Building the AI Scientist

Work loops

Built from the work loop up, not the model down

Scientific work is long-tail and context-dependent. A generic model answers common questions but struggles when the work depends on the provenance of sparse evidence, the context in which data was generated, and the assumptions behind an assay, model system or endpoint. The durable capability is built upward from validated vertical loops.

Discovery researchAvailable now

Omics data to target prioritization

From a public psoriasis RNA-seq dataset to five reviewable target hypotheses and a prioritization memo.

Walk through 10 stages

Dataset: GSE54456
Stages: 10
Checkpoints: 5
Branch points: 21
Output: Target-prioritization memo with reproducibility pack

Loops in the backlog

Each will bring its own methods, evidence standards, rubric and artifact, while reusing the generic capability underneath.

Overlapping glass microscope slides, one lifted clear and marked with a single violet dot.

planned

Translational

Dataset to biomarker hypothesis

From cohort data to a candidate biomarker with a validation design.

Scientific explanation leading to hypothesis generation

Torn paper fragments drawn along threads into one tight central cluster, two left stranded outside.

planned

Discovery research

Evidence to mechanism

From scattered published evidence to a defensible mechanistic account.

Scientific explanation

A uniform grid of pale squares with a single darker outlier, a red thread running to it.

planned

Safety

Safety-signal investigation

From a detected signal to a qualified assessment with required action.

Scientific explanation

Concentric rings of pale discs narrowing to one steel disc with a blue inset at the center.

planned

Discovery research

Disease biology to target hypothesis

From disease understanding to target hypotheses without a starting dataset.

Scientific hypothesis generation

A row of upright paper strips of differing heights, one taller and tipped with an amber marker.

planned

Strategy

Competitive pipeline assessment

From the competitive landscape to a differentiated scientific position.

Scientific explanation and scenario analysis

A paper ribbon forking into two diverging paths with a steel marker resting at the fork.

planned

Discovery research

Experimental result to next-best experiment

From a result to the experiment whose outcome would be most informative.

Scientific hypothesis generation

Why loops, and not one general-purpose scientific agent

Each loop is specialized, but most of what it needs is not. Separating the four kinds of capability is what lets validated loops accumulate into a general AI scientist instead of remaining isolated demonstrations.

Generic AI scientist capability: Reusable infrastructure. Built once, used by every work loop — runtime, memory, provenance, orchestration.
Scientific skill: A validated scientific method, reusable across related scientific problems. Versioned and testable, like a library function for science.
Loop-specific configuration: Unique to one work loop. The recipe, thresholds and rubric that make omics-to-target different from safety-signal investigation.
Organization-specific context: Your therapeutic strategy, internal datasets, platform capabilities, prior failures and scientific preferences. Cannot be bought off the shelf.