Memory use and issues with fold change output

[SofieHvid]

Memory use and issues with fold change output

[ValeSora]

Hi @SofieHvid! We found the problem, and we are working on fixing it. I'm going to ping you here as soon as it's solved.

[iprada]

Hi Sofie,

Regarding splitting the samples into small batches, your approach is definitely compatible. Regarding longer runtimes, that is probably expected, but I would need to know how long is "longer time" w.r.t to running the entire dataset.

Regarding the log fold-changes: That looks odd, I agree. In order to figure out what happened, can you give some details about how you run dgd?

best,

Iñigo

[SofieHvid]

It takes about 3.5 hours to process the entire dataset.

We run dgd as done in the tutorials, without changing anything but the input. We input our raw count expression matrix (spliced into smaller sample sizes) instead and followingly 'restack' the output before saving, else nothing is changed. Likewise we compute the log fold change as shown in the second tutorial.

Thank you for the fast response.

[iprada]

good. thank you.

We will give it a look. It is worth stating that I have never seen bulkDGD produce mostly positive log fold changes.

We will get back to you. It might take some time.

best,

Iñigo

[SofieHvid]

That sounds great, thank you.

Regarding the issue, it seems as if the variable pred_means does not get rescaled when using the function perform_dea to compute the log2 fold changes only, and instead uses the decoder output directly as pred_means. Is this correct?
It appears as if the pred_means only gets scaled if the get_p_values method also gets called.

[ValeSora]

@iprada is what SofieHvis reporting expected behavior for the log2 fold changes calculation? Otherwise, I'll implement the rescaling.

[ValeSora]

@SofieHvid, the issue should be resolved now! Let us know whether you encounter any more issues.

[SofieHvid]

Sounds good!