Elastic net regularization

Overview

Elastic net is a penalized regression method that combines L1 (LASSO) and L2 (Ridge) regularization. By balancing sparsity-inducing L1 penalty with coefficient-shrinkage L2 penalty, elastic net performs simultaneous feature selection and regularization, making it well suited to high-dimensional biomedical data where features are correlated (e.g., immune gene signatures). It is commonly used to build multi-feature predictive models from clinical, genomic, and transcriptomic inputs. When applied as a logistic model, it yields binary classifiers (e.g., responder vs. non-responder) with an interpretable, sparse feature set. When applied within a Cox framework (e.g., OncoCast-MPM), it yields survival risk scores.

Used by

  • A 25-predictor elastic net logistic regression model trained on IMvigor210 (atezolizumab-treated metastatic UC) was applied to predict ICI response in UC-GENOME (218 patients with metastatic urothelial carcinoma); achieved AUC = 0.84 (IMvigor210 validation), 0.82 (UNC-108), and 0.65 (UC-GENOME), significantly outperforming TMB alone (AUC 0.68, p = 0.038); the model integrated TMB, ECOG status, molecular subtype, and immune gene signatures PMID:36333289
  • Used to build pharmacogenomic predictive models from CCLE genomic features (mutations, copy-number, expression) across 947 cell lines and 24 drugs PMID:22460905
  • Used as one of three orthogonal classifiers (alongside ISOpure and OncoSign) to assign 88 mixed IDC/ILC breast tumors to ILC-like or IDC-like molecular classes; CDH1 mutation status was the dominant feature driving the classification PMID:26451490

Notes

  • Elastic net feature selection depends on the mixing parameter alpha (0 = Ridge, 1 = LASSO); intermediate alpha values retain correlated features.
  • Cross-validation is required for hyperparameter tuning (alpha, lambda); overfitting risk increases with small training cohorts.
  • Portability across cohorts (e.g., IMvigor210 to UC-GENOME) can be limited by differences in RNA-seq methodology (capture-based vs. total RNA-seq).
  • In the UC-GENOME application, the EN model was ICI-specific (not predictive of chemotherapy response, AUC 0.536), supporting its mechanistic basis in immune biology.

Sources

  • PMID:36333289 — UC-GENOME metastatic urothelial carcinoma study; elastic net model predicting ICI response from multivariate clinical and immunogenomic features.

This page was processed by crosslinker on 2026-05-14. - PMID:22460905

This page was processed by crosslinker on 2026-05-14. - PMID:26451490

This page was processed by crosslinker on 2026-05-14.