Leveraging mutual information in Variational Autoencoders for improved dimensionality reduction of single-cell RNA sequencing data: The scInfoMaxVAE approach

Abstract

Single-cell RNA-seq (scRNA-seq) analysis demands representations that are robust to sparsity and technical noise. We present scInfoMaxVAE, a mutual-information–maximizing variational autoencoder with a zero-inflated count likelihood tailored for scRNA-seq, designed for dimensionality reduction and cell-type classification. We evaluated the model on 12 public scRNA-seq datasets spanning multiple tissues and platforms using a unified pipeline with cell- and gene-level quality control (minimum detected genes), library-size normalization, log-transform, and reference-based cell-type annotation. Against established methods (VASC, DREAM, scVI, scDeepCluster) and conventional embeddings (e.g., t-SNE, UMAP), scInfoMaxVAE delivered competitive clustering and structure preservation across all datasets; for representative cohorts, it achieved normalized mutual information (NMI) of 0.94, matching VASC (0.94) and exceeding t-SNE (0.66), with notable gains in homogeneity (0.89 vs. 0.58 for scVI) and adjusted Rand index (0.81 vs. 0.38 for scVI). Strengths include consistent performance across heterogeneous datasets and improved preservation of neighborhood structure, attributable to information-theoretic training and explicit modeling of zero inflation. Limitations observed in our study include sensitivity to hyperparameters and modest run-to-run variance, suggesting benefits from automated tuning and further large-scale validation. Overall, scInfoMaxVAE offers a robust, reproducible alternative for representation learning in scRNA-seq workflows.