Chrombpnet model uncorrected

[linzyzhao2002]

Hello,

Hope this message finds you well. I am encountering a problem during training the chrombpnet model, where the report says "the model is uncorrected, average of the max of the profiles is 0.024". However, based on the Tn5 profile below, the maximum is below 0.001.

Additionally, I am training using a pretrained bias model, using the same peak and non-peak regions as the chrombpnet model. Please let me know why the error occurs, and any help would be deeply appreciated!

[panushri25]

Can you share your full report?

[panushri25]

Also if you can share the full output directory that will be helpful

[linzyzhao2002]

Can you share your full report?

Thank you so much for your help!! This is the output for my chrombpnet model:
train model output 19.pdf

This is the model output for my bias model:
bias model output.pdf
I trained the bias model with hyper parameter of 0.4.

[linzyzhao2002]

Also if you can share the full output directory that will be helpful
slurm_54367172.pdf
I'm attaching the full output directory.

Any help would be greatly appreciated - thank you so much for your time and help!!!

[linzyzhao2002]

Also just to add, I noticed that the pearsonr in peaks is always >0 (around 0.30) for the bias model. Could this be the problem here? Does this suggest anything wrong with the peak file or the bam file?

I generated the peak file using macs2 callpeak with qvalue cutoff of 0.01 and then removed the blacklist region. The bam files i used was deduplicated/removed MT reads. Please let me know and thank you so much for your help!!

[panushri25]

@linzyzhao2002 noticed you posted on the other thread too. Are you merging peaks across celltypes as well?

[linzyzhao2002]

@linzyzhao2002 noticed you posted on the other thread too. Are you merging peaks across celltypes as well?

No I'm not. I am training chrombpnet on only one cell type (bulk ATAC-seq for mouse T cells), sampled at different time points (0h, 2h, 5h etc). My bias model was trained with peaks called at the 5h time point. Do you have any suggestions? Thank you so much!

[linzyzhao2002]

Adding to the context, I am doing analysis on time-course ATAC-seq for in vitro activated mouse T cells. I am using a single time-point (say 5h) for training the bias model. I then use this bias model to train my chrombpnet model on the 5h timepoint bam file. However, I observed that the nobias model was uncorrected (average profile maximum for Tn5 is 0.024), and that Tn5 motifs were observed by TFmodisco. I tried sweeping the hyper-parameter (from 0.2 to 0.5) but still got the same results. Could you please let me know what could be happening here? Thank you so much!!

Sincerely,
Linzy

[panushri25]

Is the 5h timepoint the deepest dataset you have? I would recommend using the deepest time point to build the bias model and then use it across time points.

[panushri25]

Is this file here train model output 19.pdf using the 5h timepoint bam file or the bias model provided with the repo. Can you provide both reports using each of the bias models? I would generally recommend using the deepest time point to build the bias model and then use it across time points.

[linzyzhao2002]

Is this file here train model output 19.pdf using the 5h timepoint bam file or the bias model provided with the repo. Can you provide both reports using each of the bias models? I would generally recommend using the deepest time point to build the bias model and then use it across time points.

Thank you so much for your reply and happy new year!

I'm sorry if I didn't say this clear.

1. I trained a bias model using the 5h timepoint bam file
   bias model 3 output.pdf

2. I then use this bias model to train a non-bias model, using the same 5h bam file. However, i see that the model is uncorrected.
   train model output 19.pdf

Could you please suggest possible reasons why the model is uncorrected? Is there anything wrong based on the reports? Thank you so much!!!

[panushri25]

Its strange that your marginal footprints for the bias motif are different from 0.024, my guess is that the output directory is corrupted and has outputs from multiple runs.

[linzyzhao2002]

Its strange that your marginal footprints for the bias motif are different from 0.024, my guess is that the output directory is corrupted and has outputs from multiple runs.

Interesting!! Thanks for your suggestions. Does this mean that the model might in fact be "corrected", and I could use this for TFmodisco prediction? Thank you!

[panushri25]

I cant say from this because I dont know whats corrupted and whats not. Can you send a screen shot of the directories with its time stamps? Did you make two runs in the same directory?

It might be better to do a fresh run in a new directory!

[linzyzhao2002]

I cant say from this because I dont know whats corrupted and whats not. Can you send a screen shot of the directories with its time stamps? Did you make two runs in the same directory?

It might be better to do a fresh run in a new directory!

Thanks for the suggestion! I'm sorry I don't have access to the folders now since the cluster is down for maintainence, but for each model i used a new output directory (a new folder)!

[linzyzhao2002]

I cant say from this because I dont know whats corrupted and whats not. Can you send a screen shot of the directories with its time stamps? Did you make two runs in the same directory?

It might be better to do a fresh run in a new directory!

Maybe I will try TFmodisco using this nobias model and then see if the predictions make sense

[panushri25]

what is the read depth of your individual time points?

[linzyzhao2002]

what is the read depth of your individual time points?

The read depths for each time point is quite similar, around 50 million reads each! Is this sufficient?

[panushri25]

yeah thats fine! I think this more of a corruption issue ... from the time stamps you can make sure that all the files in this directory evaluation/ are dated after the creation of the chombpnet models.

If they are dated correctly you can post back the output html with a bias model provided with the repo and we can take it from there.

I will close this issue for now, do reply back or feel free to open a new issue if you have more questions.

[linzyzhao2002]

yeah thats fine! I think this more of a corruption issue ... from the time stamps you can make sure that all the files in this directory evaluation/ are dated after the creation of the chombpnet models.

If they are dated correctly you can post back the output html with a bias model provided with the repo and we can take it from there.

I will close this issue for now, do reply back or feel free to open a new issue if you have more questions.

Sounds good! Thank you so much for all your help!!