msgl

msgl implements a multinomial sparse group lasso in R to perform multiclass classification of molecular profiles and to model tissue contamination for improved variable selection and diagnostic accuracy in molecular cancer diagnostics.

Key Features:

• Sparse Group Lasso Penalty: The package fits multinomial models using a sparse group lasso penalty to promote individual variable selection and group-wise sparsity for interpretable multiclass classifiers.
• Model of Tissue Contamination: It includes an explicit contamination model to account for benign (non-cancerous) tissue mixing with cancerous samples, which can distort molecular signatures.
• Integration and Demonstrated Error Reduction: The contamination model operates independently from molecular predictor models and can be integrated with them, with reported reductions in test error from 77% to 45% on liver metastases biopsies and further to 34% when contaminated samples were included in training.

Scientific Applications:

• Primary Tumor Site Identification: Applied to identification of primary tumor sites from liver metastases biopsies, addressing challenges posed by high proportions of benign tissue in such samples.
• MicroRNA Expression Analysis: Used with microRNA expression data to improve accuracy of predicting primary tumor site by explicitly modeling contamination effects.

Methodology:

msgl fits multinomial models with a sparse group lasso penalty and implements an independent contamination model that can be integrated with molecular predictor models.