Give AI agents the environment they need

Identify cancer-associated GTPases where mutations or amplification increase the amount of active, GTP-bound protein. Group them by mechanism. Pull experimental rates and conditions comparing mutant and wild-type activity. Cross-check which mutations occur in patients and drive disease, and distinguish these from engineered laboratory mutants.

Non-RAS GTPase funnel. Helped ID which GOF/amplification events raise the GTP-bound pool, where mutant-vs-WT hydrolysis kinetics actually exist, what RAC1 P29S looks like in melanoma, and the final cut where GNAQ/GNA11 were the cleanest fit. Also flagged potential false drivers that are rare, subclonal, catalytic-dead, or amplification-only.

Search for prior studies of each drug-disease pairing. Exclude drug-disease pairings that have already been investigated. Evaluate the the current standard of care and identify specific limitations a repurposed drug would need to address.

Paperclip narrowed the field to a handful of repurposing hypotheses supported by related disease biology. For the two strongest candidates, it identified the incumbent therapy and a concrete limitation to address, such as safety-monitoring burden or a lack of controlled trial data.

Pull trials across multiple international registries — sponsor, drug, dose, phase, endpoints.

A cross-registry trial map across ClinicalTrials.gov, ICTRP/IRCT, jRCT, ISRCTN, and others, with sponsor, drug, route, dose/frequency, phase, and endpoint. It also exposed registry blind spots and identified commercially relevant newer mechanisms: clascoterone/TO-210, GED-0507, tibulizumab, engineered microbiome, and several formulation plays.

What were the Pivotal trials for (the specific corticosteroid medication)?

Paperclip found that FDA documents described the same trial as both “supporting” and “supporting pivotal.” It retrieved the passages documenting that inconsistency, helping the consultant understand why the trial’s classification was confusing and understand how it should be classified.

Search for single-variant functional measurements, saturation mutagenesis, and reporter assays covering the LDLR promoter. Read the methods to establish the experimental construct and resolution of each assay. Build a benchmark table linking each variant’s position, construct, and measured effect, and identify which measurements can test sensitivity to single-nucleotide changes.

Paperclip assembled the published measurements into a benchmark table. When the researcher tested the model against it, the model tracked large deletions but missed single-base substitutions. That was enough to reject it for this task before building the variant-scoring infrastructure.