A large-scale curated and filterable dataset for cryo-EM foundation model pre-training

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Cryo-electron microscopy (cryo-EM) is vital for near-atomic resolution 3D biomolecule reconstruction.
• Low signal-to-noise ratio (SNR) in cryo-EM data presents significant processing challenges.
• Foundation models show promise for biological imaging but lack cryo-EM specific datasets.

Purpose of the Study:

• Introduce CryoCRAB, the first large-scale dataset for cryo-EM foundation models.
• Enable development of advanced foundation models for cryo-EM data processing.
• Facilitate improvements in image denoising and feature extraction for structural biology.

Main Methods:

• Developed CryoCRAB dataset with 746 proteins and 152,385 raw movie frames (116.8 TB).
• Split cryo-EM movies into odd/even frames for paired micrograph generation for denoising.
• Stored data in HDF5 chunked format to enhance sampling efficiency and training speed.

Main Results:

• CryoCRAB provides a comprehensive resource for training cryo-EM foundation models.
• The dataset's structure facilitates efficient training and random sampling.
• Demonstrated potential for advancing cryo-EM image denoising and feature extraction.

Conclusions:

• CryoCRAB addresses the critical need for large-scale datasets in cryo-EM foundation modeling.
• The dataset supports the development of robust models for high-noise cryo-EM data.
• CryoCRAB is expected to accelerate progress in structural biology and drug discovery through improved cryo-EM analysis.