Agentic cryo-EM structure determination with CryoPilot
Abstract
Single-particle cryo-electron microscopy (cryo EM) relies on mature computational tools, but experts must still inspect intermediate results, choose among alternative processing routes, and decide when earlier choices should be revised. Here we present CryoPilot, a system of two language model agents that carries cryo-EM data from raw micrographs through three-dimensional reconstruction to validated atomic models. Rather than executing a predefined pipeline, its reconstruction agent evaluates intermediate images and quantitative diagnostics, selects and parameterizes CryoSPARC jobs, tests alternative strategies, compares their outcomes, and retains, revises, or reverts decisions accordingly. A persistent programming environment and session database preserve the complete processing history, making each decision traceable and allowing multi day workflows to proceed with minimal human intervention. Its modeling agent determines map handedness, builds atomic models with CryoZeta, removes regions the density does not support, and optimizes each model by testing refinement steps against map fit, geometry, and DAQ, a measure of local model to map agreement. Across nine public EMPIAR datasets, eight reconstructions reached 2.61 to 3.51 A resolution, and refined models from seven maps agreed with reference structures to within 0.4 to 4.9 A Calpha; RMSD. For a challenging alpha latrocrustotoxin dataset, a single change of particle picking strategy that we suggested enabled CryoPilot to improve the reconstruction from 5.9 to 4.15 A. CryoPilot establishes an adaptive and auditable framework for agent-driven cryo-EM structure determination.
Related
- Cryo-EM structure of bixafen-bound S. cerevisiae complex II unravels SDHI specificity against pathogenic fungi
- Cryo-electron microscopy structure of Jabs, a bacteriophage infecting the multidrug-resistant pathogen Mycobacterium abscessus
- CryoMV: Structure-Prior-Guided Modeling and Real-Particle Validation of Continuous Conformational Transitions in Cryo-EM
- Learning the Shared Latent Structure of Cryo-EM Inverse Problems
- Nucleocapsid-like cryo-EM structure of influenza B virus
BibTeX
@misc{punuru2026agentic,
title = {{Agentic cryo-EM structure determination with CryoPilot}},
author = {Pranav Punuru and Thomas Klose and Daisuke Kihara},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.29.762908},
url = {https://www.biorxiv.org/content/10.64898/2026.10.29.762908v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Agentic cryo-EM structure determination with CryoPilot
AU - Pranav Punuru
AU - Thomas Klose
AU - Daisuke Kihara
PY - 2026
DA - 2026/10/29
PB - bioRxiv
DO - 10.64898/2026.10.29.762908
UR - https://www.biorxiv.org/content/10.64898/2026.10.29.762908v1
AB - Single-particle cryo-electron microscopy (cryo EM) relies on mature computational tools, but experts must still inspect intermediate results, choose among alternative processing routes, and decide when earlier choices should be revised. Here we present CryoPilot, a system of two language model agents that carries cryo-EM data from raw micrographs through three-dimensional reconstruction to validated atomic models. Rather than executing a predefined pipeline, its reconstruction agent evaluates intermediate images and quantitative diagnostics, selects and parameterizes CryoSPARC jobs, tests alternative strategies, compares their outcomes, and retains, revises, or reverts decisions accordingly. A persistent programming environment and session database preserve the complete processing history, making each decision traceable and allowing multi day workflows to proceed with minimal human intervention. Its modeling agent determines map handedness, builds atomic models with CryoZeta, removes regions the density does not support, and optimizes each model by testing refinement steps against map fit, geometry, and DAQ, a measure of local model to map agreement. Across nine public EMPIAR datasets, eight reconstructions reached 2.61 to 3.51 A resolution, and refined models from seven maps agreed with reference structures to within 0.4 to 4.9 A Calpha; RMSD. For a challenging alpha latrocrustotoxin dataset, a single change of particle picking strategy that we suggested enabled CryoPilot to improve the reconstruction from 5.9 to 4.15 A. CryoPilot establishes an adaptive and auditable framework for agent-driven cryo-EM structure determination.
SN - 2692-8205
ER -