Semisupervised adversarial neural networks for single-cell classification.

Area of Science:

• Single-cell genomics
• Computational biology
• Machine learning in biology

Background:

• Accurate cell identity annotation is crucial for single-cell genomics analysis.
• Current methods face challenges in transferring annotations across diverse datasets.
• Existing bottlenecks hinder the scalability and comparability of single-cell studies.

Purpose of the Study:

• To develop a novel computational method for robust cell identity transfer in single-cell genomics.
• To improve the accuracy and efficiency of cell annotation across experiments.
• To create a tool that leverages both labeled and unlabeled data for enhanced annotation.

Main Methods:

• Development of scNym, a semisupervised adversarial neural network.
• Utilizing labeled and unlabeled datasets for learning rich cell identity representations.
• Implementing annotation transfer across experiments with biological and technical variations.

Main Results:

• scNym demonstrates superior performance in cell identity transfer compared to existing methods.
• The model effectively handles biological and technical differences between datasets.
• scNym models can integrate information from multiple datasets to boost performance.
• Achieved high accuracy, model calibration, and interpretability using saliency methods.

Conclusions:

• scNym provides a powerful and accurate solution for cell identity annotation transfer.
• The method enhances the comparability and scalability of single-cell genomics analyses.
• scNym offers a robust, interpretable, and well-calibrated approach for computational biologists.