DrugHarness
An AI agent system for computational drug discovery that integrates more than 80 scientific operations and biomedical resources. Through a conditional skill induction mechanism, it distills reusable, auditable, and revisable skills from execution trajectories, enabling the agent to continuously evolve in drug design tasks. An independent review agent verifies candidate skills to prevent self-confirmation bias, ensuring that only qualified skills enter the skill library.
Timeline (last 90 days)
DrugHarness couples composable scientific execution with conditional skill induction, with the goal of making AI Agents increasingly capable in real drug design tasks.
DrugHarness's environment integrates more than 80 scientific operations, biomedical resources, and task-specific code, and preserves the associations among candidate molecules, results, and artifacts.
After completing a task, DrugHarness induces action hypotheses with observable applicability conditions from trajectories and outcome feedback; these hypotheses then undergo source verification or targeted execution trials by an independent review Agent, and only those that pass are admitted as skills.
Without evolution enabled, DrugHarness exceeds the previously reported MolClaw-CC results on MolBench by 10.8 percentage points in affinity comparison accuracy and by 55% in the average number of top-three active hits.
After enabling skill evolution, a retrospective evaluation on skill development tasks shows that Qwen's affinity comparison accuracy increased from 62.2% to 81.1%, and the optimization success rate increased from 79.5% to 87.2%.
In a prospective HPK1 study, DrugHarness achieved more favorable median docking scores than Biomni among system-specific candidate molecules.
A synthetic compound designed by DrugHarness achieved a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference molecule, with activity comparable to selected compounds designed by two senior medicinal chemists.
The independent review Agent sets admission thresholds through source verification or targeted execution trials to control the quality of the skill library and prevent self-confirmation bias.
All Facts (8)
DrugHarness couples composable scientific execution with conditional skill induction, with the goal of making AI Agents increasingly capable in real drug design tasks.
50%UnverifiedDrugHarness's environment integrates more than 80 scientific operations, biomedical resources, and task-specific code, and preserves the associations among candidate molecules, results, and artifacts.
50%UnverifiedAfter completing a task, DrugHarness induces action hypotheses with observable applicability conditions from trajectories and outcome feedback; these hypotheses then undergo source verification or targeted execution trials by an independent review Agent, and only those that pass are admitted as skills.
50%UnverifiedWithout evolution enabled, DrugHarness exceeds the previously reported MolClaw-CC results on MolBench by 10.8 percentage points in affinity comparison accuracy and by 55% in the average number of top-three active hits.
50%UnverifiedAfter enabling skill evolution, a retrospective evaluation on skill development tasks shows that Qwen's affinity comparison accuracy increased from 62.2% to 81.1%, and the optimization success rate increased from 79.5% to 87.2%.
50%UnverifiedIn a prospective HPK1 study, DrugHarness achieved more favorable median docking scores than Biomni among system-specific candidate molecules.
50%UnverifiedA synthetic compound designed by DrugHarness achieved a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference molecule, with activity comparable to selected compounds designed by two senior medicinal chemists.
50%UnverifiedThe independent review Agent sets admission thresholds through source verification or targeted execution trials to control the quality of the skill library and prevent self-confirmation bias.
50%