CSLAN: Cross-Species Latent Alignment Network for Trauma-Related Cell-Type Classification

Rui Wu, Alan Shi, Yushi Liu, Alexandre Duprey, Ting Chen, Rui Li, Shuang Cao, Jiajie Shi

Keywords: Transfer Learning, Bioinformatics, Domain Adaptation, Latent Space Alignment, Few-Shot Learning, Cross-Species Analysis

Abstract: Transfer learning across domains with mismatched and non-mappable feature spaces is a fundamental challenge in machine learning. Existing methods often rely on brittle feature-mapping or risk catastrophic forgetting during fine-tuning. To address this, we introduce the Cross-Species Latent Alignment Network (CSLAN), a novel framework for robust knowledge transfer. CSLAN pioneers a three-pronged approach: (1) we employ a sparse regression model for principled selection of informative features in each domain, reducing noise and dimensionality. (2) We pre-train an encoder-decoder on a comprehensive source domain (mouse). (3) We introduce a biologically-inspired asymmetric fine-tuning strategy, where the pre-trained decoder and latent processor—encapsulating conserved class definitions—are frozen. A new target-specific encoder (for human) is then trained from scratch to project its distinct feature space into this preserved, semantically structured latent space. On a challenging cross-species trauma-related single-cell classification task, CSLAN achieved 95.83% accuracy using only a few hundred labeled human cells, significantly outperforming standard baselines. Our work establishes a powerful paradigm for aligning mismatched domains, demonstrating that decoupling feature projection from a conserved decision manifold is key to effective transfer.

Submission Number: 4