Claude Science Workshop PV / CSW
At the Claude science workshop we built this suite of MCP tools, so an AI agent can call our deployed model and its trust layer directly: read a tumor's transcriptome, and just as importantly, see when the model should hold back. The prediction tools return calibrated uncertainty and decline rather than guess when they cannot answer. Research and educational use only.
A confident guess on an unfamiliar input is dangerous inside an agent loop. So the prediction tools hand back their uncertainty and decline rather than guess, and the research-grade heads carry their confirm-on-assay caveats, so the agent can decide whether to rely on the answer.
Research tier
classify_tumor_origin
Reads a tumor's site of origin from RNA-seq and returns calibrated confidence, the top alternatives, and an abstain option when no site is clearly ahead.
Research tier
trust_envelope
The honesty layer in one call: a calibrated probability, a conformal candidate set, a novelty flag, and an explicit abstain decision, so an agent can reason about whether to trust the call.
Research tier
novelty_check
Flags a profile that sits off the training manifold, so the model declines inputs it should not answer instead of guessing.
Research team tier
explain_prediction
Surfaces the genes that drove a specific prediction, so the call stays reviewable rather than a black box.
Research tier
panel_coverage_report
Reports how well your genes cover the model's panel, high, partial, or low, without ever revealing the panel contents.
Research tier
model_card_facts
Returns frozen, cited model-card facts, so an agent grounds its answers in real numbers and never invents a Provotics figure.
Research tier
harmonize_expression
Confirms a profile is accepted and mapped into the model's space, so inputs from different pipelines are comparable. It never returns your per-gene values back.
Research team tier
research_report
A deterministic, research-use-only report composed from the model's own outputs, with no language model in the loop and nothing invented.
Research team tier
proliferation_index
A within-type proliferation percentile, reported as a coherence signal, never a validated prognostic score.
Research team tier
immune_microenvironment
The tumor's immune phenotype, from cold to inflamed, as a descriptive research signal.
Research team tier
cohort_harmonize_classify
Batch cohort-alignment classification for a single-site study of 20 or more samples, never a single tumor.
Research team tier
molecular_portrait
Assembles the model's multi-head research portrait in one call, gathering the site call with the research-grade heads below. Research use only, and each head carries its own caveats.
Research team tier
msi_status
An expression-based microsatellite-instability signal, site-aware and reported research-grade. Confirm on a PCR or IHC assay; not a diagnostic call.
Research team tier
molecular_subtype
A molecular-subtype signal within the predicted site, where a subtype head exists (for example PAM50 on breast). Research use only.
Research team tier
driver_alterations
Expression-imputed candidate driver alterations, surfaced as research hypotheses. Confirm on a DNA assay before any use; never a call.
Pharma tier
drug_pathway_sensitivity
A within-patient research signal of pathway sensitivity. Not a therapy recommendation, and the underlying data license is under review.
Pharma tier
chembl_compound_lookup
Looks up compound and target context from ChEMBL for this tumor's over-expressed targets, as research leads, never treatment recommendations.
Forthcoming
tumor_purity_estimate
Tumor purity estimation. Not yet live: it returns an honest forthcoming notice until the clinical purity labels land.
Forthcoming
methylation_class
A methylation-based class signal. Not yet live: blocked on a methylation data pull, and returns a forthcoming notice until then.
Forthcoming
multiomics_integrate
Paired multi-omics integration. Not yet live: pending paired data matrices, and returns a forthcoming notice until then.
Forthcoming
depmap_drug_response
A cell-line drug-response signal. Not yet live: pending the data pull; research use only, never a therapy recommendation.
Seventeen tools are live today across the research, research-team, and pharma tiers, with four more forthcoming (defined and returning an honest not-yet-available notice until their data lands). Every tool is research-use-only, returns model-derived numbers only, and keeps the gene panel proprietary. The research-grade heads (subtype, microsatellite, drivers, immune, proliferation) carry confirm-on-assay caveats, and the pharma tools are license-review-gated. Access is by paid API key.
Access
The workbench runs over MCP with a paid API key. Request access to evaluate it under our access agreement.