Inquiry on Best Practices for DEG Analysis Using scGPT Model Outputs

[ellieujin]

I have been utilizing your scGPT model to perform perturbation prediction tasks by fine-tuning it with perturb-seq data. Your model is exceptionally well-designed and has significantly advanced my research.

In my current workflow, I compare the model’s prediction results with gene expression values from control cells to perform differential gene expression (DEG) analysis. Specifically, I generate prediction values equivalent to the number of control cells before the model averages multiple predictions, and then compare these values with the control gene expressions to identify DEGs.

However, I have encountered an issue where the gene expression distribution of the model’s output differs from that of the control cells, which results in a distorted volcano plot, as shown in the figure below.

It appears that the model’s output does not adequately reflect the sparsity typically observed in single-cell data. To address this, I performed imputation on the control cells to better align their distribution with the model’s output. Despite this adjustment, DEG analyses using both Wilcoxon and MAST methods still resulted in p-value inflation and much more downregulated DEGs than up DEGs, as shown below.

Given that DEG analysis is crucial for deriving meaningful biological insights from the model, I am seeking your guidance on best practices for conducting DEG analysis with scGPT model outputs. Your expertise would be invaluable in guiding my research forward. I would greatly appreciate any recommendations or insights you could share.

Thank you very much in advance!