Leakage-safe diffusion augmentation with KAN-based models for imbalanced microarray gene-expression classification

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Gene-expression classification is crucial for functional genomics, disease subtyping, and biomarker discovery.
• Public microarray datasets often present challenges like high dimensionality, small sample sizes, and severe class imbalance.
• Existing methods can suffer from information leakage, leading to unreliable performance estimates and biological interpretations.

Purpose of the Study:

• To introduce FoDiKAN, a leakage-aware framework for robust gene-expression classification.
• To develop algorithms (SafeCV and AugTrain) that integrate fold-local diffusion augmentation with Kolmogorov-Arnold Networks (KANs).
• To ensure reliable performance estimation and interpretation in high-dimensional, imbalanced transcriptomic data.

Main Methods:

• SafeCV implements leakage-safe outer cross-validation, fitting operators on fold-local inner-training splits.
• AugTrain employs hybrid gene selection (mRMR, Boruta), trains a diffusion model, and generates synthetic minority samples using anchored DDIM sampling.
• Synthetic samples are used exclusively for training with down-weighting, integrated within a KAN backbone.

Main Results:

• The best FoDiKAN configuration achieved a macro-F1 score of 88.6% across 25 microarray datasets.
• This performance surpassed fixed-reference Gradient Boosting (by 4.2%) and XGBoost (by 2.4%) while competing with balanced baselines.
• Analyses quantified real-synthetic mismatch and assessed design choices through ablation and sensitivity studies.

Conclusions:

• FoDiKAN provides a robust and leakage-aware approach for gene-expression classification on challenging microarray data.
• The framework effectively leverages diffusion augmentation and KANs to improve classification performance and reliability.
• The study highlights the importance of leakage-aware methodologies and provides diagnostics for interpreting augmentation effects.