JIND: joint integration and discrimination for automated single-cell annotation.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Manual cell-type annotation in transcriptomics is labor-intensive.
• Automated cell annotation requires effective integration of diverse datasets and removal of batch effects.
• Existing pipelines often treat integration and classification as independent steps.

Purpose of the Study:

• To develop a novel framework for automated single-cell annotation that integrates dataset integration and cell classification.
• To improve the accuracy and efficiency of cell-type identification in transcriptomic analysis.
• To address batch effects in a way that facilitates robust classification of new datasets.

Main Methods:

• Proposed JIND (joint integration and discrimination), a neural-network-based framework.
• Implemented an asymmetric alignment approach for mapping new cells to a learned latent space.
• Incorporated cell-type-specific confidence thresholds for reliable classification.

Main Results:

• JIND demonstrated superior accuracy over existing pipelines on several batched datasets.
• The use of confidence thresholds resulted in a smaller fraction of cells being rejected as unlabeled.
• Analysis of misclassified cells provided insights into potential data outliers or annotation errors.

Conclusions:

• JIND offers a more accurate and efficient solution for automated single-cell annotation.
• The joint integration and discrimination approach effectively handles batch effects.
• The framework's ability to identify unreliably classified cells enhances data interpretation.