Incyte hands target assessment to Kosmos TA: one workday per target, and 83% of de-risking experiments match committee requests

Drug-target assessment gains an agent version that fills out the evidence package before review, and Incyte and Edison Scientific use it to judge which targets are worth pursuing. Packages were typically written by scientists with deep biology knowledge who made a compelling case for disease mechanism and intervention, yet safety, tractability, competitive intelligence, portfolio fit and strategic timing were often missing, forcing scientists to consult additional specialists and slowing evaluation by weeks or months. A rubric built from transcripts of Incyte target-review meetings served as a common standard, and human-prepared and Kosmos TA-generated assessments of the same targets were scored retrospectively; Kosmos TA scored consistently higher, with gains concentrated outside the core biological rationale.

Kosmos TA treats the assessment as a living evidence package, folding in literature, patents and other relevant publications as evidence evolves. The agent workflow is built around broader evidence coverage while staying aligned with human expert judgment, intended to bring additional perspectives in earlier rather than replace the scientist's biological expertise. In Incyte's assessments, Kosmos TA assessed a target across all categories in less than one workday on average; Pichai Raman, Sr. Director of Computational Biology at Incyte, said Kosmos produced in a morning what would have taken him three months.

The de-risking experiments committees care most about were also tested, and Kosmos TA's follow-up proposals aligned closely with expert priorities. Determining follow-up experiments for de-risking is one of the most important tasks for target assessment committees, so performance on this task is treated as critical for any AI solution. For six targets with known decisions from committee reviews, Kosmos TA was asked to propose the critical de-risking experiments and these were compared with the experiments actually requested: 83% were a perfect or conceptual match, 16% a partial match, and only 1% missed, corresponding to a single example.

The demonstrated scope is how Kosmos TA, built by Edison Scientific with Incyte, performed in Incyte's target-assessment setting: the retrospective rubric comparison, the per-target assessment time, and the de-risking experiment proposals on six targets with known committee decisions.

The results come from a retrospective comparison over a limited number of targets; the 83% match rate covers six targets with known decisions, with a single missed example, and the material reports no prospective use or performance on a larger target set.