Distribution-Independent Cell Type Identification for Single-Cell RNA-seq Data

Yuyao Zhai, Liang Chen, Minghua DengIJCAI 2024

Abstract

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular diversity, yet the manual annotation of cell types remains a significant bottleneck. Automated methods often struggle with "open-world" scenarios—where target datasets contain novel cell types not present in reference data—and the inherent "long-tail" distribution of biological samples. We present scDET (Single-Cell Distribution-Independent Universal Cell Type Identification), a deep learning framework that integrates contrastive learning with a pseudo-labeling branch. By employing dynamic distribution estimation and post-hoc logit adjustment, scDET achieves state-of-the-art performance in identifying both known and novel cell types across diverse datasets.

Introduction

The primary challenge in automated scRNA-seq annotation is the discrepancy between reference and target datasets. Specifically, existing tools often fail when encountering novel cell types, either misclassifying them or failing to distinguish between different unknown populations. Furthermore, the extreme class imbalance (long-tail distribution) in biological data often leads models to overlook rare but critical cell types. scDET addresses these hurdles by decoupling the learning process into two mutually reinforcing branches that account for distribution shifts and novel populations.

Methods: The scDET Architecture

The scDET framework utilizes a "dual-consultative" approach built upon a shared denoising autoencoder. This architecture feeds into two specialized components:

  1. Contrastive-Learning Branch: This branch performs unsupervised grouping of similar cells. It utilizes Dynamic Distribution Estimation via k-means clustering to approximate the true proportions of cell types in the target data, facilitating the discovery of novel clusters.
  2. Pseudo-Labeling Branch: Acting as a classifier, this branch uses distribution estimates from the contrastive branch to regularize its predictions. It incorporates Post-hoc Logit Adjustment to mitigate bias toward common cell types, ensuring robust performance on rare populations.
The scDET architecture features a shared autoencoder feeding into contrastive learning and pseudo-labeling branches for distribution-independent identification.

The interaction between these branches is iterative: the contrastive branch guides the classifier with distribution estimates, while the classifier provides soft pseudo-labels to help the contrastive branch refine its feature space.

Experimental Evaluation and Results

The model was rigorously tested on 10 intra-data and 10 inter-data (cross-study) scRNA-seq datasets, simulating open-world conditions by withholding 50% of cell types during training.

Superior Accuracy and Clustering

scDET consistently outperformed existing methods such as scGAD and scNym. For instance, on the Quake 10x dataset, scDET achieved an overall accuracy of approximately 86.7%, surpassing scGAD's 83.7%. A key strength of scDET is its ability to produce distinct, well-separated clusters for novel cell types.

UMAP visualizations comparing scDET with other methods (MARS, scNym, scGAD). scDET demonstrates tighter and more distinct clustering for various pancreatic cell types.

As shown in the UMAP visualizations, scDET creates much tighter and more biologically meaningful clusters for novel types compared to competitors, which often exhibit significant overlap or blurring between different populations.

Robustness to Distribution Shifts

The model maintains high performance even as the number of novel cell types increases or the amount of labeled reference data decreases. This robustness stems from its "distribution-independent" nature—re-estimating the dataset distribution periodically prevents the model from being biased by the reference data's specific composition.

Discussion and Future Work

While scDET represents a significant advancement, some limitations remain. The model currently requires a prior estimate of the total number of cell types (seen + novel). Integrating an automated estimation step for the total cluster count would enhance its autonomy. Additionally, while scDET handles batch effects effectively, extreme technical variations between sequencing technologies continue to pose challenges. Future work will focus on further optimizing the computational efficiency of the dual-branch architecture and improving its resilience to diverse technical noise.

Summary updated

BibTeX
@inproceedings{zhai2024distribution,
  author = {Yuyao Zhai and Liang Chen and Minghua Deng},
  title = {{Distribution-Independent Cell Type Identification for Single-Cell RNA-seq Data}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure