Help with ribo-seq profile prediction for CFTR

[luisub]

Dear seq2ribo authors,

I hope you are doing well. I am using your Colab notebook to predict the ribosome profile of CFTR CDS (ENST00000003084.11; 1,481 codons).
The sTASEP-only output appears distributed across the transcript, but the Mamba-polished profile becomes very small and is concentrated in the first few codons across the four cell-line checkpoints.

Could you please advise whether the released Mamba checkpoints were trained or evaluated on transcripts of this length, whether there is a recommended maximum CDS length or long-transcript workflow, and whether this behavior is expected? I have included a short Colab cell that reproduces the comparison. Also not sure, if I am loading the correct weights? please let me know if there is any error in use from my side.

Thank you very much for your time and for making Seq2Ribo available.

https://colab.research.google.com/drive/1qHMHI2YxHvNbYb6oQqsqGbSygrcwwGvC?usp=sharing

Best,

Luis

[gunkaynar]

Hi @luisub, thanks for identifying this, and for the clear reproduction. I'm working to solve it.

[luisub]

Hi @gunkaynar , thanks a lot for checking this. During the weekend, I kept working, looking at the gene, and this is what the real data from GWIPs-vis looks for in all cell lines: there are two cell lines with no counts, and the other two look that have few signals. Probably the gene is simply not expressed in the 4 cell lines? Then the model does not have enough data for inference?