You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi! This usually happens when there are some unrealistically high expressing genes in the dataset. But from I have seen the optimization still converges for the remaining genes. It is just that the total or average loss is nan. So you can either filter very high expressing genes out before, or just leave it as it is.
Thank you for the reply! Is the correction calculated by the model different for every gene in one sample? I thought it just finds a per-sample correction
No and I think the optimization never fully converges. However, the prior distributions and the starting point for the optimization are already sensible background binding values for all genes (specifically the mean of negative probes in each sample). So the values should be ok even after short optimisations.
Activity
AlexanderAivazidis commentedon Nov 22, 2022
Hi! This usually happens when there are some unrealistically high expressing genes in the dataset. But from I have seen the optimization still converges for the remaining genes. It is just that the total or average loss is nan. So you can either filter very high expressing genes out before, or just leave it as it is.
innaaver commentedon Nov 23, 2022
Thank you for the reply! Is the correction calculated by the model different for every gene in one sample? I thought it just finds a per-sample correction
AlexanderAivazidis commentedon Nov 24, 2022
yes it is different for every gene
innaaver commentedon Nov 24, 2022
Is there a way to see for which genes the model did not converge?
AlexanderAivazidis commentedon Nov 25, 2022
No and I think the optimization never fully converges. However, the prior distributions and the starting point for the optimization are already sensible background binding values for all genes (specifically the mean of negative probes in each sample). So the values should be ok even after short optimisations.