LtransHeteroGGM: local transfer learning for Gaussian graphical model-based heterogeneity analysis.

Abstract

Heterogeneity is a hallmark of both macroscopic complex diseases and microscopic single-cell distribution. Gaussian graphical models (GGMs)-based heterogeneity analysis highlights its important role in capturing the essential characteristics of biological regulatory networks, but faces instability with scarce samples from rare subgroups. Transfer learning offers promise by leveraging auxiliary data, yet existing approaches rely on unrealistic overall similarity between domains, requiring the same subgroup number and similar parameters. Numerous biological problems call for local similarities, where only some subgroups share statistical structures. In this article, we propose LtransHeteroGGM, a novel local transfer learning framework for GGM-based heterogeneity analysis. It can achieve powerful subgroup-level local knowledge transfer between target and informative auxiliary domains, despite unknown subgroup structures and numbers, while mitigating the negative interference of non-informative domains. The effectiveness and robustness of the proposed approach are demonstrated through comprehensive numerical simulations and real-world T-cell heterogeneity analysis. The R implementation of LtransHeteroGGM is available at https://github.com/Ren-Mingyang/LtransHeteroGGM.