Implementations of models for 80nt, 400nt and 2K context

[qxdu]

Hi there,

I’m working on integrating SpliceAI into my sequence design workflow, where the sequences I’m dealing with are quite short, ranging from tens to hundreds of base pairs. As such, I’m keen on utilizing the models you’ve provided for 80/400/2K nucleotide contexts. Specifically, I’m looking to apply the 80 bp flanking model.

However, I have a couple of questions that I hope you can clarify:

(1) I’m unsure which of these models performs the best in terms of accuracy. Is it the 80, 400, or 2K nt model? Additionally, would the 10,000 bp model also be highly effective for predicting short sequences?

(2) The models downloaded from Google Drive have two different suffixes, “_c” and “_g”, such as SpliceNet80_c1.h5 and SpliceNet80_g1.h5. Could you please explain the difference between these two models and how they should be applied in practice?

Your guidance on these matters would be greatly appreciated.

[kishorejaganathan]

Please use the _g models (which were trained on all canonical and GTEx splice sites), you can ignore the _c models (which were only trained on canonical splice sites). Higher context size models are more accurate in general, but if your sequence lengths are short, then it does not matter which context length you choose as long as it covers the entire sequence.