Sparse Autoencoders Enhance AI Interpretability in Biological Data

This paper investigates the phenomenon of "superposition" in neural networks, where distinct concepts are compressed into lower-dimensional latent spaces, particularly problematic in high-dimensional biological data. Superposition is known to hinder AI interpretability and, as this research demonstrates, corrupts the geometry of latent spaces. To address this, the authors utilized sparse autoencoders (SAEs) trained on over 100,000 multiplexed images of patient-derived Parkinson's disease and healthy neurons.The SAEs successfully resolved superposition, recovering geometric fidelity in the latent representations. This method bypasses the mathematical non-uniqueness often associated with feature attribution by focusing on interpretable latent representation analysis. By treating these geometrically purified representations as single-cell state vectors, the researchers adapted single-cell RNA sequencing (scRNA-seq) data analysis methodologies directly to the image domain.A key innovation is GW-map, which employs Gromov-Wasserstein optimal transport to align these image representations with authentic scRNA-seq data from scratch. This coupling enables the reconstruction of hierarchical neuronal pathology pathways, such as Calcium-AIS scaffold, without needing reference spatial transcriptomics. This establishes a scalable foundation for advancing spatial biology research.