Analyses an agent can call
Deconvolution, pathway enrichment and TF activity inference are each a separate function with a defined output. An agent calls the same function a scientist would and gets the same result.
The AI Harness for Translational Science
Evidence that keeps pace with discovery.
Modern infrastructure for translational teams. Evidence gets mapped to your program goals as the work happens, not reconstructed after the fact.
Runs inside your infrastructure. Your data never leaves it.
Live coverageSTING · solid tumourStage 3 · Preclinical lead — 30/42 criteria
The framework
42 criteria, four dimensions
Each wedge is one criterion. The four outlined wedges are open as examples: hover one to see what it measures and the evidence the program holds for it.
Biology8Target8Model14Mechanism12
Running on real programs at
The problem
Still doing cutting-edge science on old-school infrastructure?
Most translational workflows haven't changed since spreadsheets and slide decks became the norm, so evidence gets reconstructed under deadline instead of captured as it's generated.
Old-school infrastructure
- —Evidence scattered across drives, decks and notebooks
- —Every program inventing its own bar for translatable
- —A model chosen for convenience, defended from memory
- —Coverage that nobody can see until someone asks
- —Target nomination as a scramble to reassemble the past
With Pluto
- ✓Evidence structured against the framework as it is generated
- ✓Your standard for translation applied identically to every program
- ✓Model's fidelity measured and documented, not just asserted from experience
- ✓Coverage and criteria visible from day one
- ✓Nomination arrives with reasoning already assembled
Agentic analysis
A harness, not a chatbot bolted onto an analysis tool.
An agent can't do much with one large, tangled system. It needs separate tools it can call, program context that lasts between sessions, and results in a predictable format. Pluto's analysis layer has worked that way since we started. A new entrant would have to build all of it before writing their first agent.
- READReading evidence package
- CALLpathway_enrichment(sig=PRT3789_48h, db=hallmark)
- CALLcell_type_deconvolution(bulk=tumor_panel_n12)
- SCOREMulti-omic biomarker consistency
- SCOREPathway / biological program activation
- FLAGConflict kept visible: Table 1 disagrees
- WRITECoverage updated 30 → 32 of 42
- PLANProposed: donor-diverse organoid panel to close Ring 3
Each step links to its source · records whether a person or the agent ran it
Context that carries over
Evidence, rationale and decisions are saved to the program, so the next person, or the next run, picks up where the last one stopped.
Every score shows its source
Each score links to the figure, table or region it came from. When two sources disagree, both stay visible so your team can weigh the conflict.
Program states
Watch a program earn its evidence.
Coverage isn't something that should be checked at the end; it should be continuously built. From early discovery to reverse translation, Pluto tracks the whole arc.
Stage 1 · Early discovery
Public genetics and expression evidence only. The inner ring lights up first; outer rings stay open.
9of 42
criteria evidenced
criteria evidenced
The framework
42 criteria, four dimensions
Each wedge is one criterion. The four outlined wedges are open as examples: hover one to see what it measures and the evidence the program holds for it.
Biology8Target8Model14Mechanism12
Standards
Translatable shouldn't be subjective.
Stop reinventing the standard for every program. Start applying it.
42 criteria, four dimensions, one readout.
A shared definition of what counts as translatable evidence, applied the same way to every program so two assessments a year apart are still comparable.
Preview the framework →Built for how the field is actually moving.
Model fidelity assessed in step with new approach methodologies and current FDA guidance — so the evidence you generate now still counts when the expectations tighten.
See the approach →Deployment
Your data never leaves your infrastructure.
This is not a packaging option. It is the precondition for real program data entering the system at all.
Customer-hostedNo training on your dataImmutable audit trailRBACE-signatureHuman vs. AI attributionModel agnostic
The questions your IT reviewer asks
Do you train models on our data?
No. Pluto runs inside your infrastructure and your data is never used to train models — ours or anyone else's. There is no path by which proprietary program data leaves your environment.
Which models does Pluto use, and where do they run?
Pluto is model agnostic. The harness defines the bounded capabilities and enforces the output structure; which model fulfils a call is configurable, including models you host yourself, in the region you choose.
How do we tell what the AI did versus what a scientist did?
Every record carries human-versus-AI attribution, and any generated statement traces back to the dataset, run or source region behind it. The audit history is immutable.
Can an analysis be reproduced a year later?
Yes. Pipeline versions are pinned and parameters recorded, so a run can be re-executed and compared against the original result.
What can we export if we leave?
Programs, evidence, analysis outputs and audit history, in open formats. The system of record is yours, and it already lives in your infrastructure.
Start with one program.
We map it against the framework and hand back the coverage — including every criterion still open.
You bring a program, not a seat count.