Exploring simulators for particle picking in cryo-electron tomography

To understand how proteins function, we need to know the conformations that they adopt and with what they interact in their native cellular environment.Cryo-electron tomography (cryo-ET) offers a powerful tool by enabling \textit{in situ} imaging of proteins. But high noise levels and the need for expertise in particle identification limit its scalability. In this study, we present a machine learning framework for automated recognition and localization of particles in cryo-ET data. We treat particle picking as an object recognition task and employ a U-Net-based architecture for multi-class segmentation. To overcome the scarcity of annotated data, we train our model on synthetic tomograms generated by a simulator that incorporates empirical noise from publicly available cryo-ET datasets. Our results show that training on a mixed dataset containing both synthetic and empirical backgrounds provides the most effective particle-picking performance, enhancing the model’s robustness to different background types. Furthermore, we demonstrate that training exclusively on simulated particles enables the model to reliably distinguish particles from background in real tomograms, highlighting the potential of simulation-based training strategies in cryo-ET.