TL;DR
This paper introduces a more realistic challenge for single-cell RNA sequencing (scRNA-seq) analysis: simultaneously annotating known cell types and discovering novel ones within the same dataset. They propose scPOT, a novel framework leveraging optimal transport to effectively address this, achieving superior performance over existing methods.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to understand cellular heterogeneity. A crucial step in analyzing this data is cell type annotation, which assigns labels to individual cells. Historically, scRNA-seq analysis has operated under a "closed-set" assumption, where all cell types in a new dataset are expected to be present in a reference dataset. However, this is often not the case in real-world scenarios.
More recent approaches have tackled the "open-set" problem, where new datasets might contain cell types not seen in the reference, typically labeling these unknown cells as "unassigned." While an improvement, this still falls short. The authors argue that simply labeling novel cells as "unassigned" is insufficient for downstream analysis. They propose a more practical and challenging task: realistic cell type annotation and discovery. In this setting, cells from known cell types in the reference dataset are annotated with their correct labels, while cells from entirely new, unseen cell types are clustered and discovered. This dual objective is critical because existing methods that first annotate and then cluster often yield suboptimal results due to the dependency between these steps.
Methods
The core of the proposed solution, scPOT, is built upon a unified optimal transport (OT) framework. It addresses the dual task through two main components:
- OT-Based Prototypical Representation Learning for Novel Cell Type Discovery: This component focuses on discovering and clustering novel cell types. It employs a self-supervised learning paradigm using OT to learn cell representations. This encourages both global discrimination between clusters and local consistency among cells, aiming to uncover the intrinsic structure of the target data without relying solely on reference labels.
- OT-Based Prototypical Partial Alignment for Seen Cell Type Annotation: This part handles the annotation of known cell types. It uses an unbalanced OT approach to perform a partial alignment between target cells and reference prototypes. This method is designed to be robust to the presence of novel cell types in the target data, preventing misalignment.
Additionally, scPOT includes an innovative and practical solution for estimating the total number of cell types in the target data, a problem that is often assumed to be known beforehand. The overall model integrates these components with a denoising autoencoder for robust feature extraction.
Datasets
The authors conducted experiments on a variety of scRNA-seq datasets, including:
- Intra-data Annotation: 10 datasets sequenced from different organisms, with cell numbers ranging from 6462 to 110704 and cell type numbers from 9 to 45. For these experiments, they artificially split cell types into "seen" and "novel" categories.
- Cross-Data Annotation: 10 groups of datasets where batch effects are present between reference and target datasets.
They also constructed new evaluation benchmarks for this specific task, which are described as massive and highly imbalanced, reflecting real-world data complexities.
Key Insights and Ideas
- The "Realistic" Challenge: The paper convincingly argues that the traditional "open-set" annotation (labeling unknowns as "unassigned") is insufficient. The proposed "realistic" task of simultaneous annotation and discovery is a more accurate reflection of real-world scRNA-seq analysis needs.
- Optimal Transport as a Unifying Framework: The use of optimal transport is a clever way to bridge the gap between reference and target data, and to learn representations that capture both global and local structures.
- Unbalanced OT for Partial Alignment: The adaptation of unbalanced OT is crucial for handling the partial overlap between reference and target cell types, allowing for more accurate annotation of known cells.
- Automatic Cell Type Number Estimation: Providing a method to estimate the number of novel cell types is a significant practical contribution, as this number is often unknown and critical for downstream analysis.
- Robustness to Novel Cell Type Number:
scPOTdemonstrates remarkable robustness to variations in the number of novel cell types, a key advantage over other methods.
Key Results
scPOTconsistently achieved superior performance across various accuracy metrics (seen cell type accuracy, novel cell type clustering accuracy, and overall accuracy) in both intra-data and cross-data settings.- In intra-data experiments,
scPOToutperformed baselines like MARS, scNym, and scArches, which either provided sub-optimal results or struggled with novel cell type clustering. - In cross-data experiments,
scPOTshowed minimal performance degradation, indicating its robustness to batch effects. - Ablation studies confirmed the importance of each component of
scPOT, with the novel cell type discovery module (Lpro) having the most significant impact. - The proposed cell type number estimation method was validated, showing that clustering accuracy is maximized when the estimated number of cell types matches the true number.
Limitations and Future Work
- Computational Cost: While the paper mentions using PyTorch and GPUs, the computational complexity of optimal transport can be a limitation for extremely large datasets. The paper does not explicitly detail the computational time or scalability beyond the reported implementation details.
- Hyperparameter Sensitivity: The paper notes that parameters like $\epsilon$ and $\kappa$ in OT are set differently across datasets, suggesting some sensitivity. The authors also mention that the weight hyper-parameters $\lambda$ and $\gamma$ for the loss terms are set to 1.0, implying further tuning might be beneficial.
- Future Work: The paper implicitly suggests future work by highlighting the novelty of the task and the robustness of
scPOT. Further exploration of more complex batch effect scenarios, integration with other modalities, and development of even more efficient OT algorithms could be future directions.
Summary updated