Why can batch effects distort a single-cell RNA-seq analysis?

Why

Single-cell count matrices contain variability from capture, ambient transcripts, dissociation, sequencing depth, chemistry and donor biology. A low-dimensional embedding may separate otherwise similar T cells solely because they came from different runs. First inspect quality metrics, donors, chemistry, library batches and condition labels in the unintegrated space.

Apply a suitable integration method only where it matches the question, then assess mixing of matched cell types alongside preservation of known marker patterns and treatment-associated programs. For example, the emergence of interferon-responsive cells after treatment may be the primary biological finding. Forcing treated and untreated cells to overlap could remove that signal.

Integration outputs are often helpful for visualization, neighbor graphs and annotation, while differential expression may require original counts modeled with biological replicates. Perfectly confounded batches cannot be rescued just by adjusting embeddings. Report the design and validate the biological conclusion on independent samples.