External Embeddings

You can use RECOVAR's volume generation (kernel regression) with latent spaces produced by other heterogeneity methods. This is useful for:

• Improving resolution of cryoDRGN or other method's reconstructions
• Validating results from neural network methods (RECOVAR's kernel regression is transparent and produces no hallucinations)
• Combining strengths of different methods

Usage

recovar reconstruct_from_external_embedding particles.star \
    --poses poses.pkl --ctf ctf.pkl \
    -o external_output \
    --embedding z.pkl \
    --target coords.txt

Arguments

Flag	Default	Description
particles	Required	Input particles (.mrcs, .star, .cs, or .txt)
--poses	Required	Poses file (.pkl)
--ctf	Required	CTF parameters (.pkl)
-o, --outdir	Required	Output directory
--embedding	Required	External latent coordinates (.pkl, shape N x zdim)
--target	Required	Points at which to generate volumes (.txt)
--Bfactor	0	B-factor sharpening
--n-bins	50	Bins for kernel regression
--zdim1	False	Enable for 1D latent space
--tilt-series	False	Use tilt-series data

Example: using cryoDRGN embeddings

1. Run cryoDRGN to get latent coordinates (z.pkl)
2. Pick target points (e.g., k-means centers): np.savetxt("coords.txt", centers)
3. Generate volumes using cryoDRGN's latent space with RECOVAR's kernel regression:

recovar reconstruct_from_external_embedding particles.mrcs \
    --poses poses.pkl --ctf ctf.pkl \
    -o cryodrgn_recovar \
    --embedding z.24.pkl \
    --target coords.txt --Bfactor=50

The resulting volumes use RECOVAR's transparent kernel regression for volume generation but follow cryoDRGN's latent space structure.