In single-cell genomics, the data we collect is a mixture of two opposing forces: the genuine biological state of a cell and the technical "noise" introduced by the experiment itself. This noise, commonly known as a batch effect, arises from differences in laboratory protocols, sequencing hardware, or even the time of day a sample was processed.
Often, these technical variations are so large that they dwarf the biological signal. If you sequence the same tissue in two different labs, the cells will often cluster by "Lab A" and "Lab B" rather than by their actual cell types. To build a unified atlas of human biology, researchers must find ways to "disentangle" these factors, stripping away the experimental context to reveal the underlying cellular identity.
The Geometry of Integration
The core challenge of data integration is finding a shared "alignment space" where cells of the same type from different experiments overlap, while cells of different types remain distinct.
Early approaches relied on identifying Mutual Nearest Neighbors (MNN)—finding pairs of cells that are each other's closest matches across different datasets. However, unsupervised MNN can sometimes "mis-match" similar but distinct cell types. To solve this, Yuchen Yang and colleagues developed SMNN and its successor iSMNN, which use cell-type information to guide the alignment, ensuring that a T-cell in one batch is only ever mapped to a T-cell in another.
Adversarial Disentanglement
Modern methods often treat batch correction as a game of hide-and-seek played by neural networks. In an adversarial framework, one part of the model (the encoder) tries to represent the cell's data in a compressed latent space, while another part (the discriminator) tries to guess which batch that cell came from.
If the discriminator can’t tell the batches apart, the encoder has successfully learned a "batch-invariant" representation. Jia Zhao and the team behind Portal used this adversarial domain translation to integrate millions of cells across species. Similarly, ResPAN uses a Wasserstein Generative Adversarial Network (WGAN) to merge datasets while preserving delicate biological signals that simpler methods might "over-correct" or smooth away.
From Cleaning to Counterfactuals
The most recent shift in the field moves beyond simply "cleaning" the data toward modeling the experimental process itself. Instead of viewing the batch as a nuisance to be deleted, researchers like Pierre Boyeau and Michael Jordan use deep generative models like MrVI to ask counterfactual questions: "How would this specific cell from Patient A look if it had been treated with Drug B?"
By using a hierarchical latent variable approach, these models can separate sample-level heterogeneity (like a patient's response to a drug) from the technical noise of the sequencing run. This allows for "true online" integration, as seen in SCALEX by Lei Xiong, where new data can be projected into an existing reference atlas without needing to retrain the entire model.
The Limits of Foundation Models
As the field moves toward "foundation models" trained on hundreds of millions of cells—such as Tahoe-100M—the stakes for disentanglement have never been higher. If a model is trained on data where batch effects are not properly handled, it may learn to represent "experimental artifacts" rather than "biological laws."
Indeed, Katarzyna Kedzierska and Alex Xijie Lu have shown that massive zero-shot models often struggle to outperform simple statistical methods because they fail to distinguish technical noise from true biology. This has led to a renewed focus on "semi-mechanistic" models, as argued by Luka Kovacevic in No Foundations without Foundations, which suggest that we must incorporate the physical sequence of experimental actions into our AI architectures to truly understand regulatory biology.