A1405
Title: Correlation-informed adaptive regularization for multiclass glioma subtype prediction
Authors: Mina Norouzirad - Center for Mathematics and Applications (NOVA Math), NOVA School of Science and Technology (Portugal) [presenting]
Marta Lopes - Department of Mathematics NOVA School of Science and Technology (Portugal)
Tomas Bandeira - NOVA University Lisbon (Portugal)
Abstract: Gliomas display substantial molecular heterogeneity, making accurate subtype classification critical for prognosis and treatment selection. A multiclass prediction approach for astrocytoma, glioblastoma, and oligodendroglioma was developed using RNA-seq profiles from 619 The Cancer Genome Atlas (TCGA) samples with 20,504 genes. In this ultra-high-dimensional regime, standard penalized multinomial logistic regression may yield unstable selection because many genes are strongly correlated and convey overlapping signals. A two-phase correlation-informed adaptive regularization strategy is proposed. First, gene-specific relevance is estimated with a correlation-based scoring procedure for multinomial logistic regression, explicitly accounting for predictor dependence and attenuating redundant contributions. Second, these scores are converted into adaptive penalty weights by taking The inverse of their absolute values and fitting an adaptive LASSO multinomial model, promoting sparse, interpretable gene signatures while prioritizing strongly associated genes. This approach is evaluated against LASSO, ridge, and elastic net under cross-validation, focusing on predictive performance and The interpretability of selected biomarkers for glioma subtype classification.
