Xaira's X-Scientist AI Agents for Drug Discovery

Xaira Therapeutics, the AI drug-discovery company, has unveiled X-Scientist, a team of AI agents meant to work alongside its researchers from early target hunting to experiment planning. Its headline result comes from a wet lab, not a leaderboard: in a test on liver cells, genes picked by the agents worked about as often as genes picked by human scientists, the company says.

Xaira calls the findings “initial evidence”. They are its own results, published on its website rather than in a journal.

How the agents divide the work

X-Scientist pairs frontier language models with biology tools, Xaira’s internal data and its own models, including X-Cell for target identification, X-Design for molecular design and X-Patient for patient stratification. A lead agent plans an investigation, hands tasks to worker agents and checks the evidence they bring back, while a task graph keeps track of the plan as it changes.

On two public benchmarks, DrugDiscoveryBench and the life-sciences part of Terminal-Bench-Science 0.1, Xaira says its system beat other set-ups using the same underlying model, even with its own biology tools switched off for the tests. With GPT-6-Astra, it beat the Codex harness by 7.4 points and mini-SWE-agent by 4.1. With Claude Sonnet 5, for example, it scored 13.4 percentage points higher on DrugDiscoveryBench than Anthropic’s Claude Code.

The liver-cell test

The more telling experiment used HepG2 liver cells treated with TGF-β, a signalling protein that drives fibrosis in chronic liver disease. Xaira compared 381 genetic perturbations nominated by X-Scientist with 321 chosen by human scientists and 433 chosen at random, then measured each in the lab with Perturb-seq to see which pushed the cells back towards an untreated state. A perturbation counted as a hit if the shift reached a z-score of 3 or more.

The agents’ hit rate was 7.1%, against 6.9% for the human-curated set and 0.9% for random picks. X-Scientist produced its list in about 20 minutes, against more than three hours for people, which Xaira describes as “approximately 10x faster”. Besides well-known pathway genes, the agents found less obvious ones. Three of these beat the median effect of the canonical hits, and none appeared in the human list.

In practical terms, the agents matched people on accuracy rather than beating them, and the gain is speed and breadth.

What comes next

Xaira wants to close the loop between reasoning and the lab. It says it will soon share more on X-Automate, its AI robotics platform, and how it will feed experiment results straight back into X-Scientist. It has also run an early pilot of the system doing autonomous research on its own. Its stated aim is a “virtual drug hunter”. For now, the evidence is one cell type, one pathway and one company’s numbers.