Discrepancy between paper and code: additional output layer for enrichment odds

[vpwang]

Hi,

I'm reproducing the results from Her et al. 2026 and noticed a potential discrepancy between the paper and the code.

Paper description:

"we extended RBPNet with an additional output layer that predicts the enrichment odds ratio"

Current implementation in RBPNet.py:

d_log_odds = self.dlogodds_bn(
    torch.sum(x_signal[:, self.mask:-self.mask-1], dim=-1).unsqueeze(dim=-1)
).squeeze()

d_log_odds is derived from x_signal (f_target) rather than from a dedicated layer from the bottleneck.

I have a higher crosslinking correlation and a lower enrichment correlation than in the paper. This pattern is consistent with the ablation study in the paper, suggesting the current code may correspond to the version without the additional output layer

Could you clarify whether the current d_log_odds computation is the intended implementation, or if there is a dedicated layer from the bottleneck that matches the paper's reported metrics?

Thank you!

[algaebrown]

Hi Chris,
Thanks for using this repo. I checked the repo is up to date
Here is the snakemake rules I made to train the models.
https://github.com/YeoLab/skipper/blob/dev/charlene/2025revision/rules/train_rbpnet.smk
I can confirm the additional layer is added on top of x-signal. The intuition is that the clean signal should inform how strong (d-log-odds) the binding is. I should have written it more accurately.

In the loss function, there is a parameter that weights w the enrichment loss vs the crosslinking loss. The weight is 30 by default, and is used by the model training
https://github.com/algaebrown/RBPNet/blob/main/losses.py#L68

In alblation experiment w=0 to disable learning the enrichment loss https://github.com/algaebrown/RBPNet/blob/dev/train_original_rbpnet.py

Did you set w to a different value? I do see w affects the performance a lot.

[algaebrown]

Hi Chris, there is also one detail that might result in lower correlation
in genomic bins with low coverage, the signal is extremely noisy. I recommend calculating the enrichment with genomic bins with at least 50 reads IP+INPUT, as noted in the manuscript "we retained only models exhibiting decent performance on the unseen test set (enrichment correlation > 0.5 in genomic bins with more than 50 reads; [Table S2]

[algaebrown]

Hi Chris,

I lost access to train stuffs on my original server :(, but I did a quick experiment on google colab and pulled the dataset from huggingface (I did upload most of them) .

Note this experiment does not fully capture the max 100 epochs and the full training enhancement jiggling and such data enhancements strategies, but these are all minor details.

Where in the original table S2
https://docs.google.com/spreadsheets/d/1AdFSRs3GrtpfoI-OkeElTGQBMQr-XZno/edit?gid=1718444976#gid=1718444976
RBFOX2_HepG2_ENCSR987FTF
dlogodds pearson(total>50) is 0.5213659024
I highlighted the row in light red. the enrichment odds columns are in bright yellow and notice the coverage did change the performance (due to noisiness)

The quick experiment I got a 0.58 slightly better than published.
Particularly:
dlogodds pearson(total>10) 0.498594
dlogodds pearson(total>100) 0.681309
dlogodds pearson(total>200) 0.743436
dlogodds pearson(total>50) 0.588760

I put my colab notebook here.
https://github.com/algaebrown/RBPNet/blob/main/rbpnet_colab_tutorial.ipynb

Hope this is somewhat helpful.