Can the standard reference genome be directly used as input?

[zhangjy859]

Hi,

I noted you said: This script takes a fasta file containing 1000 bp records and outputs an hdf5 file containing the predictions for each record. for predict_ensemble.py, I want to know is it possible to run predict_ensemble.py with a known whole reference genome (such as hg38) for a genomewide search or its limited to short records (like 1000bp) in fasta file?

Best.

Zhang,

[adamyhe]

Hi Zhang,

No, the model itself only supports predicting on 1kb input sequences. In principle, you could fragment the entire genome and stitch the predictions together (keeping in mind that CLIPNET only outputs a prediction for the middle 500 bp), but I wouldn't really recommend it because:

1) It would be pretty compute intensive. 2) Most of the genome does not have active transcription initiation. It would be a lot more efficient to predict at, say, 1kb regions around cCREs from ENCODE or ATAC-seq peaks. 3) Most enhancers and promoters are cell-type specific, with their cell-type specific activity being driven by non-sequence factors like chromatin state, TF expression, etc. You'll end up with erroneous initiation predictions at many of these inactive regulatory elements.

If you have a specific use case in mind, please let me know and I can brainstorm some ideas.

Adam

[zhangjy859]

Hi,

thank you for your answer. I understand. I noticed your article before, and it was very interesting. I have some confuse about the description mentioned in this issue of the document. At the same time, I am also interested in whether it can work in non-model organisms with poorly annotated transcriptomes. You have answered my questions. Thank you!

[adamyhe]

Oh, yeah, I would not use this to call regulatory elements in poorly annotated genomes. You'll likely need some functional data to reliably call enhancers (promoters you can probably call from gene annotations).