Le Zhang

Bio: Le Zhang is currently the vice dean of the School of Computer Science at Sichuan University, a professor. His research mainly focuses on artificial intelligence and computational biology. He has published more than 150 research papers with an H-index of 31 and over 3,000 citations. He has led multiple research projects, including Major Special Projects of the Ministry of Science and Technology and projects funded by the National Natural Science Foundation of China. His achievements have been recognized with awards such as the Wu Wenjun Artificial Intelligence Science and Technology Award and Sichuan Science and Technology Progress Award.

Title: CryoPromptSeg: Prompt-guided Segmentation with Integrated Denoising for Cryo-EM Particle Picking

Abstract: 

Motivation: Cryo-electron microscopy (Cryo-EM) single particle analysis (SPA) is a key technique for revealing the structure of biomacromolecules by three-dimensional reconstruction. Achieving high-resolution reconstruction relies on the acquisition of a large number of authentic particles; however, manual particle picking is inefficient and inadequate for the demands of reconstruction, making automated particle picking a major research focus. Although the foundational segmentation model Segment Anything Model (SAM) has recently advanced automated particle picking, its segmentation advantages have not been fully realized in cryo-EM applications. Moreover, cryo-EM images often have significant noise. Conventional denoising decreases noise but frequently overlooks high-level semantic information, leading to oversmoothed particle regions and reduced particle distinguishability.

Results: To address these challenges, we propose CryoPromptSeg, which employs prompt-guided SAM for particle picking while integrating a semantically enhanced image denoiser. Specifically, by performing domain adaptation fine-tuning of SAM and incorporating prompts generated by the proposed automatic prompt generator, it achieves precise segmentation of cryo-EM particles. In addition, it employs a parallel multi-task framework to jointly train the denoiser and the prompt generator, incorporating particle semantic information from the prompt generator into the denoiser to suppress noise while preserving highly distinguishable particle structures. To lower the barrier to practical application, we developed a user-friendly online prediction platform for particle picking. Experimental results demonstrate that CryoPromptSeg outperforms existing mainstream methods in both particle picking accuracy and image denoising quality, thus providing a novel solution for the automation of particle picking.