Fine-Tuning AI Models for Cryo-Electron Microscopy Data Collection

Letian Deng

Mentor: Dr. Jana Ognjenovic, Frederick National Laboratory for Cancer Research.

Date/Time: August 23rd, 2024 at 2:00pm.

Abstract: This summer internship focused on the application and fine-tuning of artificial intelligence (AI) methods to enhance cryo-electron microscopy (cryo-EM) data analysis. Over the first five weeks, I explored the cryo-EM workflow using two prominent tools: cryoSPARC and RELION. Through tutorials and the analysis of real betaGal data, I achieved a final resolution of less than 1.8  with cryoSPARC, while in RELION, time constraints limited the resolution to 3.3  for betaGal 3D model.

In the subsequent four weeks, I fine-tuned the Florence2 model to improve cryo-EM data collection. The model was refined to identify the thinnest part of the ice within a hole, improving upon previous methods that centered the entire hole. This approach successfully identified contaminated data that previous methods had missed. Although this work was not fully completed, I documented my steps and provided future improvement suggestions.