CryoSift: an accessible and automated CNN-driven tool for cryo-EM 2D class selection скачать в хорошем качестве

CryoSift: an accessible and automated CNN-driven tool for cryo-EM 2D class selection

Video interview with the authors. Published in Acta Cryst. F https://doi.org/10.1107/S2053230X2500... CryoSift: an accessible and automated CNN-driven tool for cryo-EM 2D class selection Single-particle cryo-electron microscopy (cryo-EM) has become an essential tool in structural biology. However, automating repetitive tasks remains an ongoing challenge in cryo-EM data-set processing. Here, we present a platform-independent convolutional neural network (CNN) tool for assessing the quality of 2D averages to enable the automatic selection of suitable particles for high-resolution reconstructions, termed CryoSift. We integrate CryoSift into a fully automated processing pipeline using the existing cryosparc-tools library. Our integrated and customizable 2D assessment workflow enables high-throughput processing that accommodates experienced to novice cryo-EM users. 00:00 Main editor intro: 01:00 Stephen Muench intro 01:10 Jan Hannes-Schäfer intro 01:30 Scott M. Stagg intro 01:45 Michael A. Cianfrocco 01:58 What specific gap in cryo-EM workflows were you seeking to plug with CryoSift? 03:00 How does CryoSift combine image features with metadata to improve class quality prediction? 04:30 When training the CNN, what training data were used and how did you ensure diversity across the samples? 06:04 Why do you recommend a cut-off of 3.5 and what is the trade-off? 07:34 How does CryoSift integrate into the different platforms, such as CryoSPARC or Relion? 08:49 Which evaluation metrics were use to evaluate reconstruction quality and why were they chosen? 10:55 How much does CryoSift help reduce subjectivity between practitioners? 12:10 What major limitations would you highlight for future improvement? 13:30 Given the rapid improvement in CryoEM and machine learning, how automated do you think the process is going to become? 15:00 How did the tool and the paper come together? 16:20 MX is know for large software suites such as CCP4 or PHENIX; will CryoSift be a step in that direction for CryoEM? 18:30 Unexpected features in data or metadata might be wrong, or right and interesting. How might the software prevent the suppression of interesting features? 20:33 What drove you to develop CryoSift?