BioBatchNet: A Dual-Encoder Framework for Robust Batch Effect Correction in Imaging Mass Cytometry

Haiping Liu, Shaojie Zhang, Shengzhong Mao, Qian Zhao, Yuxi Zhou, Andrew Gilmore, Mauricio A. Álvarez, Hongpeng ZhoubioRxiv 2025 · DOI

TL;DR

This paper introduces BioBatchNet, a novel dual-encoder framework that uses adversarial training to effectively disentangle biological signals from batch effects in single-cell data, particularly Imaging Mass Cytometry (IMC). It offers controllable batch correction and an integrated clustering method, outperforming existing approaches on multiple datasets.

What Problem Does This Paper Address?

Single-cell technologies like Imaging Mass Cytometry (IMC) are revolutionizing our understanding of biological systems by allowing high-resolution analysis of protein expression within tissues. However, a major hurdle in analyzing this data is the presence of "batch effects." These are systematic biases introduced during sample processing, experimental variations, or differences between labs, which can obscure true biological signals and lead to incorrect conclusions. Existing methods, often developed for scRNA-seq data, struggle with IMC data's unique characteristics and can either over-correct (removing valuable biological information) or under-correct (leaving residual batch effects). Furthermore, IMC data often contains complex cell populations that are imbalanced and overlap, making downstream analyses like clustering particularly challenging. This paper aims to address these limitations by providing a robust, controllable, and generalizable method for batch effect correction and improved clustering.

Methods

BioBatchNet employs a dual-encoder Variational Autoencoder (VAE) architecture. The core innovation lies in its adversarial training strategy to achieve disentanglement:

  • Bio-Encoder: This encoder learns biological latent representations. It's trained in an adversarial manner against a batch discriminator using a gradient reversal layer. The goal is for the Bio-encoder to generate representations that the discriminator cannot use to identify the batch source, thus preserving biological signals independent of batch.
  • Batch-Encoder: This encoder learns batch-specific latent representations. It's trained with a batch classifier to ensure it captures batch-related information.
  • Decoder: Reconstructs the original data using a combination of the biological and batch-specific latent representations.
  • Orthogonal Loss: An additional loss function is introduced to enforce independence between the biological and batch-specific latent representations.
  • Controllable Correction: The degree of batch effect correction can be precisely tuned by adjusting a weighting factor ($\lambda_1$) in the overall loss function.

In addition to batch correction, the paper introduces a Constrained Pairwise Clustering (CPC) method. This approach leverages weak prior knowledge (must-link and cannot-link constraints) to improve clustering performance, especially on imbalanced and overlapping datasets. It involves pre-training an autoencoder on the biological latent representations from BioBatchNet and then refining it using these pairwise constraints.

Datasets

The study evaluated BioBatchNet on a variety of datasets:

  • Imaging Mass Cytometry (IMC) Datasets:
    • IMMUcan_Cancer (4 patients, 4 batches)
    • Damond_Pancreas (1 dataset, 12 batches)
    • Hoch_Melanoma (1 dataset, 39 batches) - This dataset highlights the need for handling a large number of batches.
  • scRNA-seq Datasets:
    • Human pancreas (1 dataset, 8 batches)
    • Human immune (1 dataset, 10 batches)
    • Macaque retina (1 dataset, 4 batches)
    • Mouse brain (1 dataset, 2 batches)

The paper notes that for scRNA-seq data, a tailored loss function was developed to account for its distinct characteristics (discrete counts vs. continuous intensities in IMC).

Key Insights and Ideas

  • Adversarial Disentanglement for IMC: Applying adversarial training specifically for disentangling biological and batch signals in IMC data is a novel approach, addressing the unique challenges of this modality.
  • Controllable Correction is Crucial: The ability to precisely control the extent of batch effect correction is vital. Over-correction can be as detrimental as under-correction, and BioBatchNet's $\lambda_1$ parameter offers this fine-tuning capability.
  • Weak Supervision for Clustering: Using simple "must-link" and "cannot-link" constraints as weak supervision for clustering is a practical and effective way to improve accuracy, especially when full annotations are unavailable or unreliable.
  • Generalizability Across Modalities: Demonstrating strong performance on both IMC and scRNA-seq data highlights BioBatchNet's versatility and potential as a unified solution for single-cell data integration.

Key Results

  • Superior Batch Effect Correction: BioBatchNet consistently outperformed seven benchmarking methods on three IMC datasets, achieving higher batch effect correction scores. For example, on the IMMUcan_Cancer dataset, BioBatchNet achieved a score of 0.87, compared to Harmony (0.82) and scDREAMER (0.83).
  • Effective Biological Signal Preservation: The method also demonstrated robust biological signal conservation, balancing correction with the preservation of true biological variation.
  • Improved Clustering: The integrated CPC method showed significant improvements in clustering performance (ARI, NMI, ACC) compared to K-means, Leiden, and scDCC on IMC datasets.
  • Scalability: The Damond_Pancreas dataset, with over 250,000 cells, was used to evaluate scalability, where BioBatchNet showed competitive performance.
  • Generalization: BioBatchNet achieved competitive results across four scRNA-seq datasets, demonstrating its broad applicability.

Limitations and Future Work

  • Limitations: While BioBatchNet shows excellent performance, the paper acknowledges that the choice of weighting factor $\lambda_1$ for controlling correction requires careful tuning. The current CPC method relies on generating pairwise constraints, which, while cost-effective, still requires some manual effort or a separate strategy for constraint generation.
  • Future Work: The authors propose several exciting directions:
  • Integrating Spatial Information: Incorporating spatial context into BioBatchNet, potentially by combining its latent embeddings with spatial data for Graph Neural Networks, to uncover deeper biological insights beyond clustering.
  • Data Imputation and Generation: Exploring BioBatchNet's potential for other tasks like data imputation and data generation, further expanding its utility.

Summary updated

Abstract

Abstract Motivation Imaging Mass Cytometry (IMC) is a cutting-edge technology for analysing spatially resolved protein expression at the single-cell level. However, its downstream analyses are often hindered by batch effects, which introduce systematic biases and obscure true biological variations. Existing correction methods, largely developed for scRNA-seq data, struggle to achieve precise control, leading to either over-correction by removing critical biological information, or under-correction by leaving residual batch effects. Moreover, these methods face challenges in adapting to IMC data due to differences in data characteristics. Furthermore, IMC data often feature imbalanced and overlapping cell populations, complicating clustering and downstream analysis. These challenges underscore the need for a robust and controllable batch effect correction approach tailored to IMC data. Results We present BioBatchNet, a dual-encoder framework utilising adversarial training to explicitly disentangle batch-specific and biological signals. BioBatchNet enables controllable batch effect correction, effectively balancing correction with the preservation of biological variation. We evaluated BioBatchNet on three IMC datasets, where it outperformed seven benchmarking methods with robustness in both correction and biological signal conservation. Additionally, we developed a Constrained Pairwise Clustering (CPC) method, which employs constrained pairs to improve clustering performance, even in datasets with imbalanced and overlapping cell populations. To validate its generalisability, BioBatchNet was also applied to four scRNA-seq datasets, where it delivered competitive performance compared to eight typical methods. These results demonstrate BioBatchNet’s generalisability and robustness in correcting batch effects across diverse single-cell datasets and underscore its potential for large-scale biological analyses.
BibTeX
@misc{liu2025biobatchnet,
  author = {Haiping Liu and Shaojie Zhang and Shengzhong Mao and Qian Zhao and Yuxi Zhou and Andrew Gilmore and Mauricio A. Álvarez and Hongpeng Zhou},
  title = {{BioBatchNet: A Dual-Encoder Framework for Robust Batch Effect Correction in Imaging Mass Cytometry}},
  year = {2025},
  doi = {10.1101/2025.03.15.643447},
  howpublished = {bioRxiv},
}
Figure