[Question] - Recommendations on usage for ASVs?

[jolespin]

I wanted to generate embeddings for amplicon sequence variants (ASV) and was wondering if this model would work well for highly similar sequences that are around 150bo. In particular, the V3-V4 region of prokaryotic genomes.

If so, I had a few questions:

• are there any parameters or commands you would recommend when only using DNA and not multimodal?
• what is the resulting embedding dimensionality?
• are there any transformations recommended on the outputs? For example, dnabert-s recommends averaging.

[scopello]

Thanks for the question @jolespin!

• You should be able to only input DNA, but this may be slightly out of distribution for the model as it was trained on contigs.
• For the 650M param model, the output dimension after mean-pooling is 1280.
• We recommend mean pooling, same as dnabert-s.

[jolespin]

Awesome, thank you for the additional context. What is the maximum sequence length (is this the context window?) that is accepted by the 650M param model?