Coganu AI

COGANU AI DESKTOP · RESEARCH USE ONLY

Candidate rankings on your lab's PC

From efficacy evidence to clinical-entry review: installable desktop software that weighs each candidate's efficacy evidence (target, binding, cell response) together with its developability (ADMET, safety, public clinical information) to estimate candidate rankings for the next step. The computation finishes on the lab's own PC, and customer research data is not sent off it.

See plans and pricing Try it in the browser first What "stays on your PC" means
The Efficacy Axis (TRG) screen of Coganu AI desktop. Running PCSK9 against hypercholesterolemia returns TRG-1 genetics HIGH (causal 0.954), TRG-2 mechanism toxicity LOW (risk 0) and a calibrated PoS of favorable (0.1, preclinical), alongside a notice that this is not a definitive success prediction.
One target, run in the installed software. Each badge carries the number behind it, and the line underneath says what the number is not. A real run on PCSK9 and hypercholesterolemia. We do not invent numbers — only what was computed is shown, and the result itself states that it is not a clinical-success prediction. Tap the screen to open it at full size — zoom in on a small display.
Research Use Only installable software for researchers. Not a medical device; we do not sell or ship drugs, chemicals or any physical goods. Rankings are estimates for researchers to review and do not predict clinical success. Not medical advice, and not for diagnosis, treatment decisions or patient care.

We publish the numbers

Measured by us on public benchmarks. What each number measured, and its limits, are stated in an evidence file that ships with the software.

+16net cases gained by pose selectionPoseBusters 427 · 30 improved / 14 degraded · p=0.023
1.04 Åmedian pose errorAstex 46 · full product path (was 1.39 Å)
21/28"high" confidence badges correct"low" badges: 8 of 32 — on a check set not used for tuning
{
  "schema_version": "coganu.pose-rescorer-evidence.v1",
  "weights_sha256": "9325869f7ecefd55…a798c3e0aca973dc67",
  "benchmark": { "n": 427, "vina_top1": 202, "rank_sum": 218,
                 "wins": 30, "losses": 14, "mcnemar_p": 0.023 },
  "astex_46": { "vina_top1": 27, "rank_sum": 31,
                "measures": "re-ranking of frozen Vina poses only (no re-docking)" }
}
Extract from genscore_clean_v1.evidence.json, shipped inside the software. It records the weight hash and what the number measured. The 1.04 Å above is the full product path (SMILES → docking → pose selection); the 27 → 31 in the file is a separate comparison that re-ranks frozen poses without re-docking.

It follows the order research actually takes

Not a list of features. Four questions that narrow a candidate set, each with the tool that answers it, in one workspace.

  1. Is the target worth it?

    No GPU

    Summarises target evidence from public genetics and cell-line dependency data, and ranks which targets to look at first.

  2. Does it bind?

    Desktop

    Runs structure prediction and docking, then selects which generated pose to trust, with a confidence badge.

  3. Does the cell respond?

    Desktop

    Estimates cellular response from expression signatures and ranks which cell lines to assay first.

  4. Can it be made, and is it safe?

    Desktop

    Collects ADMET predictions with their applicability domain, plus public clinical and safety data, into material for the go/no-go call.

All four steps run in the same workspace of the installed software. Step 1 needs no GPU. You can also try step 1 first in the web app.

What it does, concretely

Target and efficacy evidence

Summarises target evidence and cell-line dependency from public genetics and cell-line data.

ADMET and lead optimisation

Shows absorption, metabolism and toxicity predictions together with their applicability domain.

Structure and docking

Protein–ligand pose generation, docking and computational-chemistry helpers.

Repurposing and combinations

Ranks repurposing and combination candidates from expression signatures and public data.

Clinical and safety information

Collects public clinical-trial and safety data into material for risk review.

ADC and cancer-vaccine design aids

Helper tools for organising linker, payload and epitope candidates.

All outputs are reference material for researchers to review. Uncertainty and applicability domain are shown with each prediction; accuracy is not guaranteed.

Your research data does not leave your PC

Structures, sequences, assay values, queries and results are all processed on the customer's PC. Nothing is uploaded to us for computation and sent back.

Computation finishes locally
The models are installed on the PC too. No request carries research data.
Account traffic carries no research data
Only sign-in, licence checks and updates. A PC with no internet access is activated with a signed licence file.
Public-registry lookups are the exception, and we say so
Features that query public registries — clinical trials, adverse events, cancer genomics — send the name you typed to that organisation. The default is not to send; each lookup has to be allowed separately. The software lists exactly what can leave and where it goes.
We do not train on customer data
Our models are trained on public data. Nothing a customer enters is reused.

This is not a claim of perfect security. It means an advantage where computation must not use the network.

Pricing

One seat = one registered PC. Because computation runs on your PC, nothing is charged automatically by usage — you add more only when you choose to.

Academic research · 1 PC

$30–$100 / month

For researchers at universities and public or non-profit research institutions · Basic $30 · Plus $100 · monthly card subscription

Researchers and companies · 1 PC

$200–$800 / month

Commercial use included · Pro $200 · Pro+ $400 · Max $800 · monthly card subscription

Annual institutional contract · lab (3 PCs)

Company $27,600/year · non-profit $14,000/year (invoice, PO, bank transfer)

In Korea: company KRW 40m/year (excl. VAT) · non-profit KRW 20m/year (incl. VAT)