Is It Biology or Just a Different Lab

When single-cell RNA-seq datasets from different labs, instruments, or processing batches are integrated, systematic technical variations (batch effects) confound biological signals. Current batch correction methods face a fundamental tradeoff: aggressive correction removes technical noise but also erases genuine biological differences between conditions, while gentle correction leaves technical artifacts that are misinterpreted as biology. A 2025 study found that batch correction methods that modify count data introduce significant artifacts in downstream differential expression analysis. No method can reliably distinguish technical from biological variation when the two are correlated — which they usually are, because patients from different conditions are typically processed in different batches.

Single-cell RNA-seq is the primary technology driving cell atlas projects (Human Cell Atlas), disease mechanism discovery, and drug target identification. The field is building integrated datasets comprising millions of cells from hundreds of labs. If batch effects masquerade as biological signal — or if genuine biological signal is erased during correction — the resulting cell atlases and disease models contain systematic errors. Downstream drug targets and biomarkers derived from confounded data may be artifacts, not biology.

Computational normalization methods (Harmony, Seurat integration, scVI, LIGER, scANVI) can align cells by type across batches, but alignment appearance does not guarantee biological fidelity. Adversarial learning approaches designed to remove batch effects simultaneously remove biological signals correlated with batch membership. Increasing model complexity (VAEs, adversarial networks) helps with integration visualization but can manufacture false biological signals. A 2025 finding showed that batch correction methods are "often poorly calibrated" — their confidence estimates do not match their actual accuracy. The fundamental information-theoretic barrier is that when batch and biology are confounded, no computational method can separate them without external information.

Experimental design solutions that prevent confounding rather than trying to correct it computationally: balanced batch assignment (randomizing samples across batches), spike-in reference samples processed across all batches to calibrate technical variation, and split-pool designs where the same sample is processed multiple times. Development of computational methods that provide explicit uncertainty estimates about which signals are batch-driven vs. biological. Single-cell platforms with inherently lower technical variability (reducing the problem at the source). The key insight is that this is primarily an experimental design problem, not a computational one.

A team could take a publicly available multi-batch scRNA-seq dataset (e.g., from the Human Cell Atlas), apply multiple batch correction methods, and systematically compare which "biological" findings are consistent across methods vs. which are method-specific artifacts. Building a benchmark using datasets with known ground truth (e.g., cell line mixtures processed in multiple batches) would be a stronger contribution. Bioinformatics, statistics, and genomics skills would be most relevant.