Robust Cell Type Annotation in Single-Cell RNA-Seq via Contrastive Domain Adaptation

Abstract

Single-cell RNA sequencing (scRNA-seq) has revolutionized the understanding of cellular heterogeneity by allowing transcriptional profiling at single-cell resolution. However, the automated annotation of cell types across different datasets remains a formidable challenge due to substantial batch effects—systematic technical variations arising from differences in sequencing protocols, capture platforms, and donor variability. These non-biological variances often confound standard supervised learning algorithms, leading to poor generalization on unseen target data. This paper introduces a novel framework, Contrastive Domain Adaptation for Single-Cell (CDA-SC), designed to robustly transfer knowledge from labeled source atlases to unlabeled target datasets. By integrating domain-adversarial training with supervised contrastive learning, CDA-SC aligns the feature distributions of source and target domains while explicitly enforcing intra-class compactness and inter-class separability in the latent space. We demonstrate that this dual-objective approach effectively mitigates batch effects and prevents the misalignment of biologically distinct but transcriptionally similar cell types. Extensive experiments on multiple cross-protocol benchmarks indicate that CDA-SC outperforms state-of-the-art baselines in classification accuracy and macro F1-scores, providing a scalable and reliable solution for automated cell type identification in large-scale integrative studies.