Recommendation of method for inferring gene expression trends along latent time

[Evenlyeven]

Hi VeloVAE team,

Thanks for developing this tool, it is my favorite RNA velocity analysis tool so far!

I wanted to infer and plot a trend line for gene expression along latent time. Do you recommend a specific method (e.g. GAM, splines, LOESS) that aligns the best with VeloVAE's framework? Or if there is a built-in way to do so, kindly let me know.

Thanks for your help in advance!
Di

[g-yichen]

Hi, Di:

Thank you for your interest in our tool! Currently, there is limited support for line plotting in a gene fitting (u/s vs. time) plot. Our package depends on the loess package to generate curves. If you only need a trend line, please try setting the parameter frac in the plotting function. frac represents the fraction of points to consider in the local approximation (Please see the loess documentation). Please try the following code piece:

import velovae as vv
vv.post_analysis(
    adata,
    < EXPERIMENT ID >,
    ['VeloVAE'],
    [< KEY >],
    compute_metrics=False,
    genes=[< GENE_1 >, < GENE_2 > ...],
    plot_type=['gene'],
    frac=< FRAC >
)

Please be aware that this plotting function fits a loess curve for each cell type to reflect possible branching. Thus, it may not be great for visualization at the moment. The latent time is usually stored at .obs[< KEY >_time] and predicted cell-by-gene unspliced and spliced values are stored at .layers[< KEY >_uhat] and .layers[< KEY >_shat]. Here < KEY > is a prefix string for saving purpose when you call vv.save_anndata and should be 'vae' in the example notebook. You are welcome to take these results and plot them using better visualization methods.

Best,
Yichen