An auditable harness for biomolecular dynamics simulations by AI agents

A new tool called MDClaw has been developed to provide an auditable harness for biomolecular dynamics simulations carried out by AI agents. This innovation addresses the challenge of AI-generated biological hypotheses, which are often too complex to test quickly. MD simulations can examine hypotheses at the atomic level, but without a reliable process record, trust in simulation results becomes questionable when no reference exists.

Existing MD agents typically delegate the record-keeping to language models, leaving gaps in the simulation process. MDClaw aims to fill these gaps by incorporating validating tools and a directed acyclic graph (DAG) check stage before execution. This ensures that every stage of the simulation is recorded and monitored.

To validate MDClaw's effectiveness, the tool was tested on the MDDataBench platform, which grades the outputs against deposited simulations. Four open-weight models, when used with MDClaw, achieved success rates ranging from 81-100% across 31 test systems. The DAG complements the outcome check by ensuring that simulations are carried out correctly and in the right order.

An ablation study, which compared the results with and without the DAG, revealed that an MDClaw-free approach resulted in 11 successful runs that had inadvertently skipped essential steps like minimization or equilibration. In some cases, these runs simulated different structures altogether. By employing the DAG, MDClaw ensures that all simulations are executed accurately and that the trajectories can be traced back to their original inputs and settings, providing a reliable and auditable framework for AI-driven biomolecular dynamics simulations.