AI Health Spark Seminar Series: Entering the Era of AI-driven Cryo-EM with SmartScope

August 2, 2022
12:00 pm to 1:00 pm
Virtual

Event sponsored by:

AI Health
+DataScience (+DS)
Biostatistics and Bioinformatics
Computer Science
Department of Radiology
Electrical and Computer Engineering (ECE)
Information Initiative at Duke (iiD)
Pratt School of Engineering

Contact:

Duke AI Health

Share

Speaker:

Presented by: Alberto Bartesaghi, Associate Professor of Computer Science, Duke University with host Maciej Mazurowski, PhD; Associate Professor in Radiology, Duke University
Despite the recent success of AI-based strategies for protein structure prediction, single-particle cryo-electron microscopy (EM) is still considered a gold-standard technique for determining three-dimensional structure of proteins at resolutions that are sufficient to visualize molecular level interactions. Determining a structure with this method, however, requires collecting images for several days and subjecting them to lengthy data processing protocols that can take weeks to complete. Moreover, a single biomedical study typically requires the determination of multiple protein structures, only adding to the logistical, resource and computational complexity of single-particle cryo-EM projects. In this talk, I will describe our efforts to employ AI-based techniques for feature recognition to drive the development of strategies to accelerate structure determination using cryo-EM. For example, finding the conditions to stabilize a macromolecular target for imaging remains the most critical barrier to determining its structure. SmartScope is the first framework to streamline, standardize, and automate specimen evaluation using deep-learning-based object detection, allowing it to perform specimen screening in a fully automated manner. During the downstream data analysis, randomly distributed copies of the protein of interest need to be identified, extracted and averaged in 3D to obtain a high-resolution structure. Existing neural-network-based detection algorithms require extensive labeling and are very slow to train. Leveraging positive unlabeled learning and consistency regularization, we propose a novel framework that is able to identify particles much faster than previously possible while still using very few labels. Altogether, these advances in AI-driven cryo-EM greatly facilitate and accelerate the structural analysis of important biomedical targets, thus lowering the barrier of adoption of this powerful technique for protein structure determination. This session is a part of the monthly seminar series organized by SPARK: AI Health Initiative for Medical Imaging. The seminar will highlight outstanding work in medical imaging at Duke and beyond. The seminar recordings will be publicly available. The SPARK initiative focuses on development, validation, and clinical implementation of artificial intelligence algorithms for broadly understood medical imaging by bringing together the technical and clinical expertise across Duke campus.