How would you evaluate whether integrated clusters retain treatment-associated biology?

Why

A visually compact UMAP with perfectly mixed batches can be misleading. Identify comparable cell types across independent donors and inspect whether technical batches still dominate neighbor structure. At the same time, test whether expected lineage markers, validated treatment-response genes and rare genuine states remain distinguishable.

Suppose an interferon treatment creates a reproducible activated T-cell program. If integration forces the treated T cells directly onto resting T cells and removes their signature, it has overcorrected. Compare cluster stability and labels before and after integration, quantify biological conservation and batch mixing with more than one metric and inspect marker expression using original counts.

Where treatment and batch are inseparable, report that limitation instead of claiming successful correction. For differential expression, work with sample-aware models and an appropriate count representation rather than blindly testing integrated latent coordinates. External replication or targeted protein validation strengthens the finding.