Expression reference dataset giving high noise and tumor proportion

[katiedaco-diamondage]

Hello,

I'm running numbat on bone marrow samples from subjects with a heme malignancy. I am using a public dataset of the same tissue and technology types as the normal reference expression, and I get a warning of high MSE, and the results give very high proportion of tumor cells - way more than we would suspect based on orthogonal measurements.

I'm wondering if anyone has ideas on how to improve the results using this reference expression set. I tried providing the annotated cell types and cluster id's as groups, and both give similar response. My understanding is that numbat takes raw counts, and there shouldn't be a need for any normalization, but I'm wondering if there is some pre-processing that can be done so that my samples map better to this reference set (using harmony outside of numbat, my samples indeed map well to this public dataset).

Could it be that the fact that these are bone marrow cells, with a range of cell types that is making this happen?

I appreciate any advice or thoughts

[teng-gao]

Hi,

High MSE typically indicates the reference expression profile doesn't match your target sample well. A few things to check:

1) Reference format:
The reference must be a gene × cell-type matrix created using numbat::aggregate_counts() from raw counts + cell type annotations. Don't use Seurat normalization or other methods (#29, #97).

2) Cell type specificity:
For bone marrow with diverse cell types, ensure the reference includes the relevant normal cell types present in your sample. You can differentiate cells of the same type from different samples in the group column if needed (#205).

3) Exclude tumor cells from reference:
Make sure no tumor/malignant cells are included in your diploid reference (#62).

4) Consider a custom reference from your own normals:
If you have normal bone marrow cells from the same experiment, using those as an internal reference often works better than external datasets (#98).

5) CNV calling thresholds:
Adjusting CNV calling thresholds can help reduce false positive CNV calls in normal cells, which can cause them to be misclassified as tumor:

• Increase min_LLR (default 5) for more stringent pseudobulk CNV calling
• Lower max_entropy for stricter filtering of ambiguous CNV calls

Related issues: #29, #62, #97, #98, #205