CombiCast: Human-in-the-Loop Multi-Agent Decision Support for Combination-Therapy Partner Selection

Choosing drug combinations requires evidence on effects, safety, and suitability for the intended patients, drawn from papers, trial registries, and labels. We present CombiCast, a decision-support system that connects candidate discovery, evidence assessment, structured AI review, and human review. Across four discussion phases, agents link assessments to source records and record rationales, disagreements, and evidence gaps. A fixed-rule Chair ranks research priorities separately from the evidence requirements for advancement. In a multiple-myeloma case, CombiCast prioritized isatuximab and daratumumab as partners for carfilzomib-dexamethasone, while identifying unresolved patient-selection and drug-interaction questions. We compared four discussion setups in 32 runs across eight cancer-treatment contexts using the same supplied candidates and evidence within each case. Generalists reproduced the peer-review team’s full ordered shortlist in five of eight contexts using 12–13% of the team’s input-plus-output tokens. Specialists contributed additional trial-planning detail in a separate repeated study: they calculated required trial event counts and identified missing information needed to estimate enrollment and follow-up. CombiCast records why candidates were retained and what evidence is missing to support human review and further evidence collection.