Knowledge-based inductive bias and domain adaptation for cell type annotation.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Cellular data analysis faces challenges with domain gaps across measurement batches.
• The effectiveness of cell annotation methods depends on their inductive biases and biological relevance.

Purpose of the Study:

• To develop a method that enhances cross-batch cell type annotation by integrating biological knowledge.
• To improve the generalizability and interpretability of cell annotation models.

Main Methods:

• Proposed KIDA (Knowledge-based Inductive bias and Domain Adaptation) method.
• Incorporated knowledge-based inductive bias to align model predictions with biological functional patterns.
• Utilized pseudo labels and self-knowledge distillation for domain adaptation in multi-batch query datasets.

Main Results:

• KIDA demonstrated accurate cross-batch cell type annotation in benchmark experiments.
• The knowledge-based inductive bias improved model generalization to unseen batches.
• Domain adaptation effectively reduced the distribution gap in query datasets.

Conclusions:

• KIDA offers a robust approach for accurate cell type annotation across diverse data batches.
• Aligning inductive biases with biological knowledge is crucial for effective computational biology tools.
• The method shows promise for advancing single-cell data analysis and interpretation.