Clarification on ChromBPNet Tracks Used in Extended Fig. 2B

[tristan-ka]

Hi, thanks for releasing ChromBPNet. I’m trying to reproduce the Pearson correlations shown in Extended Fig. 2B, but the values I obtain from ENCODE do not match, even when evaluating on the full chr11.

1. Which ChromBPNet output was used in Extended Fig. 2B?

For each assay, was the figure computed using:

• Predicted-signal (e.g., ENCFF081XKZ, ENCFF584WBW, ENCFF097JRL), or
• Bias-corrected signal (e.g., ENCFF772NHP, ENCFF264HQR)?

The bias-corrected tracks give much higher correlations than what appears in the figure, so I’d like to confirm which one is correct for reproducing the results.

2. Why do ENCODE profiles for H1 not match the paper’s Pearson values?

Using H1 DNase-seq as an example:

• Ground truth: ENCSR000EMU
• Predicted-signal: ENCFF081XKZ
• Bias-corrected: ENCFF772NHP

Computing Pearson correlation on full chr11 (raw or log1p) gives values that do not match those shown in the figure.
Are any additional preprocessing steps applied (e.g., smoothing, normalization, clipping, filtering)?

Thanks — happy to share our code or full reproduction details if helpful.

[akundaje]

The count head performance eg Ext Fig 2B are using total predicted coverage log(counts) in 1000 bp around peak regions vs observed log(counts) for the same regions. Each point in the scatter plot is a peak. You would have to take the "predicted signal track" and sum the total reads in each 1 KB region at each peak. It's not at base pair resolution. The peak region coordinates are also provided as the portal. The profile prediction performance is on bp resolution normalized probability profiles in peaks. Take a look at the ChromBPNet code base to see how the performance is computed. Anusri can point you to the notebooks that can help you replicate the results.

…

On Wed, Dec 10, 2025, 8:30 AM Tristan Karch ***@***.***> wrote: *tristan-ka* created an issue (kundajelab/chrombpnet#270) <#270> Hi, thanks for releasing ChromBPNet. I’m trying to reproduce the Pearson correlations shown in Extended Fig. 2B, but the values I obtain from ENCODE do not match, even when evaluating on the full chr11. 1. Which ChromBPNet output was used in Extended Fig. 2B? For each assay, was the figure computed using: - Predicted-signal (e.g., ENCFF081XKZ, ENCFF584WBW, ENCFF097JRL), or - Bias-corrected signal (e.g., ENCFF772NHP, ENCFF264HQR)? The bias-corrected tracks give much higher correlations than what appears in the figure, so I’d like to confirm which one is correct for reproducing the results. 2. Why do ENCODE profiles for H1 not match the paper’s Pearson values? Using H1 DNase-seq as an example: - Ground truth: ENCSR000EMU - Predicted-signal: ENCFF081XKZ - Bias-corrected: ENCFF772NHP Computing Pearson correlation on full chr11 (raw or log1p) gives values that do not match those shown in the figure. Are any additional preprocessing steps applied (e.g., smoothing, normalization, clipping, filtering)? Thanks — happy to share our code or full reproduction details if helpful. — Reply to this email directly, view it on GitHub <#270>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABDWEJ4GS7WVGHKSCXZLFL4BBDB5AVCNFSM6AAAAACOUK4Y4GVHI2DSMVQWIX3LMV43ASLTON2WKOZTG4YTKOJTG43TCMI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>