it needs, plans the steps with you, then runs the commands, reads the output and tries another way when one fails. Every command opens as a trace you can read.

The problem

From a decade-long funnel to one team's next compound.

A medicine is what is left after many thousands of candidates are narrowed down to one. The narrowing starts with questions that public data and scientific models can help answer.

  1. 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 nine in ten candidates that enter clinical trials are never approved.

    years from discovery to approval
    10-15+
    years from discovery to approval
    for each approved drug
    $1-2B+
    for each approved drug
    clinical candidates fail
    ~9 in 10
    clinical candidates fail
    Where data and models helpPreclinical and clinical
    TargetScreeningOptimisationPreclinicalPhase IPhase IIPhase IIIApproval
    >10,000~25010-20~6~4~21

    One medicine

    Screening
    >10,000
    Optimisation
    ~250
    Preclinical
    10-20
    Phase I
    ~6
    Phase II
    ~4
    Phase III
    ~2
    Approval
    1
    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.
  2. 02

    Before any trial,years of choosing.

    A team picks a target, screens more than 10,000 compounds and optimises about 250, before 10-20 go on to preclinical tests. Each of those first steps takes about a year and a half.

    1. 01 · Target

      Is this protein worth pursuing?

      Open TargetsUniProt
    2. 02 · Screening
      >10,000

      Which of thousands of compounds could bind it?

      ChEMBLRCSB PDBDocking
    3. 03 · Optimisation
      ~250

      Which analogues are worth making next?

      ADMETGenerative design
  3. 03

    A practical questionshould not need a computational team.

    Which compound on this list is worth testing next on this target, and why? In a large company, a computational group answers it. A team without one has the biology; the technical work is what stops it. Today it has four ways to get that work done.

    Each with its own format

    • UniProtFASTA
    • RCSB PDBmmCIF
    • AlphaFold DBPDB
    • ChEMBLTSV
    • DockingSDF
    • ADMETCSV
    KaimeiOne recordCanvas, chat and every command
    • Today

      Stitched together by hand

      Today: UniProt, PDB or AlphaFold DB, ChEMBL, RDKit, ADMET models and docking: five to ten sources and tools, each with its own format, install and code.

      WithKaimei

      Ask in plain language: the agent fetches the data, writes the code and runs it, with nothing to install. When you ask, it places the steps on your canvas, where you can change and rerun them by hand.

    • Today

      A workflow fixed in advance

      Today: A workflow tool needs every branch declared up front. In discovery the next step depends on what the last one returned.

      WithKaimei

      The plan you approve names the steps, not every branch. The agent writes each command after reading what the last one printed, so a failed command or an unexpected result changes what it does next.

    • Today

      Answers from memory

      Today: A general chatbot answers from memory and can make up its sources: in one peer-reviewed study, chatbots invented 28.6-91.4% of the references they gave for systematic reviews, depending on the model.

      WithKaimei

      The system, not the model, checks the ids and figures in a report's tables against your words and your session's output, and holds back a report with one it cannot find there.

    • Today

      A platform that picks for you

      Today: A commercial agent platform chooses the models and sets what each run costs you.

      WithKaimei

      You bring your own keys and choose the model for each session, and every session runs on a token budget with its meter in view. The choice and the spend stay with you.

    Chatbot references: Chelli M. et al., J Med Internet Res, 2024; models of 2023.

How it works

One question, from a library to leads.

An example session on EGFR, the lung-cancer target: screen a library against known actives, check the hits against what the tools printed, then optimise the leads on your canvas. The chat and the canvas share one record, so you can follow each step and change any of them by hand.

Ask in plain language, next to your canvas, with the model you pick. Before it plans, the agent asks for what it needs.

Step 1 of 9. The person asks for the best starting points against EGFR, screening 4,741 kinase compounds from ChEMBL against the known drugs. The agent asks which EGFR structure to work with, and 1M17, with erlotinib bound, is chosen.

An illustrative session. The target, the structure and the tools are real; the compounds and results are examples.

Why Kaimei

Quick to start, careful all the way.

The agent brings the evidence. You make the scientific call.

Seconds, not setup

There is nothing to install and no GPU to find. In our own runs a structure prediction came back in about two seconds, and an agent turn in six to eight. Hand the agent a question and it keeps working in the background around the clock, running and refining its experiments while you are away.

a structure prediction
~2 s
a structure prediction
an agent turn
6-8 s
an agent turn
in the background
24/7
in the background

Asks, plans, runs, checks

It asks you for what only you can give - a sequence, a SMILES, a structure id. It shows its plan before any work, runs one command at a time in a private workspace and reads what each printed before the next. Before a report reaches you, the ids and figures in its tables are checked against that output.

  1. Ask
  2. Plan
  3. Run
  4. Check

Chat and canvas, one record

Ask in chat and the agent builds the pipeline on your canvas: the nodes, their settings and the wiring. You check them and press Run. Pick a row in a results table and the 3D viewer and the plots follow it.

  1. Chat
  2. Canvas
  3. Run
  4. Results

15+ tools in one workspace

Structure prediction, docking, protein and molecule design, ADMET and the public databases, set up and wired in. Ten of them need no key of your own, and for the rest you bring yours: neither the agent nor its code ever sees it.

  • Boltz-2
  • ESMFold
  • OpenFold2
  • DiffDock
  • RFdiffusion
  • ProteinMPNN
  • MolMIM
  • ADMET-AI
  • RDKit
  • AlphaFold DB
  • RCSB PDB
  • UniProt
  • ChEMBL
  • PubChem

Timings measured on Kaimei's own runs: five Boltz-2 structure predictions on NVIDIA NIM took 1.5-3.7 s; agent turns took 6-8 s with a small model. Yours depend on the model and the input.

Kaimei means "to bring to light"

Bring one real question.

Kaimei is in early access for drug discovery teams without their own computational group: university labs, institutes, startups and companies. Request access with your organization and team, and each request is reviewed before an account opens.

Request accessSee how it works