TensorTarget · An AI co-scientist for drug-target discoveryIn development

Drug-target hypotheses from genetics and omics data, with an AI co-scientist and human review

TensorTarget is being built to turn human genetics, single-cell, and other omics data into a ranked set of testable drug-target hypotheses, each grounded in evidence and reviewable by a human scientist.

It is not finished, and it makes no claim of validated or autonomous results. I'm looking for a few design partners to help shape it.

Become a design partner →

What it is

Most target lists come from a single lens, such as a genetics score or a pile of differentially-expressed genes, with the other layers stitched together by hand, one gene at a time. No one can weigh genetics, expression, networks, druggability, and literature across thousands of genes by hand.

The approach starts from genes with human genetic causal evidence, maps each to the cell type where its biology is dysregulated, traces signalling to effectors through protein networks, and layers in druggability. An AI co-scientist then generates, critiques, and refines hypotheses, citing the evidence for every claim, and human scientists review and curate each one.

The aim is to apply the reasoning a target-ID expert would use across every layer and every gene at once.

TensorTarget architecture for IBD drug-target discovery: collector agents gather evidence in parallel (genetics, single-cell, network, druggability), a co-scientist reasoning loop of generation, critique, and evolution overseen by a supervisor, producing ranked testable target hypotheses such as NOD2 to CASP1 in BEST4+ enterocytes.
Collector agents gather evidence in parallel; a co-scientist loop then proposes candidate target hypotheses (each a target, its mechanism, and the cell type it acts in), stress-tests them against the evidence, and refines the survivors into a ranked, testable set. IBD is shown here as a worked example; the same approach applies to other indications.

What it is designed to produce

  • A ranked set of candidate target hypotheses for your indication
  • For each hypothesis: the cell type it acts in, the mechanism, the genetic + expression + network evidence, known drugs & druggability, and a suggested validation experiment
  • Full provenance and caveats, including what was considered and why it was set aside
  • A walkthrough call with the analysis

Design principles

Genetics-first

causal human-genetic evidence, not just correlation or expression.

Cell-type-resolved

where a target acts, not only whether it matters.

Grounded reasoning

citations + adversarial self-critique, not a black-box score.

AI co-scientist + human review

human scientists review and curate every hypothesis the AI produces.

You’re in the loop, too

Your team’s data and judgment shape the analysis. Bring your own data and your own hypotheses into the loop.

Bring your own data

Add your internal omics, genetics, or clinical datasets and I run the analysis over them alongside the public evidence.

Seed your own hypotheses

Have a target or mechanism in mind? Hand it to the co-scientist as a starting hypothesis. It gets generated, critiqued, and refined with the rest, and you see the evidence for and against.

Interested in shaping it?

TensorTarget is in development. Be one of the first design partners: bring a real indication and help shape it.

Become a design partner