deepsignal2 vs deepsignal ?

[pterzian]

Hi Peng,

Glad to see a new version of deepsignal ! I have some questions:

• Did you already do a comparison of both tools (deepsignal vs deepsignal2) ? Did you use the same dataset to build the model ? Is it faster ? etc)
• Can we use this new version to train models the same way we did before ? Are you still using the bisulfite-based strategy to find high confidence sites ?

Best, Paul

[PengNi]

Hi Paul,

• We used the same dataset (HX1 nanopore reads) to train the models of deepsignal and deepsignal2. deepsignal2 has sligtly higher performance than deepsignal: 0.8879 vs 0.8824 pearson correlation with BS-seq using 5x nanopore reads. We did not test runnning time, but deepsignal2 should be a little faster.
• Yes, we used the bisulfite-based strategy to find high confidence sites, and the train processes of the two tools are basicly the same. Please check generate_samples_and_train.sh for more details.

Best, Peng

[pterzian]

Thank you for these answers. I am trying to understand if I can use features files extracted with the first version of deepsignal to train a model with deepsignal2. Or, call previously extracted features with your new model. I see you made some modifications to the features format and I would like to be sure of what I understood. From these lines of code :

https://github.com/PengNi/deepsignal2/blob/2136cb258f156046ac7163f6dfd257ca09891d28/deepsignal2/extract_features.py#L316

https://github.com/PengNi/deepsignal2/blob/2136cb258f156046ac7163f6dfd257ca09891d28/deepsignal2/call_modifications.py#L69

I understand you split the raw signal of one base with , and sets of raw signal with ; ?

If 1 base = 3 signal values it would be written out like this :

0.0,0.0,0.0;0.0,0.0,0.0;etc

Is this the only change between both features format file ?

[PengNi]

Yes, it's the only change. In deepsignal, we extract at most 360 signal values centered in the raw signals of 17mer; in deepsignal2, we extract at most 16 signals of each base, to form 17*16 matrix.

Best, Peng