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

A Label-Free Multi-Metric Pipeline for Benchmarking Single-Cell RNA-Sequencing Clustering and Testing the Reproducibility of Cell-Type Heterogeneity

A discovered sub-population from single-cell transcriptomic data is only meaningful if it is reproducible, yet clustering is usually done with one method on one embedding and rarely tested. We present a label- free, multi-metric pipeline that reframes clustering as an auditable, methods-blind decision and separates two notions of stability that are commonly conflated: reproducibility under cell resampling (bootstrap) and reproducibility under re-embedding (retraining the representation). The pipeline evaluates seven clustering configurations across cluster counts using five non-redundant quality metrics. As a whole- dataset control on a mouse retinal atlas, it recovers an eight-cell-type annotation at 96.3% accuracy (adjusted Rand index, ARI = 0.91) without labels. We then validate the discovery mode on two cell types with opposite ground truth. On bipolar cells, which have well-established subtypes, the pipeline accepts the sub-structure: across-embedding reproducibility rises with cluster number to a high plateau (mean pairwise ARI ∼0.93 near the ∼15 known bipolar subtypes), with quality metrics improving in parallel. On rod photoreceptors, treated as homogeneous, it rejects over-clustering: the metric-selected partition passes a bootstrap-stability check but is not reproducible when the embedding is retrained (mean pairwise ARI = 0.69), and the metrics do not improve with cluster number. On synthetic data, the test recovers real structure down to a 5% subpopulation while rejecting null data (high sensitivity and specificity). Bootstrap stability alone is therefore insufficient evidence for sub-population; the across-embedding test discriminates real sub-structure from over-clustering and applies to any cell type as a reproducible alternative to single-method, single-embedding clustering.

iDCF: Interpretable deconvolution of cell fractions via biologically-informed deep learning using scRNA-seq data.

Precise resolution of cellular heterogeneity within complex tissues is fundamental to deciphering disease etiologies from bulk transcriptomic profiles. While computational deconvolution offers a scalable alternative, current deep learning methods predominantly operate as "black boxes," neglecting the structural constraints of biological laws. This reliance on purely data-driven feature extraction often yields biologically incoherent predictions and limited mechanistic interpretability. iDCF (Interpretable Deconvolution of Cell Fractions) is a novel framework that enforces biological topology onto deep neural networks. The iDCF architecture employs a dual-stream design, synergizing a standard deep network with a knowledge-based sparse neural network (KSNN) explicitly masked by pathway definitions and protein-protein interaction (PPI) networks. In comprehensive benchmarks, iDCF achieves top-tier performance, consistently ranking among state-of-the-art methods in accuracy and robustness. iDCF integrates the SHapley Additive exPlanations (SHAP) framework, bridging the gap between computational inference and biological intuition. The model's decision logic is governed by established biological mechanisms rather than spurious statistical correlations, validating its reliability. Validations across clinical contexts, including Alzheimer's disease, ovarian cancer, and diabetes, demonstrate iDCF's ability to recover disease-relevant cellular dynamics. iDCF offers a high-performance, interpretable, and biologically grounded tool for deconvolving cell-type proportions, facilitating deeper insights into tissue heterogeneity in health and disease.

MOFUN-CCC: A Multi-omics Intermediate Fusion Network for Digital White Blood Cell Count Prediction.

A novel multi-modal deep learning model with intermediate fusion: multi-omics fusion neural network- computational cell counting (MOFUN-CCC) designed to predict absolute cell counts directly by integrating gene expression and DNA methylation data within a supervised framework, assuming that the underlying true cell components are shared across the two omics data.

Cross-modal mapping of cancer stem-like cell plasticity using deep learning

Abstract Cancer stem-like cells (CSCs) play a pivotal role in driving tumor heterogeneity, therapeutic resistance, and disease progression. Despite the power of single-cell RNA sequencing (scRNA-seq) to resolve intratumoral hierarchies, there remains a need for robust, scalable tools to consistently profile CSCs across both single-cell and bulk transcriptomic data. To address this, we developed ACSCeND—a unified, machine learning–based framework that enables high-resolution CSC state classification and tissue-level deconvolution. ACSCeND comprises (i) a supervised classifier trained on curated scRNA-seq datasets to assign cells into pluripotent-like, multipotent-like, or unipotent-like states, and (ii) an attention-guided autoencoder that deconvolves CSC subtype proportions from bulk RNA sequencing data. Compared to existing tissue deconvolution tools, ACSCeND achieves superior performance, with higher accuracy across synthetic and real-world samples. Applied to over 25 000 tumor profiles from The Cancer Genome Atlas (TCGA), PREdiction of Clinical Outcomes from Genomics (PRECOG), tumor-relapse, and checkpoint inhibitor studies, ACSCeND reveals that CSC abundance strongly correlates with poor disease-free survival and reduced immunotherapy efficacy. Moreover, it uncovers distinct CSC-state-specific molecular programs, offering insights into CSC-driven heterogeneity and tumor evolution. The model also recapitulates known developmental hierarchies in noncancerous tissues, supporting its broader biological relevance. By integrating single-cell precision with bulk-level applicability, ACSCeND offers a robust, interpretable approach to profiling CSC dynamics and establishes CSC state as a clinically meaningful, pan-cancer biomarker for guiding stemness-informed therapies. ACSCeND is available as a python package (through pip) at https://pypi.org/project/ACSCeND/.

scLGGCL: Label-Guided Graph Contrastive Learning for Single-Cell Fusion Clustering.

scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating

Deep clustering models for single-cell RNA sequencing often assign cells through latent or centroid-based mechanisms that are difficult to inspect. We introduce scDNM-VAE (single-cell Dendritic Neuron Model Variational Autoencoder), a deep clustering framework that combines a variational autoencoder with a dendritic neuron-inspired head. Cluster assignments are governed by learnable signed synaptic weights and thresholds: the weight sign determines the direction of a gate's response to a latent coordinate, its magnitude controls steepness, and the weight-threshold pair determines the transition location. The trained clustering function can therefore be inspected directly without fitting a post-hoc explanation model. We benchmark scDNM-VAE on four datasets spanning immune, cortical, cardiac, and hematopoietic cells against scVI followed by KMeans and an MLP-DEC ablation. scDNM-VAE performs better than scVI on PBMC3k, comparably on the Human Heart Cell Atlas and Paul15, and worse on Zeisel, while producing biologically coherent marker-gene signatures. Ablating each cluster's three highest-magnitude synaptic dimensions causes numerically greater reassignment than random-dimension ablation across all datasets, but the margins are modest and negligible on Zeisel. These results show that signed dendritic gating supports competitive clustering with a parameter-inspectable decision function, while indicating that decision-relevant information is distributed across the latent space.