Department of Computer Science, Johns Hopkins University, Baltimore, MD, 21218, USA.
Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD, 21218, USA.
F1000Res. 2024 Feb 29;11:1460. doi: 10.12688/f1000research.126215.2. eCollection 2022.
Multi-view datasets are becoming increasingly prevalent. These datasets consist of different modalities that provide complementary characterizations of the same underlying system. They can include heterogeneous types of information with complex relationships, varying degrees of missingness, and assorted sample sizes, as is often the case in multi-omic biological studies. Clustering multi-view data allows us to leverage different modalities to infer underlying systematic structure, but most existing approaches are limited to contexts in which entities are the same across views or have clear one-to-one relationships across data types with a common sample size. Many methods also make strong assumptions about the similarities of clusterings across views. We propose a Bayesian multi-view clustering approach (BMVC) which can handle the realities of multi-view datasets that often have complex relationships and diverse structure. BMVC incorporates known and complex many-to-many relationships between entities via a probabilistic graphical model that enables the joint inference of clusterings specific to each view, but where each view informs the others. Additionally, BMVC estimates the strength of the relationships between each pair of views, thus moderating the degree to which it imposes dependence constraints. We benchmarked BMVC on simulated data to show that it accurately estimates varying degrees of inter-view dependence when inter-view relationships are not limited to one-to-one correspondence. Next, we demonstrated its ability to capture visually interpretable inter-view structure in a public health survey of individuals and households in Puerto Rico following Hurricane Maria. Finally, we showed that BMVC clusters integrate the complex relationships between multi-omic profiles of breast cancer patient data, improving the biological homogeneity of clusters and elucidating hypotheses for functional biological mechanisms. We found that BMVC leverages complex inter-view structure to produce higher quality clusters than those generated by standard approaches. We also showed that BMVC is a valuable tool for real-world discovery and hypothesis generation.
多视图数据集越来越普遍。这些数据集由不同的模态组成，这些模态提供了对同一底层系统的互补描述。它们可以包括具有复杂关系、不同程度缺失和各种样本大小的异构类型的信息，这在多组学生物学研究中经常发生。聚类多视图数据可以让我们利用不同的模态来推断底层的系统结构，但大多数现有的方法都局限于实体在视图中是相同的或在数据类型之间具有明确的一一对应关系并且具有相同的样本大小的情况。许多方法还对视图之间的聚类相似度做出了很强的假设。我们提出了一种贝叶斯多视图聚类方法（BMVC），可以处理多视图数据集的实际情况，这些数据集通常具有复杂的关系和多样的结构。BMVC 通过概率图形模型来处理实体之间已知的和复杂的多对多关系，该模型可以联合推断特定于每个视图的聚类，但每个视图也可以为其他视图提供信息。此外，BMVC 还估计了每对视图之间的关系强度，从而适度地强加了依赖约束的程度。我们在模拟数据上对 BMVC 进行了基准测试，以表明当视图之间的关系不限于一对一对应时，它可以准确地估计不同程度的视图间依赖关系。接下来，我们展示了它在波多黎各飓风玛丽亚后对个人和家庭的公共卫生调查中捕捉直观的视图间结构的能力。最后，我们表明，BMVC 聚类可以整合乳腺癌患者数据的多组学特征之间的复杂关系，从而提高聚类的生物学同质性，并阐明功能生物学机制的假设。我们发现，BMVC 利用复杂的视图间结构生成了比标准方法生成的聚类质量更高的聚类。我们还表明，BMVC 是现实世界发现和假设生成的有价值的工具。
