A Curvature Guided Composite Kernel Framework for Differential Gene Selection in Cancer Transcriptomics

Technical Summary: A Curvature Guided Composite Kernel Framework for Differential Gene Selection in Cancer Transcriptomics

Problem Statement
Differential Gene Expression (DEG) analysis is fundamental for biomarker discovery and drug target identification in cancer transcriptomics. However, existing methods face significant limitations depending on the data modality (bulk RNA-seq vs. single-cell RNA-seq) and the underlying statistical assumptions.

• Parametric methods (e.g., DESeq2, edgeR) rely on negative binomial distributions and generalized linear models. They often struggle with class imbalance, small sample sizes, and the specific sparsity characteristics of single-cell data.
• Non-parametric methods (e.g., SAMseq, NOIseq) handle variable sequencing depths but can be computationally expensive due to resampling or may suffer from data loss and reduced biological insight.
• Machine Learning approaches using standard kernels (Linear, Polynomial, RBF) often fail to capture the complex, non-linear, and hierarchical biological relationships inherent in high-throughput NGS data. RBF kernels, for instance, may oversmooth sharp biological features or become unstable at trend separation lines.

The paper posits that current methods lack a unified, modality-agnostic framework that can effectively model the dual nature of gene states (normal vs. cancerous) while preserving the intrinsic geometric structure of the data without relying on rigid distributional assumptions.

Methodology
The authors propose a "Curvature Guided Composite Kernel Framework" that integrates quantum mechanical analogies, differential geometry, and physics-inspired optimization. The methodology proceeds through the following stages:

1. Quantum-Inspired State Representation:

   • Gene expression states are modeled using the first postulate of quantum mechanics, representing genes as wavefunctions $|\Psi\rangle$ in a Hilbert space.
   • A Composite Kernel is constructed based on the Linear Combination of Atomic Orbitals (LCAO) approach. The gene state vector is partitioned into two orthogonal regimes: a "normal" regime ($|\Phi_i\rangle$) and a "cancer" regime ($|\Xi_i\rangle$).
   • The framework imposes axioms ensuring orthogonality between normal and cancer wavefunctions, treating a gene not as a single point but as a dual entity capable of simultaneous projection into healthy and pathological states.
   • A composite inner product is defined to map the expression matrix into a Reproducing Kernel Hilbert Space (RKHS), generating a Positive Semi-Definite (PSD) Gram matrix.

2. Geometric Analysis via Ricci Curvature:

   • The induced Hilbert space is modeled as a Riemannian manifold. The authors compute the Oliver-Ricci Curvature (ORC) on the discrete network of gene transcripts.
   • Unlike continuous differential operators which are unstable on noisy, finite datasets, ORC relies on the discrete topology of the network and transition probabilities to aggregate local geometric structures into a density value for each gene.
   • This curvature spectrum identifies geometric anomalies; critical DEGs are hypothesized to exhibit highly negative Ricci curvature (bottlenecks), while redundant genes show flat or positive curvature.

3. Gravitational Search Optimization:

   • To autonomously determine the optimal number of DEGs ($k$) without visual heuristics, a Gravitational Search Algorithm (GSA) is employed.
   • A representational loss function $L_{rep}(S)$ is minimized, comprising three terms:
     1. Empirical classification error ($\epsilon_{class}$).
     2. A linear penalty for subset size (promoting sparsity).
     3. A geometric regularization term based on the exponential of the Ricci curvature ($\exp(Ric)$). This term heavily penalizes the inclusion of flat/positively curved genes while minimizing the cost of including negatively curved bottleneck genes.
   • Genes are treated as particles with "mass" determined by their fitness contribution to minimizing the global loss. A dynamic gravitational force moves particles in the feature space to converge on the optimal gene subset.

Key Contributions

• Novel Framework: Introduction of a quantum-inspired composite kernel that explicitly models the duality of gene states (normal vs. cancer) within a Hilbert space, differing from standard linear combinations of kernels.
• Geometry-Aware Selection: Utilization of Ricci curvature on a Riemannian manifold to identify topological anomalies (bottlenecks) associated with differentially expressed genes, moving beyond statistical significance alone.
• Physics-Informed Optimization: Development of a gravitational search mechanism that dynamically selects the optimal gene subset by minimizing a loss function that balances classification accuracy, sparsity, and geometric relevance.
• Modality Agnosticism: The method is designed to function across different data modalities, including bulk RNA-seq (NGS) and microarray data, without requiring modality-specific parameter tuning.

Results
The framework was validated on multiple datasets from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO/CuMiDa), covering cancers such as Lung Adenocarcinoma (LUAD), Head and Neck Cancer (HNSC), Breast Cancer (BRCA), Liver, and Pancreatic cancer.

• Classification Performance:
  • On TCGA datasets, the proposed method consistently improved classification accuracy compared to using the original full transcriptomic feature space. For SVM with RBF kernels, accuracy increased from ~98% to ~99% in BRCA and LUAD.
  • XGBoost also showed gains, particularly in BRCA (accuracy rising from 95% to 97.85%).
  • The method outperformed or matched established benchmarks (DESeq2, Voom-Limma, NOISeq) in most scenarios. Notably, while NOISeq achieved perfect accuracy on one dataset, it failed to identify significant DEGs in microarray cohorts, whereas the proposed method remained robust.
• External Validation:
  • The framework demonstrated strong cross-platform generalizability. DEGs selected from TCGA (NGS) data maintained predictive performance when applied to independent GEO (microarray) datasets, and vice versa.
  • In breast cancer external validation, using TCGA-derived DEGs on GSE data improved XGBoost accuracy from 79.89% to 87.21%.
• Biological Enrichment:
  • Gene Set Enrichment Analysis (GSEA) and KEGG pathway analysis confirmed that the selected genes were biologically relevant.
  • Identified pathways consistently included DNA damage response, cell-cycle control, apoptosis, immune modulation, and major oncogenic signaling (e.g., p53, Notch, Hedgehog, RAS), aligning with known cancer biology across heterogeneous tissue types.
• Dimensionality Reduction:
  • PCA and UMAP visualizations showed that the selected DEG subsets created low-dimensional manifolds with improved intra-class compactness and inter-class separability compared to the original data, effectively removing noise and redundancy.

Significance and Claims
The paper claims that this work presents a new paradigm for DEG identification that moves beyond traditional distributional assumptions. By leveraging quantum mechanical axioms to define gene states and differential geometry to analyze the latent manifold, the framework successfully extracts compact, informative, and biologically coherent gene signatures.

The authors emphasize that the method is modality-agnostic, capable of handling both NGS and microarray data with consistent performance. The external validation results suggest that the selected features capture disease-associated expression signatures that are robust to platform variations and dataset-specific artifacts. The study concludes that this geometry-aware, quantum-inspired strategy offers a promising tool for biomarker discovery and disease classification, providing a stable and biologically meaningful representation of transcriptomic data. The authors note that further validation on larger independent cohorts and single-cell datasets is required to fully assess clinical translational potential.