Question about time of cell fraction estimation (RLR)

[QiongZhao]

Hi Weixu,

I am using ENIGMA to estimate cell fraction to estimate cell type-specific gene expression. I created an ENIGMA object using the 'single_cell' type, where the bulk is a matrix with 971 samples and 28,285 genes, and the reference is from 10 other samples, containing 12,506 cells and 28,285 genes. However, when I use the created ENIGMA object with the RLR method to estimate cell fractions, it shows that the program needs three days to run (ncores=5). I would like to ask if this is normal.

At the same time, I tried selecting 20-1000 genes in the Reference to get results faster. But an error occurs: Error in rlm.default(x, y, weights, method = method, wt.method = wt.method, : 'x' is singular: singular fits are not implemented in 'rlm'.

Kindly regards,

Zhao

[WWXkenmo]

Hey,

Sorry for being late response, I am busy on my submission these weeks.

Is definately not normal, how are you actually implement deconvolution, could you share your code please?

All the best,
Weixu

[QiongZhao]

Hi, Weixu
I have already found the answer to my previous question. I am working with single-cell format data and have not performed batch effect correction.

I have a question about normalization method in the ENIGMA_L2_max_norm function for deconvolution.

1. Both my single-cell (reference) and bulk data are already TPM- ormalized.
2. I noticed the function has a do_cpm parameter, which defaults to TRUE.
3. Through my experiments, I found that when the data is TPM input, setting do_cpm=T and preprocess='log' yields the best results.
4. I am confused about whether the do_cpm parameter would re-normalize TPM-transformed reference data using CPM, and whether it would apply log transformation to TPM-input bulk data. Moreover, I couldn't find the implementation of the do_cpm parameter in the source code.

Could you kindly provide some advice on the following:

1. What are the most reasonable parameter settings when my reference and bulk data are already TPM-normalized?

2. Should I instead use raw count data and then set do_cpm=T and preprocess='log'?

3. Do the reference and bulk data need to be normalized using exactly the same method?

I would greatly appreciate your insights and guidance.

Thank you very much for your help!

Best regards,
Qiong Zhao