DrugHarness studies how a scientific agent can learn reusable guidance without losing the conditions that made a previous action useful. The research paper combines a computational drug-discovery environment with a process for proposing, reviewing and revising skills from execution experience.
The point is not simply to give an agent more tools. A recommendation must connect the right scientific evidence to an observable situation, remain executable and be revisable when later results contradict it.
Build skills around applicability and evidence
The environment integrates more than 80 scientific operations, biomedical resources and task-specific code. It preserves connections between molecular candidates, results and the artifacts produced during a run.
From trajectories and outcome feedback, the learner proposes action hypotheses containing applicability conditions, recommended actions and supporting evidence. An independent review agent checks candidate skills against sources and can run targeted execution trials. Accepted skills guide later tasks, with new feedback available for revision.
That structure makes the condition part of the stored knowledge. A strategy that helps a set of candidates lacking diversity need not be applied to a different set whose immediate problem is a failed property constraint.
Keep benchmark and prospective evidence separate
On MolBench, DrugHarness without evolution reports 91.9% affinity-comparison accuracy and 1.24 mean top-three active hits. The paper compares those figures with reported MolClaw-CC results. Retrospective skill-development experiments raise Qwen's affinity-comparison accuracy from 62.2% to 81.1%, with optimisation success moving from 79.5% to 87.2%.
The paper also includes a prospective HPK1 study. A synthesised compound has a reported cellular IC50 of 14.19 nM, described as a 2.64-fold potency improvement over the best input reference. That is experimental evidence for a particular compound and assay. It is not evidence of clinical efficacy, patient safety or an approved treatment.
Evaluate whether the learned rule transfers
Our interpretation is that the review process is as important as the skill generator. A scientifically plausible rule can still use the wrong endpoint, confuse a prediction with a measurement or assume conditions that no longer hold.
For a reproduction, inspect one admitted skill from source evidence through its applicability test and final execution. Then use a task where the condition is deliberately absent. The question is whether the agent knows when to leave a learned recommendation unused, as well as whether it can repeat a successful action. The reported benchmarks and laboratory result offer different pieces of evidence for that broader goal.
Comments