Potential impact of shift mismatch between bias model and training sample

[ypauling]

Hi ChromBPNet team,

Thanks for developing and maintaining this tool!

I have a question regarding potential shift mismatches between a trained bias model and the ATAC dataset used for bias-corrected model training.

Here is my setup:

• I trained a bias model using a HOMER tag directory generated from ATAC-seq data.
• I converted the BED files in the tag directory into tagAlign format for ChromBPNet bias model training.
• Initially, the pipeline raised the following error:

Input shift is non-standard ({:+}/{:+}). Please post an Issue.

• To proceed, I manually shifted fragment positions by +4/-5, following the ENCODE ATAC-seq pipeline convention.
• After this adjustment, the estimated shift reported by ChromBPNet became +4/-4.

However, when I later used this bias model to train a bias-corrected model on another ATAC-seq dataset:

• The estimated shift for the new sample was 0/0.
• The resulting Tn5 profile score consistently remained >0.003.

From my understanding, this may indicate incomplete bias correction.

My question is:

Could a discrepancy between the inferred shifts of the bias model (+4/-4) and the target ATAC sample (0/0) negatively affect bias correction performance and lead to elevated Tn5 profile scores?

Or alternatively, does the 0/0 shift estimation suggest a preprocessing issue in the input data itself?

Thanks in advance for any suggestions or clarification!

Best,
Bing