TL;DR
This paper introduces scCRT, a novel dimensionality reduction model designed to improve single-cell RNA sequencing (scRNA-seq) trajectory inference. By integrating cell-level and cluster-level contrastive learning, scCRT learns more accurate cell representations that better capture dynamic biological processes, outperforming existing methods on a large number of real and synthetic datasets.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) allows researchers to study dynamic biological processes like cell differentiation by analyzing gene expression at the individual cell level. A key task in this analysis is trajectory inference, which aims to reconstruct the developmental paths cells take. This process typically involves dimensionality reduction to simplify the high-dimensional gene expression data into a lower-dimensional space where trajectories can be visualized and analyzed. However, existing dimensionality reduction methods (like PCA or UMAP) are not specifically designed for trajectory inference and often fail to fully leverage prior biological information, such as known cell states. This can lead to inaccurate trajectory reconstructions. The paper addresses this by proposing a new dimensionality reduction approach that is optimized for trajectory inference and can effectively incorporate prior cell state information.
Methods
scCRT employs a two-pronged approach to dimensionality reduction, integrating two novel feature learning components:
- Cell-level Pairwise Module: This module focuses on learning accurate cell representations in a reduced-dimensionality space while preserving the relative positions of cells as they appear in the original, high-dimensional space. It achieves this by creating "positive" and "negative" pairs of cells based on their distances in the original space. The module then trains to minimize the distance between positive cell pairs and maximize the distance between negative cell pairs in the learned low-dimensional space. This helps maintain the local structure of the data.
- Cluster-level Contrastive Module: This module addresses the potential issue of cell representations becoming too scattered in the low-dimensional space, which can hinder trajectory inference. It leverages prior cell state information (e.g., cell type labels from upstream analysis) to aggregate similar cells. It creates augmented clusters from each cell cluster and uses contrastive learning to maximize the similarity between augmented clusters derived from the same cell state and minimize similarity between clusters from different states. This helps to group related cells together, preventing excessive dispersion.
The overall loss function for scCRT is a combination of the losses from these two modules. After learning the cell representations, a Minimum Spanning Tree (MST) is constructed based on these learned features to infer cell lineages, and principal curves are used to estimate pseudotime.
Datasets
The authors evaluated scCRT on a substantial collection of 135 datasets:
- 54 real scRNA-seq datasets
- 81 synthetic scRNA-seq datasets
These datasets were sourced from the dynverse project and varied in size (from dozens to over 10,000 cells) and complexity, including linear, nonlinear, and multiple branching trajectories. The synthetic datasets were generated using simulators like dyntoy, prosstt, and splattersimulators, providing ground truth for pseudotime and lineage. The real datasets were generated using different protocols (smart-seq, Drop-seq, 10x Chromium).
Key Insights and Ideas
- Dimensionality Reduction is Key for Trajectory Inference: The paper strongly advocates that improving the dimensionality reduction step is more critical for accurate trajectory inference than solely focusing on the inference algorithms themselves.
- Leveraging Prior Information is Crucial: Explicitly incorporating known cell states (like cell types) during the feature learning process significantly enhances the ability to reconstruct accurate trajectories.
- Balancing Local and Global Structure: The dual-module approach effectively balances preserving local cell-cell relationships (cell-level module) with grouping similar cell states (cluster-level module), preventing both over-dispersion and over-clustering.
- Contrastive Learning for Biological Data: The successful application of contrastive learning, a self-supervised technique, demonstrates its power in learning meaningful representations from complex biological data like scRNA-seq.
Key Results
- Superior Performance on Real Datasets: scCRT outperformed eight commonly used or state-of-the-art trajectory inference methods across three key metrics (HIM, F1-branches, F1-milestones) on 54 real datasets.
- Accurate Pseudotime Estimation: On 81 synthetic datasets, scCRT achieved the best average scores for trajectory inference (HIM, F1-branches, F1-milestones) and pseudotime estimation (Pearson Correlation Coefficient - PCC), indicating its ability to accurately determine developmental order.
- Visual Evidence of Accuracy: Visualizations on specific datasets (e.g., placenta trophoblast differentiation) showed scCRT accurately inferring multiple cell lineages and cell feature distributions that closely mirrored the ground truth, unlike some baselines that produced incorrect lineages or inaccurate cell feature placements.
- Ablation Study Confirms Module Importance: An ablation study showed that removing either the cell-level or cluster-level module (or using standard UMAP) resulted in diminished performance compared to the full scCRT model, highlighting the synergistic contribution of both modules.
Limitations and Future Work
The authors acknowledge a few limitations:
- Requirement for Cell Type Information: scCRT requires prior cell type information as input for the cluster-level module. If this information is noisy, it can negatively impact the learned features and subsequent trajectory inference.
- Discrete Outlier Clusters: The presence of discrete outlier clusters in the input data can affect trajectory prediction.
Future work suggested by the authors includes:
- Reducing Sensitivity to Noisy Labels: Developing methods to mitigate the negative impact of inaccurate prior information.
- Handling Outlier Clusters: Refining the approach to better handle discrete outlier clusters.
Summary updated