When Does Deep Representation Learning Help Single-Cell Clustering? A Sensitivity-Aware Diagnostic Benchmark for Biomedical AI Pipelines

This paper investigates whether incorporating deep representation learning—such as autoencoders or variational autoencoders (VAEs)—into single-cell RNA sequencing (scRNA-seq) clustering pipelines provides a measurable performance benefit over traditional pipelines based on Principal Component Analysis (PCA). The authors argue that because deep models require significant hyperparameter tuning and compute resources, practitioners need a data-driven framework to decide when these complex models are actually necessary.

To address this, the authors benchmarked nine clustering pipelines across ten diverse scRNA-seq datasets. They employed a rigorous statistical protocol that included Optuna-based hyperparameter optimization, repeated-run robustness checks, and Sobol total-order sensitivity analysis to determine which hyperparameters most influence clustering outcomes.

The study found no single "best" method. While a contrastive autoencoder achieved the highest mean Adjusted Rand Index (ARI), statistical testing revealed that it did not significantly outperform the strongest classical baselines. Instead, the authors identified three reproducible performance regimes:

• Small Datasets: VAE variants excel here, as the latent-space regularization acts as an inductive bias that prevents overfitting to noise.
• Mid-Scale Datasets: Deep autoencoders perform best when the data contains complex batch effects or many fine-grained cell types, as they can better preserve neighborhood structure.
• Linear-Dominant Datasets: Classical PCA pipelines remain highly competitive when the primary biological variation is captured well by linear projections, making the additional compute cost of deep models unnecessary.

Furthermore, the Sobol sensitivity analysis revealed that learning rate and latent dimensionality are the primary drivers of performance variance. This suggests that practitioners with limited tuning budgets should prioritize these two parameters over others.

This work shifts the conversation in biomedical AI from seeking universal model superiority to promoting compute-conscious, dataset-aware pipeline selection. By demonstrating that simpler, faster methods are often sufficient, the authors provide a practical decision framework that helps researchers allocate limited computational resources more effectively while maintaining high-quality analytical results.