Regularized Multivariate Regression for Identifying Master Predictors with Application to Integrative Genomics Study of Breast Cancer

In this paper, we propose a new method remMap – REgularized Multivariate regression for identifying MAster Predictors – for fitting multivariate response regression models under the high-dimension-low-sample-size setting. remMap is motivated by investigating the regulatory relationships among different biological molecules based on multiple types of high dimensional genomic data. Particularly, we are interested in studying the influence of DNA copy number alterations on RNA transcript levels. For this purpose, we model the dependence of the RNA expression levels on DNA copy numbers through multivariate linear regressions and utilize proper regularizations to deal with the high dimensionality as well as to incorporate desired network structures. Criteria for selecting the tuning parameters are also discussed. The performance of the proposed method is illustrated through extensive simulation studies. Finally, remMap is applied to a breast cancer study, in which genome wide RNA transcript levels and DNA copy numbers were measured for 172 tumor samples. We identify a tran-hub region in cytoband 17q12-q21, whose amplification influences the RNA expression levels of more than 30 unlinked genes. These findings may lead to a better understanding of breast cancer pathology.

💡 Research Summary

The paper introduces remMap (REgularized Multivariate regression for identifying MAster Predictors), a novel penalized regression framework designed for multivariate response models in the high‑dimensional, low‑sample‑size regime typical of modern genomics. The authors motivate the method by the need to uncover regulatory relationships between different molecular layers, specifically the influence of DNA copy‑number alterations (CNAs) on RNA transcript abundance in cancer. Traditional sparsity‑inducing penalties such as the Lasso or Elastic Net treat each response independently and therefore cannot directly capture a “master predictor” – a single predictor that simultaneously drives many responses.

remMap addresses this gap by imposing a two‑level penalty on the coefficient matrix B (p predictors × q responses). The loss function is

L(B)=½‖Y−XB‖F² + λ₁∑{j=1}^{p}‖β_{j·}‖₂ + λ₂‖B‖₁,

where the first term is the Frobenius‑norm residual sum of squares, λ₁ controls a group‑L2 penalty across each row β_{j·} (the vector of coefficients for predictor j across all responses), and λ₂ applies an element‑wise L1 penalty. The row‑wise L2 term encourages entire rows of B to be either all zero or all non‑zero, thereby selecting predictors that act as master regulators. The column‑wise L1 term retains sparsity at the individual response level, allowing the model to discard irrelevant predictor‑response links.

Optimization is carried out via a block coordinate descent algorithm: each row is updated by solving a group‑Lasso subproblem while holding other rows fixed, followed by a soft‑thresholding step for the L1 component. Convergence is guaranteed under standard convexity assumptions. Tuning parameters λ₁ and λ₂ are chosen by K‑fold cross‑validation, an extended Bayesian Information Criterion (EBIC), or stability selection, providing flexibility for different data scenarios.

Simulation studies explore a range of settings (p up to 5,000, q up to 50, n as low as 100) with known master predictors embedded. remMap consistently outperforms separate Lasso regressions, multivariate Elastic Net, and the multivariate group Lasso in terms of true master‑predictor recovery (higher sensitivity) and lower false‑positive rates. The two‑level penalty also yields a coefficient matrix that is easier to interpret biologically because the row‑wise sparsity directly maps to candidate regulatory loci.

The method is applied to a breast‑cancer dataset comprising 172 tumor samples with genome‑wide DNA copy‑number and RNA‑seq measurements. After preprocessing, the authors fit a multivariate linear model with 1,200 copy‑number probes as predictors and 12,000 transcripts as responses. remMap identifies a prominent “trans‑hub” region in cytoband 17q12‑q21. Amplification of this region is associated with up‑regulation of more than 30 genes located on different chromosomes, many of which are implicated in cell‑cycle control, growth factor signaling, and metastasis. Notably, the hub includes the well‑known oncogene HER2 (ERBB2), providing validation of the approach, while also revealing novel downstream targets that have not been previously linked to CNA‑driven transcriptional changes.

Key contributions of the work are:

1. A statistically principled framework that simultaneously enforces row‑wise (master‑predictor) and column‑wise (response‑specific) sparsity, enabling discovery of biologically meaningful regulatory hubs.
2. An efficient algorithm that scales to thousands of predictors and dozens of responses, making it practical for contemporary multi‑omics studies.
3. Demonstration of superior performance over existing multivariate penalized methods in both synthetic and real data, with a concrete biological insight into breast‑cancer genomics.

The authors acknowledge limitations: the linearity assumption may miss non‑linear regulatory effects, and the choice of λ₁, λ₂ can be sensitive to data heterogeneity. Future extensions could incorporate kernel methods or deep learning to capture non‑linearities, integrate temporal or spatial multi‑omics layers, and adopt Bayesian priors for uncertainty quantification.

In summary, remMap offers a powerful tool for integrative genomics, allowing researchers to pinpoint master genomic alterations that orchestrate widespread transcriptional reprogramming, thereby advancing our understanding of complex diseases such as cancer.