Semi-supervised scArches about scarches

[M0hammadL]

Hey, thanks for using the model :) I would suggest you include more datasets in the reference building. There are two factors in having query integration:

1. The reference model needs to have seen multiple dataasts in order to understand the meaning of batch
2. The quality of the datasets are important and need to include multiple cell types and enough cells in each batch

[chsher]

Thanks for your helpful feedback. I tried splitting the reference into two batches (10x v2 vs 10x v3 sequencing technology), but the improvement was marginal (accuracy of 61%).

• The unsupervised trainer trains for 332 out of 500 epochs, and the semisupervised trainer trains for only 12 out of 200 epochs. Is there anything I can change about the training regime to improve performance?

• In the reference dataset, I have approximately 200k cells, 400 genes, and 95 cell types. Is this too few cells and/or too few genes given the number of cell types?

[M0hammadL]

Is the problem solved now? happy to follow up on this.

[chsher]

Thanks @M0hammadL the problem is unresolved.