VillageNet: Graph-based, Easily-interpretable, Unsupervised Clustering for Broad Biomedical Applications

Clustering complex, large-scale biomedical data is essential for precision medicine applications. Because biomedical data may reveal latent biological patterns or subgroups with significant clinical outcomes, clustering these data is important for downstream tailoring of medical therapies for distinctive patient subgroups. Complexity in biomedical data may originate from inherently variable features of datasets and/or heterogeneous sources of information, such as electronic health records or physiological, cellular, and/or molecular assays. The more novel and/or complex a biomedical dataset is, the less is usually known at the offset about its inherent features, e.g. labels, linearity, etc. that can limit the initial selection of suitable clustering techniques. Building upon our previous work (i.e., MapperPlus), we introduce VillageNet, an unsupervised clustering framework that integrates topological principles, graph-based community detection, and random-walk analysis to derive data-driven knowledge in an unsupervised context. VillageNet autonomously infers the number of clusters directly from the data and demonstrates a robust ability to identify clusters with non-linear separation, thereby avoiding restrictive assumptions about cluster geometry, a commonly unknown feature of biomedical datasets. VillageNet was evaluated on an extensive suite of non-biomedical benchmark datasets with known ground-truth labels, as well as four heterogeneous biomedical datasets (flow cytometry, tissue imaging, single-cell gene expression, and image-derived data). VillageNet achieved overall superior performance when assessed using normalized mutual information and an adjusted Rand index, and favorable computational properties, with runtime scaling linearly with both dataset size and dimensionality-thereby eliminating the need for dimension-reduction procedures. Together, these findings establish VillageNet as a scalable, topology-informed, and broadly generalizable framework for clustering complex biomedical datasets, especially during the discovery phase when most features about complex datasets may still be unknown.