An agent writes its execution experience into reviewable skills: DrugHarness designs and synthesizes an HPK1 compound with 14.19 nM cellular IC50
Synopsis
Computational drug-discovery agents usually rebuild their tool calls for every task, so a successful strategy rarely carries over. DrugHarness places more than 80 scientific operations and biomedical resources in one workspace, induces action hypotheses with explicit applicability conditions from execution trajectories, and admits them as skills only after an independent review agent checks sources and runs targeted trials when needed. On MolBench, the configuration without evolution reaches 91.9% affinity-comparison accuracy, 10.8 percentage points above reported MolClaw-CC results; in retrospective evaluation, frozen learned skills raise Qwen's optimization success from 79.5% to 87.2%.
Interpretation
Computational drug-discovery agents are starting to turn execution experience into reusable skills rather than only completing single tasks. Earlier agent systems mainly addressed tool access and workflow orchestration, and experience reuse appeared mostly as historical trajectories or rules without explicit conditions for when a recommendation applies. On MolBench, the DrugHarness configuration without evolution reaches 91.9% affinity-comparison accuracy and 1.24 mean Hits@3, 10.8 percentage points and 55% above reported MolClaw-CC results.
Skills are not written in directly; they begin as conditional action hypotheses and are admitted only after independent review. The learner compares decision situations across task objectives, execution steps, visible observations, deliverables and permitted outcome feedback, merging consistent recommendations and splitting or narrowing conflicting ones. The review agent checks sources and premises with the same scientific tools and code environment as the executor, running targeted trials when needed; on retrospective skill-development tasks, frozen skills raise Qwen's affinity-comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%.
The computational design is carried into experimental validation rather than stopping at docking scores. In a prospective HPK1 design, DrugHarness achieves a more favorable median docking score than Biomni among system-specific candidates, and one compound was then synthesized and tested by collaborating medicinal chemists. That compound reaches a Jurkat p-SLP76 IC50 of 14.19 nM, a 2.64-fold improvement over the best input reference at 37.47 nM, with activity competitive with two compounds independently designed and selected by medicinal chemists.
Perspective
The demonstrated scope is molecular screening, optimization and constrained design on MolBench and SMDD-Bench, a retrospective in-domain skill evaluation, and one prospective HPK1 design starting from four experimentally characterized references; that design modifies only terminal amide nitrogen substituents while preserving the shared core and specified double-bond geometry.
Skill evolution is evaluated on a single benchmark with a limited number of tasks under a retrospective, in-domain setting, and its effectiveness at larger scales and generalization to unseen targets and task families remain to be established; experimental validation is also limited in scale and coverage of developability properties, and systematic comparisons of runtime, computational cost and resource requirements are still open.