Joint deep learning for batch effect removal and classification toward MALDI MS based metabolomics

Jingyang Niu, Jing Yang, Yuyu Guo, Kun Qian, Qian WangBMC Bioinformatics 2022 · DOI

TL;DR

This paper introduces a novel deep learning framework that simultaneously tackles batch effect removal and classification in metabolomics data, specifically for MALDI MS. By jointly learning to align data across batches and predict disease states, the method significantly improves diagnostic accuracy compared to existing approaches.

What Problem Does This Paper Address?

Metabolomics, the study of small molecules in biological systems, holds immense promise for understanding diseases and developing personalized treatments. However, a major hurdle in analyzing metabolomics data, especially from high-throughput technologies like MALDI MS, is the "batch effect." This refers to systematic variations in measurements that arise from differences in experimental conditions (e.g., different plates, reagents, or time points) and are unrelated to the actual biological differences of interest. These batch effects can obscure true biological signals, leading to inconsistent or inaccurate conclusions, particularly when trying to build predictive models that generalize across different experimental runs or datasets. Existing methods for batch effect correction often focus solely on removal, potentially at the cost of reducing the very biological differences needed for classification, or they make assumptions that don't always hold true in practice.

Methods

The authors propose an end-to-end deep learning framework designed for joint batch effect removal and classification. The core of their approach involves three interacting neural network modules:

  • Calibrator (Encoder): This module takes input data from different batches and maps them into a common latent space. It's designed to align the distributions of these batches, effectively removing batch-specific variations. Batch normalization is used in the first layer to aid convergence.
  • Reconstructor (Decoder): To ensure that the latent space captures meaningful biological information and doesn't lose crucial data, reconstructors are used. These modules take the latent representations and attempt to reconstruct the original input data, forming an encoder-decoder backbone for self-learning.
  • Discriminator: This module is trained on the source batch (with known labels) to classify samples. Crucially, it's trained to distinguish biological labels, not batch labels. Once trained, it's used to predict labels for the calibrated target batch data.

The framework is trained by minimizing a combined loss function that includes terms for calibrator loss (minimizing MMD between batch distributions in the latent space), reconstruction loss (ensuring accurate data recovery), and discriminator loss (binary cross-entropy for classification).

Datasets

The proposed method was evaluated on two datasets:

  1. Public CyTOF Dataset: A subset of publicly available mass cytometry (CyTOF) data was used for simulation experiments. This dataset involved Peripheral Blood Mononuclear Cells (PBMCs) from two patients, collected on different days, creating two distinct batches. Each sample had 25 features.
  2. Private MALDI MS Dataset: A newly collected dataset of serum samples from healthy controls (HCs) and patients with Systemic Lupus Erythematosus (SLE) was used. This dataset involved 598 subjects across three different MALDI MS target plates, serving as three distinct batches. Each sample had 814 features (m/z values).

Key Insights and Ideas

  • Joint Learning is Key: The paper argues that simultaneously addressing batch effect removal and classification, rather than treating them as separate steps, leads to superior performance. This allows the model to learn latent representations that are both batch-invariant and discriminative for the task at hand.
  • Latent Space Alignment for Classification: By forcing different batches into a common latent space, the model can then leverage a discriminator trained on one batch to effectively classify samples from another, even if they were processed under different conditions.
  • Reconstruction as a Regularizer: The inclusion of reconstructors is vital for ensuring that the learned latent space retains biologically relevant information, preventing the model from simply "forgetting" important features during batch correction.
  • Discriminator Focuses on Biological Labels: A critical aspect is that the discriminator is trained to distinguish between disease states (biological labels) rather than batch origins, which is crucial for accurate downstream prediction.

Key Results

  • CyTOF Simulation: The method demonstrated effective batch effect removal, as evidenced by reduced MMD values compared to other state-of-the-art methods. Visualization using t-SNE showed that after calibration, samples from different batches were well-entangled, and samples were grouped by class labels rather than batch. Classification performance also improved significantly.
  • MALDI MS Real-Case Study: The framework achieved the highest diagnostic accuracy on the private MALDI MS dataset, with an average increase of 5.1% to 7.9% in accuracy compared to other methods. For specific batch combinations, improvements reached as high as 13.6% for ACC, 9.7% for F-score, 15.3% for AUC, and 26.9% for MCC. The results were also comparable to or even exceeded in-batch classification performance, which serves as a benchmark for batch-effect-free data.

Limitations and Future Work

The authors acknowledge several limitations:

  • Hyperparameter Optimization: Achieving optimal performance for all compared algorithms might require exhaustive hyperparameter tuning.
  • Single-Source-Single-Target Scenario: The current framework is designed for a single source batch and a single target batch. Real-world scenarios often involve multiple batches.
  • Domain Knowledge Requirement: The approach is not entirely unsupervised, requiring domain knowledge (e.g., labels for the source batch).
  • Labeled Data Dependency: Achieving reasonable performance might necessitate a substantial amount of labeled data.

Future work could focus on adapting the trained model for multiple unlabeled target batches, utilizing multiple source batches for training, and exploring ways to reduce the reliance on labeled data.

Summary updated

BibTeX
@article{niu2022joint,
  author = {Jingyang Niu and Jing Yang and Yuyu Guo and Kun Qian and Qian Wang},
  title = {{Joint deep learning for batch effect removal and classification toward MALDI MS based metabolomics}},
  journal = {BMC Bioinformatics},
  year = {2022},
  doi = {10.1186/s12859-022-04758-z},
}
Figure