Question about RDA pretraining

[yunhak0]

Hello,

I am deeply impressed by your scFoundation model in the field of single-cell transcriptomics. I would like to ask about the underlying principles and insights of the Read-Depth-aware modeling pretraining method.

1. How are the two total count indicators, T (‘target’) and S (‘source’), utilized within the model?

As I understand, the target count indicator (T) is not directly used in the loss function. Thus, I am curious about how the model uses T to reinforce sequencing depth at higher counts. Given that T is used as an additional input through simple concatenation, how does the model interpret this input as the intended sequencing depth to be enhanced? Once T is recognized as the target sequencing count, by what mechanism does the model operate to improve overall sequencing depth?

2. What is the basis of the insight that led to employing this technique?

I fully agree with the motivation for addressing sequencing depth variation across datasets due to different experimental conditions. However, I am interested to know the rationale or evidence that guided the insight that using S and T as additional inputs would improve performance.

Thanks.