matrans: an R package for parameter-transfer learning of multi-source cancer prognosis data

This is the authors' abstract. We don't add key points for this paper.

The rapid growth of multi-source molecular profiling data has created new opportunities for integrative cancer prognosis modeling. However, heterogeneity across datasets often limits model generalizability. Transfer learning provides a powerful strategy to leverage auxiliary information from related source domains. Parameter-transfer learning based on frequentist model averaging is a leading-edge approach that has been widely used in practice. Nevertheless, the lack of accessible and user-friendly software has limited its broader adoption. We introduce matrans , an R package that implements a parameter-transfer learning framework based on frequentist model averaging within a semiparametric setting. The proposed procedure circumvents a separate source-selection step by adaptively weighting the target and source estimators according to their predictive performance on the target data. Furthermore, the source models can be fitted locally, and the transfer process requires only source-level parameter estimates rather than the centralized pooling of individual-level source data. Simulation studies and a real-world multi-source cancer prognosis case study demonstrate that matrans achieves robust estimation accuracy, improved predictive performance, and high computational efficiency. The package is freely available at https://CRAN.R-project.org/package=matrans . matrans offers a convenient and efficient tool for parameter-transfer learning from multi-source cancer prognosis datasets, bridging the gap between methodological advances and practical biomedical applications.