long-read alignments

[MichelMoser]

Dear @mchowdh200 ,

I would like to test your tool to filter SV calls using long-read alignments. As i understood, your model was trained with short-read alignments, correct? Do you think the model will perform as well for nanopore-read alignments or should i consider re-training the model?

Cheers, Michel

[mchowdh200]

Hi Michel! The samplot visualizations for paired end sequencing and long reads is a bit different. Things like depth of coverage will look the same, but long read visualizations lack elements like discordant read pairs and split reads, but include things like alignment gaps. These differences might confound the existing classifier. With that said, we actually are working a version that is trained on long read data.

Thanks, Murad

[volcano1998]

Hi dear @mchowdh200 , I read your paper and am impressed by your performance of removing false positives of long reads alignment. However, I did not find the pre-trained model for long reads specifically here. When I tried to use your default model on long reads, the performance is not so good. I wonder were you able to get the long read model working? Thank you!

Best, Can

[mchowdh200]

Hi Can, Sorry for the long delay in response. During the development of Samplot-ML, we trained a long-read model, but had very limited training data so performance was still poor -- and as you've already seen the short-read model didn't do well either (short read samplot visualizations contain elements that you wouldn't see in long read images (ie split/paired end reads). We are currently investigating other methods of training a long read model in the presence of scarce training data, but are not ready to show results at this time.

[volcano1998]

Gocha. Thank you anyway! Hope you could get a good result on long reads filter training, that would be very helpful! Let me know when you got it and I can test it out!