A fmLASSO framework for multi-omics cancer prediction and biomarker discovery

Maximizing predictive accuracy while minimizing the number of selected biomarkers represents two conflicting objectives in Multi-Omics cancer modeling. In this context, high-dimensionality and block heterogeneity often compromise model performance. Prediction accuracy strongly depends on the relevance and redundancy of features within each omics block, as irrelevant variables introduce noise and weaken interpretability. To address these challenges, this study presents a hybrid regularization framework for Multi-Omics cancer prediction that integrates standard Least Absolute Shrinkage and Selection Operator (LASSO), mbLASSO, and fmLASSO clustering. The proposed approach introduces adaptive fuzzy block weighting, in which soft memberships are assigned to each block prior to blockwise regularized regression, thereby balancing information across heterogeneous data sources. Nested cross-validation and block level optimization are employed to enhance model stability and robustness. Experimental evaluation on the RNA Pan-Cancer dataset from The Cancer Genome Atlas demonstrates that the proposed fuzzy Multiblock framework achieves high predictive performance, attaining a classification accuracy of 99.75 % , an AUC of 1.00 , and an RMSE of 0.0409 , outperforming both standard LASSO and mbLASSO in terms of predictive stability and sparsity. Moreover, functional enrichment analysis confirms that the selected genes are strongly associated with known cancer related pathways, supporting the biological relevance of the proposed method. These results highlight the effectiveness and competitiveness of the proposed hybrid regularization framework for high-dimensional cancer prediction.