SCT integration vs Harmony for large dataset

[PantehaK]

I'm working on a single-cell RNA-seq dataset of PBMCs from a case/control study involving 32 individuals, with plans to expand to ~80 participants (n=40 each group). Due to technical difficulties, we multiplexed 4 samples together and therefore sequenced multiple batches independently, and there is substantial variability in the number of recovered cells per donor. We performed SCTransform normalisation per sample, and then merged the objects and applied Harmony (RunHarmony()) to correct for sample-specific effects during dimensionality reduction. Another option we were looking at was to use SCT-integration with the anchor-based method to account for the sample-specific variability. Does correcting for sample impact biological signal? I am worried we are over-correcting and therefore missing out on real signal.

My other question is, let's say we've gone ahead with SCT-integration, for differential gene expression analysis between cases and controls within specific biological groups (CD4 T cells, CD8 T cells, etc), which assay should I be using? SCT, Integrated, or the original RNA? I understand that the integrated assay (from Seurat’s integration workflows) is designed for clustering and dimensionality reduction, not for expression-based analyses. But if we use Harmony, we do not have an integrated assay- only SCT. Given this context, is it correct to use the SCT assay with slot = "data" for DEG analysis and visualisation across PBMC groups and donors, even in the presence of sample-level heterogeneity?

I’d appreciate guidance on best practices for large-cohort datasets where the intention is to perform DEG between cases and control.