CPSWin Lab · Kansas State University · May 2024 – Jul 2024
An inductive GNN framework for single-cell RNA-seq imputation and clustering on glioblastoma data. The visualisation below follows the pipeline from sparse counts to clustering metrics.
01 Imputation and LTMG
Created an inductive GNN framework for scRNA-seq imputation and clustering that recovers dropout zeros while retaining cellular manifold structure. LTMG fits each gene a left-truncated mixture of Gaussians over its regulatory states, and the result is assembled into tensors for the model.
02 VGAE
Extended the pipeline with VGAE preprocessing for GBM-specific analysis: a GCN encoder maps the cell graph to a latent Gaussian, sampling by reparameterisation with z = μ + σε, and an inner-product decoder reconstructs the adjacency as  = σ(ZZᵀ).
03 MST ensembles
Latent samples drawn from μ and σ are combined through MST-based ensemble embeddings: a minimum spanning tree over the embedding links cells into clusters that respect the manifold rather than imposing spherical groups.
04 Metrics
The framework improves downstream clustering while cutting runtime: +60% Silhouette and +14.9% adjusted Rand index, with better heterogeneity analysis accuracy on glioblastoma test data.
The pipeline drawn above: a gene by cell count matrix with dropout zeros imputed by an inductive GNN, LTMG fitting a left-truncated mixture of Gaussians per gene, a VGAE whose GCN encoder yields μ and σ with reparameterised sampling and an inner-product adjacency decoder, MST-based ensemble embeddings forming clusters, and unsupervised metrics, Silhouette and adjusted Rand index, scoring the result.
Each scroll hands its output to the next stage: the matrix and mixture become tensors, the tensors enter the VGAE, the latent samples become the clustered points, and the clusters are what the metrics measure.