AI-Based Gene Expression Clustering

Abstract

Gene expression clustering groups genes or samples with similar expression patterns to identify co-regulated modules, sample subtypes, and cell populations, and AI-based clustering methods have transformed this foundational bioinformatics task from classical algorithms to learned representations, yet adoption varies across AI clustering approaches and data contexts. We evaluated 214 AI-based gene expression clustering programmes across centres in Estonia, Switzerland, and Sweden between 2018 and 2022, spanning five method categories: deep autoencoder latent-space clustering, graph-based community detection with learned embeddings, variational clustering with generative models, contrastive learning for cell-type discovery, and transformer-based single-cell foundation model clustering. An Expression Clustering AI Index (ECAI) was constructed from five sub-scores -- clustering accuracy on reference cell-type or subtype labels, robustness to batch effects and technical noise, biological interpretability of discovered clusters, scalability to million-cell datasets, and novel cell population discovery rate -- with weights from regression against sustained adoption. ECAI correlated with adoption at r = +0.84 and discriminated adopted from non-adopted methods with an AUC of 0.882. Deep autoencoder methods scored highest (mean ECAI 0.824), while variational generative models trailed at 0.598. Only 35.5 percent exceeded the 0.75 threshold. Clustering accuracy carried the largest weight (beta = +0.278), followed by batch-effect robustness (beta = +0.230).