AI agent supervision of structural model building and refinement

AI agents can assist in the building and refinement of structural models for biological macromolecules. This process involves numerous decisions that may require the expertise of an experienced researcher familiar with the various programs and experimental data. To test this concept, researchers utilized Claude Code with Claude Opus 5.5 to oversee 24 cryo-EM and X-ray cases over the course of approximately 36 hours on a single workstation.

The AI agent managed the preparation of inputs, executed established programs, examined validation results, compared candidate models, and ultimately determined the next step in the process. For 15 deposited cryo-EM models with suboptimal starting geometry, the median MolProbity score improved significantly from 2.62 to 1.76. In three crystal structures, the AI-generated models were found to be in close proximity to the published Rfree values.

It is important to note that while the agent can handle the placement and verification of chemical information such as ligands, ions, and modifications, this data must be provided by the researcher, similar to the sequence. The results indicate that agent supervision is a viable approach when every decision can be thoroughly reviewed by the researcher, who retains the final scientific judgment.

The researchers suggest practical requirements for implementing this method, with the potential for future AI agents to contribute to the creation of high-quality and reliable structures of biological macromolecules.