v4.2.0: Introducing cryoDRGN-AI ab-initio reconstruction

[michal-g]

In this major release we are introducing the cryodrgn abinit command which implements the cryoDRGN-AI ab-initio reconstruction method originally released in its own repo.

More info about running this command can be found at our documentation for this release. Note that the existing ab-initio commands abinit_homo and abinit_het have been marked for deprecation as abinit_homo_old and abinit_het_old.

We are also making the following other minor improvements to the codebase in this release:

• support for Python 3.13 as well as PyTorch 2.9
• bug fixes in parse_relion (Bug fixes for parse_relion #470)
• more GPU memory-efficient pose search by taking evaluation of candidate poses:
  

  cryodrgn/cryodrgn/pose_search.py

  Line 152 in 8833e1a

  y_hat = self.model(x)

  
  and doing it in chunks:
  

  cryodrgn/cryodrgn/pose_search.py

  Lines 154 to 157 in 3789c6c

  	y_hat = [
  	self.model(x_chunk).view(x_chunk.shape[0], 1, -1, YX)
  	for x_chunk in x.chunk(self.nkeptposes)
  	]

• adding ability to control number of volumes sampled along each principal component in cryodrgn analyze using --n-per-pc
• implementing keepreal for ImageDataset