mlrMBO

The R package mlrMBO is a software tool that implements a subgroup-based weighted likelihood approach for survival prediction using high-dimensional genetic covariates, such as gene expression data. This approach aims to address the challenges of building reliable prediction models for specific cancer subgroups or cohorts with limited sample sizes and potentially high censoring rates in survival analysis.

Key features and functionality of mlrMBO:

1. Weighted likelihood approach: When predicting survival for a specific subgroup, mlrMBO assigns individual weights to observations from other subgroups. These weights determine the strength with which the observations from each subgroup contribute to the model-building process.

2. Model-based optimization (MBO): mlrMBO employs MBO to efficiently identify the best prediction model in the presence of many hyperparameters. MBO is used to optimize the weights for additional subgroups in a Cox proportional hazards model, commonly used for survival analysis.

3. Identification of similar subgroups: By optimizing the weights, mlrMBO can identify cancer subgroups with similar relationships between covariates (e.g., gene expression) and the target variable (survival). It allows for the inclusion of only the most relevant subgroups in the model-building process, reducing bias due to heterogeneity between cohorts.

4. Evaluation of lung cancer cohorts: The approach implemented in mlrMBO has been evaluated using a set of lung cancer cohorts with gene expression measurements. The resulting models have shown competitive prediction quality and reflect the similarity of the corresponding cancer subgroups through the assigned weights (close to 0, close to 1, or medium weights).