scEVOLVE: cell-type incremental annotation without forgetting for single-cell RNA-seq data.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-resolution analysis of cellular heterogeneity.
• Accurate cell-type annotation is crucial for downstream scRNA-seq data analysis.
• Existing automatic annotation methods struggle with continuous learning and expanding cell-type knowledge from data streams, leading to limitations in scalability and adaptability.

Purpose of the Study:

• To introduce a novel framework for cell-type incremental annotation to address the limitations of static annotation models.
• To develop a method that can continuously acquire knowledge from incoming data streams without catastrophic forgetting.
• To enhance the capacity of annotation systems for ever-increasing cell-type concepts.

Main Methods:

• Proposed scEVOLVE, an incremental annotation method utilizing contrastive sample replay and partition confidence maximization.
• Implemented a prototypical learning objective to mitigate cell-type imbalance, replacing cross-entropy.
• Introduced a cell-type decorrelation strategy to uniformly scatter feature representations for improved model training.

Main Results:

• Demonstrated scEVOLVE's ability to incrementally learn numerous cell types over extended periods on constructed benchmarks.
• Showcased superior performance compared to other strategies that exhibit rapid failure in incremental learning scenarios.
• Validated the framework's simplicity and ease of integration with deep softmax-based annotation methods.

Conclusions:

• scEVOLVE provides a robust solution for the challenging task of cell-type incremental annotation.
• The methodology effectively overcomes catastrophic forgetting and cell-type imbalance in continuous learning settings.
• This work represents the first end-to-end algorithm framework for practical incremental cell-type annotation in scRNA-seq data.