scPlantAnnotate: an accurate and robust transformer-based model for plant cell type annotation.

Area of Science:

• Plant biology
• Computational biology
• Genomics

Background:

• Accurate cell type annotation is crucial for plant single-cell RNA sequencing (scRNA-seq) but is hindered by a lack of specialized tools.
• Existing computational frameworks are often adapted from animal studies, leading to suboptimal performance on plant data and limiting plant cell atlas development.

Purpose of the Study:

• To develop and evaluate scPlantAnnotate, a Transformer-based framework specifically designed for plant scRNA-seq data.
• To benchmark scPlantAnnotate against current state-of-the-art deep learning and conventional methods across diverse plant species.

Main Methods:

• Trained species-specific scPlantAnnotate models using curated scRNA-seq datasets from Arabidopsis thaliana, Zea mays, Oryza sativa, and Glycine max.
• Evaluated scPlantAnnotate against leading methods using standard random-split and a stringent leave-one-dataset-out cross-validation approach to assess robustness.

Main Results:

• scPlantAnnotate consistently outperformed existing methods across all four species in standard evaluations.
• Demonstrated superior robustness to dataset shifts and heterogeneity, achieving the highest performance metrics in challenging leave-one-dataset-out scenarios for Arabidopsis thaliana.
• Results highlight improved handling of batch effects and dataset variations, critical for real-world plant scRNA-seq data.

Conclusions:

• scPlantAnnotate offers a plant-specific, Transformer-based solution for single-cell annotation with state-of-the-art performance.
• The framework enhances robustness to unseen datasets, addressing a key limitation in current plant scRNA-seq analysis.
• Facilitates the creation of comprehensive plant cell atlases and promotes wider adoption of single-cell genomics in plant science.