A Topology-Feature Disentangled Deep Clustering Method for scRNA-seq Based on Graph-to-MLP Distillation.

MOTIVATION: Graph neural networks (GNNs), owing to their ability to model topological relationships between cells, have gradually emerged as a mainstream approach in the clustering analysis of single-cell RNA sequencing (scRNA-seq) data. However, their practical application still faces the following challenges: (1) excessive smoothing undermines cellular heterogeneity; (2) pre-constructed graphs introduce end-to-end dependencies; (3) large-scale cellular graph computations are computationally intensive. RESULTS: To fully utilise the expression profiles and topological information contained within scRNA-seq data and overcome these limitations, this paper proposes a Graph-to-MLP (G2M) knowledge distillation method incorporating AdaBoost principles, and builds upon this to develop a graph-disentangled single-cell clustering framework-scAdaG2M. This framework utilises a GNN teacher model to learn higher-order topological relationships and local neighbourhood structures between cells on a cellular graph, whilst multiple lightweight MLP student models perform feature learning in an ungraphed setting. By combining this with an AdaBoost-style dynamic reweighting strategy, the framework enhances the model's ability to characterise cells that are difficult to distinguish, cells with blurred boundaries, and potentially rare cell populations, thereby mitigating the dilution of cellular heterogeneity caused by over-smoothing. At the same time, this framework effectively decouples graph-based modelling from graph-free reasoning, reduces the end-to-end reliance on pre-constructed graphs, and significantly reduces the computational overhead associated with large-scale cellular graphs. AVAILABILITY: All datasets used in this study are publicly available. Detailed information about all the single-cell datasets analyzed in this paper is provided in Supplementary Table 1. The source code is available at https://github.com/gxl1110/scAdaG2M and https://doi.org/10.5281/zenodo.21512950. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Authors

Institutions

Publication Details

Journal
PubMed
Published
2026-11-13
DOI
https://doi.org/10.1093/bioinformatics/btag827
Primary Topic
Single-cell and spatial transcriptomics
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Topology-Feature Disentangled Deep Clustering Method for scRNA-seq Based on Graph-to-MLP Distillation.

高喜来, Chen Cheng, Xiaoyi Lv, Yang Zhao et al.
3 links
PubMed
Single-cell and spatial transcriptomics
article

A Topology-Feature Disentangled Deep Clustering Method for scRNA-seq Based on Graph-to-MLP Distillation.

高喜来, Chen Cheng, Xiaoyi Lv, Yang Zhao, Xiangwei Zhao, Hao Liu
article en

Abstract

MOTIVATION: Graph neural networks (GNNs), owing to their ability to model topological relationships between cells, have gradually emerged as a mainstream approach in the clustering analysis of single-cell RNA sequencing (scRNA-seq) data. However, their practical application still faces the following challenges: (1) excessive smoothing undermines cellular heterogeneity; (2) pre-constructed graphs introduce end-to-end dependencies; (3) large-scale cellular graph computations are computationally intensive. RESULTS: To fully utilise the expression profiles and topological information contained within scRNA-seq data and overcome these limitations, this paper proposes a Graph-to-MLP (G2M) knowledge distillation method incorporating AdaBoost principles, and builds upon this to develop a graph-disentangled single-cell clustering framework-scAdaG2M. This framework utilises a GNN teacher model to learn higher-order topological relationships and local neighbourhood structures between cells on a cellular graph, whilst multiple lightweight MLP student models perform feature learning in an ungraphed setting. By combining this with an AdaBoost-style dynamic reweighting strategy, the framework enhances the model's ability to characterise cells that are difficult to distinguish, cells with blurred boundaries, and potentially rare cell populations, thereby mitigating the dilution of cellular heterogeneity caused by over-smoothing. At the same time, this framework effectively decouples graph-based modelling from graph-free reasoning, reduces the end-to-end reliance on pre-constructed graphs, and significantly reduces the computational overhead associated with large-scale cellular graphs. AVAILABILITY: All datasets used in this study are publicly available. Detailed information about all the single-cell datasets analyzed in this paper is provided in Supplementary Table 1. The source code is available at https://github.com/gxl1110/scAdaG2M and https://doi.org/10.5281/zenodo.21512950. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PubMed
People's Hospital of Xinjiang Uygur Autonomous Region (CN), State Key Laboratory of Digital Medical Engineering (CN), Beijing National Research Center for Information Science and Technology (CN), Southeast University (CN), Xinjiang University (CN), Tsinghua University (CN)
Single-cell and spatial transcriptomics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon