RMKL

RMKL implements multiple kernel learning (MKL) algorithms in R to integrate heterogeneous genomic and clinical datasets for predictive modeling and biomarker discovery.

Key Features:

• Multiple MKL implementations: Includes SimpleMKL (Rakotomamonjy et al., 2008), Simple and Efficient MKL (Xu et al., 2010), and Dual Augmented Lagrangian MKL (Suzuki and Tomioka, 2011).
• Cross-validation for SVM models: Provides functions to perform cross-validation to identify optimal kernel shapes and hyperparameter combinations and to construct candidate kernels for MKL.
• Automated kernel prioritization and parameter tuning: Implements an automated scheme for prioritizing kernels and tuning parameters for model selection and optimization.
• Integration of clinical and miRNA expression data: Supports integration of clinical information with miRNA gene expression data, as demonstrated in ovarian cancer analyses.
• Prognostic prediction using TCGA: Applied to identify gene sets involved in prognostic prediction across 15 cancer types using The Cancer Genome Atlas (TCGA) gene expression data.
• Simulated data support: Can handle simulated data for model selection and evaluation.
• Convex kernel combination and SVM-based modeling: Facilitates convex combinations of candidate kernels to build SVM classifiers or regressors.

Scientific Applications:

• Integrative miRNA-clinical analysis (ovarian cancer): Integrates clinical variables with miRNA expression to construct unified predictive models in ovarian cancer studies.
• Cancer prognostic biomarker discovery (TCGA): Identifies gene sets associated with prognosis across 15 cancer types using TCGA gene expression data.
• Model selection and benchmarking with simulated data: Uses simulated datasets to evaluate and select MKL model configurations.
• Predictive modeling for diagnosis, prognosis, and treatment stratification: Builds multi-source predictive models from high-throughput genomic and clinical data for clinical research applications.

Methodology:

Employs modern optimization techniques within the MKL framework, facilitating convex combinations of candidate kernels to construct an optimal hyperplane for classification or regression tasks.