Dimensionality reduction and data integration for scRNA-seq data based on integrative hierarchical Poisson factorisation

Thomas Wong, Mauricio BarahonabioRxiv 2021 · DOI

TL;DR

This paper introduces Integrative Hierarchical Poisson Factorisation (IHPF), a new method for analyzing single-cell RNA sequencing (scRNA-seq) data that excels at integrating datasets from different sources (batches) while reducing dimensionality. IHPF produces interpretable "latent factors" that help in clustering cells and understanding gene relationships, even when batch effects are present.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) generates massive datasets where each cell is described by the expression levels of thousands of genes. This high dimensionality, combined with the inherent sparsity and noise in the data, makes it challenging to extract meaningful biological insights using traditional dimensionality reduction techniques like PCA or NMF. A further complication arises when integrating data from different experiments or labs (batches), as technical variations (batch effects) can obscure true biological differences. Existing methods often struggle to disentangle these technical artifacts from biological signals, leading to inaccurate cell clustering and interpretation.

Methods

The core contribution is Integrative Hierarchical Poisson Factorisation (IHPF). IHPF is an extension of Hierarchical Poisson Factorisation (HPF) designed to handle data integration. It models scRNA-seq data as a Poisson distribution, where the observed gene counts are decomposed into:

  • Cell scores ($\Theta$): Representing how much each cell expresses each latent factor.
  • Batch-specific gene scores ($\Delta^{(s)}$): Capturing gene expression patterns unique to a particular batch.
  • Shared gene scores (B): Representing gene expression patterns common across all batches.

A key innovation is the introduction of a noise ratio hyper-parameter ($\alpha$). This parameter allows IHPF to adaptively balance the influence of batch-specific effects versus shared biological signals, effectively controlling the degree of data integration. The model uses a hierarchical Gamma prior structure, which has been shown to be more flexible for sparse, zero-inflated data like scRNA-seq. The inference of the model parameters is performed using mean-field variational inference.

Datasets

The authors used 16 publicly available scRNA-seq datasets from the literature. These datasets were combined to create three distinct data integration scenarios (Scenarios A, B, and C) designed to represent different levels of alignment between cell types and experimental batches.

  • Scenario A: Datasets from the same technology with similar cell types and good overlap.
  • Scenario B: Datasets from different technologies with diverse cell types and varying overlap.
  • Scenario C: Datasets from the same technology with distinct cell types, where cell types are highly concentrated within specific batches.

Key Insights and Ideas

  • Adaptive Batch Correction: IHPF's noise ratio hyper-parameter is crucial for effectively integrating data. It allows the method to adapt to different scenarios, either performing strong integration when needed (Scenario B) or minimal integration when batches are similar or cell types are well-aligned (Scenarios A and C).
  • Dual Block Structure: IHPF extracts latent factors that exhibit a "dual block structure" in both the cell and gene spaces. This means that the cell scores group cells by type, and the gene scores identify groups of genes associated with those cell types, leading to enhanced biological interpretability.
  • Interpretability over Complexity: Unlike some deep learning or complex multi-step pipelines, IHPF achieves strong performance with a relatively simple model, making its outputs more interpretable.
  • Robustness to Hyperparameters: The performance of IHPF is shown to be robust to changes in its cell and gene scale hyper-parameters, simplifying its application.

Key Results

  • Superior Clustering: IHPF consistently achieved the highest Silhouette Coefficients (SC) for cell clustering across all three scenarios, indicating well-separated cell clusters. For example, in Scenario A, IHPF achieved an SC of 0.963, significantly outperforming other methods like Scanorama (0.447) and INMF (0.458).
  • Accurate Cell Type Identification: The Adjusted Mutual Information (AMI) scores for cell clustering versus true cell types showed that IHPF performed best, consistently outperforming INMF, BBKNN, and scVI.
  • Effective Batch Effect Mitigation: IHPF demonstrated a strong ability to reduce batch effects while preserving biological signals, as evidenced by the low AMI scores between cell clusterings and batch labels, especially when compared to the AMI of true cell types versus batches.
  • Interpretable Gene Clusters: IHPF successfully identified biologically meaningful gene clusters, as confirmed by high Biological Homogeneity Index (BHI) scores and low p-values, indicating that the gene groupings correspond to coherent biological functions.
  • Visualisation Benefits: TSNE plots generated from IHPF latent factors showed better separation of cell groups and were less confounded by batch labels compared to plots generated from PCA components.

Limitations and Future Work

  • Categorical Batch Effects: The current IHPF model is designed to handle categorical batch effects. The authors suggest that future work should extend the model to accommodate continuous factors, such as population effects (e.g., age, sex).
  • Sophisticated Downstream Analysis: The latent factors extracted by IHPF could serve as input for more advanced clustering and classification methods, such as variational autoencoders or graph-based multi-scale clustering.
  • Further Model Extensions: The authors mention that IHPF can only correct differences between data sets that are categorical. Extending it to handle continuous factors is a key direction.

IHPF is made available as a Python package.

Summary updated

Abstract

Abstract Single-cell RNA sequencing (scRNA-seq) data sets consist of high-dimensional, sparse and noisy feature vectors, and pose a challenge for classic methods for dimensionality reduction. Such problems are compounded when dealing with composite data sets formed by different batches. We introduce Integrative Hierarchical Poisson Factorisation (IHPF), an extension of HPF that makes use of a noise ratio hyper-parameter to tune the variability attributed to batches vs . biological sources (cell phenotypes). We exemplify the application of IHPF under different data integration scenarios with varying alignments of batches and cell diversity, and show that IHPF produces latent factors that can be advantageously applied for cell clustering and visualisation. In addition, the extracted factors have a dual block structure in both cell and gene spaces with enhanced biological interpretability.
BibTeX
@misc{wong2021dimensionality,
  author = {Thomas Wong and Mauricio Barahona},
  title = {{Dimensionality reduction and data integration for scRNA-seq data based on integrative hierarchical Poisson factorisation}},
  year = {2021},
  doi = {10.1101/2021.07.08.451664},
  howpublished = {bioRxiv},
}
Figure