Enhancement of network architecture alignment in comparative single-cell studies

Clemens Schächter, Martin Treppner, Maren Hackenberg, Hanne Raum, Joschka Boedecker, Harald BinderbioRxiv 2024 · DOI

TL;DR

This paper introduces "scSpecies," a novel deep learning method that aligns single-cell RNA sequencing (scRNA-seq) data across different species. By mapping diverse animal model data to a unified latent space, scSpecies enables more accurate comparisons of human and animal biology, facilitating better translation of research findings.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized biological research, allowing us to study cellular heterogeneity at an unprecedented resolution. Animal models are crucial for understanding human diseases, but translating findings from these models to humans is challenging due to biological differences between species. Specifically, scRNA-seq datasets from different species often have varying gene sets (due to lack of orthologs) and different gene expression patterns even for homologous genes. This makes it difficult to directly compare or transfer knowledge between these datasets. Existing methods like "architecture surgery" can align network architectures but struggle with cross-species comparisons, particularly when dealing with missing genes or significant biological differences.

Methods

The core of scSpecies is a deep learning approach that combines variational autoencoders (specifically, the scVI model) with a data-level nearest neighbor search. Here's a breakdown:

  • Architecture Transfer: scSpecies pre-trains a scVI model on a "context" dataset (e.g., mouse data). Then, the encoder layers of this pre-trained model are transferred to a new scVI model for the "target" dataset (e.g., human data). This allows the model to leverage learned representations from the context species.
  • Mid-Level Feature Alignment: Instead of aligning at the raw data level, scSpecies focuses on aligning "mid-level features" learned by the neural network's intermediate layers. This is inspired by computer vision techniques and aims to capture more abstract biological properties less affected by species-specific noise.
  • Nearest Neighbor Search: To guide the alignment of these intermediate representations, scSpecies performs a nearest neighbor search on homologous genes between the context and target datasets. This helps identify similar cells across species and incentivizes the model to map biologically related cells into similar regions of the latent space.
  • Conditional Variational Autoencoder: The underlying scVI model uses a conditional variational autoencoder framework to learn a latent representation that captures biological variability while accounting for technical factors like batch effects.

Datasets

The authors evaluated scSpecies on several publicly available scRNA-seq datasets:

  • Liver: Datasets from mice (healthy and non-alcoholic fatty liver disease) and humans.
  • White Adipose Tissue: Datasets from mice and humans.
  • Glioblastoma: Datasets from mice (immune response) and humans.
  • Multi-species Liver: Datasets from mice, humans, pigs, monkeys, chickens, and hamsters.

The paper does not introduce new benchmarks but rather uses existing datasets to evaluate the performance of scSpecies against established methods like scArches and a data-level k-nearest neighbor classifier.

Key Insights and Ideas

  • Mid-level feature alignment is crucial: Aligning intermediate neural network layers, rather than just input layers, captures more robust biological signals and is less sensitive to species-specific variations.
  • Nearest neighbor search guides alignment: Using homologous genes to find similar cells across species provides a powerful signal to guide the latent space alignment, improving cell type transfer accuracy.
  • Robustness to gene set differences: scSpecies performs well even with limited shared genes or when datasets have significant differences in observed gene sets, a common challenge in cross-species comparisons.
  • Bridging the evolutionary gap: The method effectively maps animal model data to a human context, providing valuable insights into how well animal models represent human biology.

Key Results

  • Improved Cell Label Transfer: scSpecies significantly improved cell label transfer accuracy compared to data-level nearest neighbor search and existing architecture surgery approaches. For example, broad cell type labels achieved accuracies of 92% (liver), 82% (adipose tissue), and 80% (glioblastoma).
  • Biologically Meaningful Alignment: Visualizations (UMAP plots) showed biologically meaningful alignment of cells from different species, with cell types without direct counterparts forming distinct clusters or aligning with related cell types.
  • Robust Performance: The method demonstrated robust performance even with small datasets and large differences in gene sets.
  • Cross-species Liver Atlas: scSpecies successfully aligned liver cell datasets from multiple species (mice, humans, pigs, monkeys, chickens, hamsters), creating a unified landscape.
  • Gene Expression Comparison: The method enabled detailed comparison of gene expression profiles between species, identifying genes with systematic differences and commonalities.

Limitations and Future Work

  • Cell types unique to target dataset: scSpecies may align cell types unique to the target dataset with biologically close cell types in the context dataset, potentially leading to misinterpretation of species-specific populations.
  • Context dataset completeness: If a cell type is not present in the context dataset, it cannot be aligned across species. The authors suggest that the context dataset should ideally encompass all suspected cell types.
  • Future Directions:
    • Adaptation to other CVAE-based models.
  • Extension to handle multimodal datasets (e.g., integrating scRNA-seq with protein expression data).
  • Development of direct metrics to identify unique cell types and detect misclassifications due to noisy nearest neighbor search.

Summary updated

Abstract

1 Abstract Animal data can provide meaningful context for human gene expression at the single-cell level. This context can improve cell-type or cell-state detection and clarify how well the animal models human biological processes. To achieve this, we propose a deep learning approach that identifies a unified latent space to map complex patterns between datasets. Specifically, we combine variational autoencoders with a data-level nearest neighbor search to align neural network architectures across species. We visualize commonalities by mapping cell samples into the latent space. The aligned latent representation facilitates information transfer in applications of liver, white adipose tissue, and glioblastoma cells from various animal models. We also identify genes that exhibit systematic differences and commonalities between species. The results are robust for small datasets and with large differences in the observed gene sets. Thus, we reliably uncover and exploit similarities between species to provide context for human single-cell data.
BibTeX
@misc{schachter2024enhancement,
  author = {Clemens Schächter and Martin Treppner and Maren Hackenberg and Hanne Raum and Joschka Boedecker and Harald Binder},
  title = {{Enhancement of network architecture alignment in comparative single-cell studies}},
  year = {2024},
  doi = {10.1101/2024.08.30.608255},
  howpublished = {bioRxiv},
}
Figure