SCMG: A Multi-Omics and Stage-Resolved Single-Cell Framework for Cancer Driver Gene Prioritization

Technical Summary: SCMG – A Multi-Omics and Stage-Resolved Single-Cell Framework for Cancer Driver Gene Prioritization

Problem Statement
Current computational methods for prioritizing cancer driver genes largely rely on bulk molecular profiles (mutation burden, DNA methylation, and gene expression) measured at the time of tumor sampling. While network-based approaches and graph neural networks (GNNs) have improved the integration of multi-omics data and gene interaction topologies, they often overlook the dynamic transcriptional changes occurring across disease stages. Specifically, for epithelial tumors progressing through Normal, Precancer, and Cancer states, existing methods fail to fully utilize stage-resolved single-cell RNA sequencing (scRNA-seq) data. This data contains critical information regarding epithelial-cell expression levels, the proportion of expressing cells, and mean expression changes across stages, which could serve as auxiliary inputs to better organize and interpret high-scoring candidate genes.

Methodology
The authors propose SCMG (Single-Cell Multi-omics Graph), a graph-learning framework designed to integrate bulk multi-omics features, stage-resolved epithelial single-cell data, and biological gene interaction networks (GINs). The framework operates through a dual-branch architecture with a late-logit fusion strategy:

1. Data Integration:

   • Bulk Branch: Inputs 48-dimensional features derived from TCGA data (mutation frequency, DNA methylation, and gene expression across 16 cancer types). These are processed through a "disagreement-aware bulk multi-view encoder" that combines local gene features with graph-propagated information from one of six GINs (CPDB, STRING, IRefIndex, PCNet, PathNet, GGNet).
   • scRNA Branch: Inputs stage-resolved epithelial expression profiles from three datasets (GSE251990, GSE161277, GSE218170) covering Normal, Precancer, and Cancer states. The model constructs "stage-source tokens" comprising mean log-normalized expression, detection rates, and expressing-cell means. These are processed via a "stage-source scRNA progression encoder" using Transformers to model direct stage-source relationships and progression patterns (differences between adjacent stages).
   • Static-Dynamic Integration: The scRNA branch further combines static expression summaries with dynamic progression representations using a hybrid encoder and Feature-wise Linear Modulation (FiLM).

2. Model Architecture:

   • The two branches (bulk and scRNA) are optimized independently using weighted binary cross-entropy.
   • A learnable linear late-logit fusion layer combines the logits from both branches to produce the final cancer driver score.
   • The model utilizes graph convolutional networks (GCNs) and Transformers to capture both local gene features and global topological/progression signals.

3. Post-Hoc Interpretation:

   • To interpret high-scoring candidates, the authors calculate an Epithelial Alteration Score (EAS). This metric quantifies the balance of expression changes between the Normal-to-Precancer and Precancer-to-Cancer transitions.
   • A background-corrected residual EAS is used to organize high-scoring genes into two operational regions: Q1 (residual EAS ≥ 0, indicating larger Precancer-to-Cancer changes) and Q2 (residual EAS < 0, indicating larger Normal-to-Precancer changes).

Key Results

• Pan-Cancer Performance: SCMG was evaluated against five baseline methods (HGDC, EMOGI, DeepCDG, TGCN, SGCD) across six different gene interaction networks. SCMG achieved the highest Area Under the Precision-Recall Curve (AUPRC) on all six networks, with values ranging from 0.8763 (GGNet) to 0.9339 (PathNet).
• Cancer-Type Specificity: In a pan-cancer analysis across 15 cancer types using the CPDB network, SCMG achieved the highest AUPRC in 9 types (including BRCA, BLCA, LIHC, COAD) and the highest AUROC in 7 types. The most significant relative improvement was observed in UCEC (14.65% over the next best method).
• Ablation Studies: Removing single-cell features reduced AUPRC from 0.9113 (full model) to 0.9018 (bulk-only), confirming the incremental value of stage-resolved data. Removing protein-protein interaction (PPI) propagation caused the largest performance drop (AUPRC 0.8092), highlighting the importance of network topology.
• External Validation:
  • Driver Catalog Enrichment: SCMG rankings were significantly enriched for IntOGen driver genes outside the training positive-label set (Top 500 OR = 9.40, FDR = 2.28 × 10⁻²⁰). This enrichment persisted even after excluding TCGA and PCAWG-derived driver records.
  • Clinical Association: Signatures derived from Q1 and Q2 genes were associated with patient outcomes (Overall Survival, Progression-Free Interval, Disease-Specific Survival) in TCGA cohorts, with both high-score groups showing poorer survival outcomes.
  • External Cohort: In an external cutaneous squamous cell carcinoma (cSCC) single-cell cohort, selected candidates (e.g., FN1, APP, FYN, RYR3) showed consistent tumor-versus-normal expression patterns in basal cell subtypes.

Significance and Claims
The paper claims that SCMG provides a novel framework for integrating stage-resolved epithelial context into cancer driver gene prioritization. Its primary contributions are:

1. Methodological: Demonstrating that combining bulk multi-omics with stage-resolved single-cell summaries via a dual-branch graph-learning framework improves driver gene prioritization over methods relying solely on bulk data.
2. Interpretability: Establishing a coordinate system (Prediction Score vs. Residual EAS) that organizes high-scoring candidates into distinct groups (Q1 and Q2) based on their expression trajectories across disease stages. Q2 genes showed a progressive decrease in expression from Normal to Cancer, while Q1 genes decreased in Precancer but recovered in Cancer.
3. Biological Relevance: Showing that these distinct groups are enriched for specific cancer-related functional themes (e.g., PI3K–RTK/MAPK signaling, cell cycle) and that their signatures correlate with patient survival.

The authors maintain a modest tone, noting that the single-cell inputs represent three available epithelial contexts rather than a unified pan-cancer trajectory, and that the negative labels form a conservatively selected background. They emphasize that while the framework supports candidate selection, functional testing is required to confirm driver activity.