TiCoNE 2: A Composite Clustering Model for Robust Cluster Analyses on Noisy Data

TL;DR

TiCoNE 2 introduces a composite clustering model that integrates multiple data types, including time-series and network data, to improve robustness and accuracy in noisy biomedical datasets.

Contribution

It extends the original TiCoNE approach to seamlessly incorporate multiple data types, enhancing cluster recovery and noise robustness in biomedical data analysis.

Findings

Successfully recovers embedded cluster patterns in noisy data.

More robust to noise than single data type models.

Effectively distinguishes foreground from background clusters.

Abstract

Identifying groups of similar objects using clustering approaches is one of the most frequently employed first steps in exploratory biomedical data analysis. Many clustering methods have been developed that pursue different strategies to identify the optimal clustering for a data set. We previously published TiCoNE, an interactive clustering approach coupled with de-novo network enrichment of identified clusters. However, in this first version time-series and network analysis remained two separate steps in that only time-series data was clustered, and identified clusters mapped to and enriched within a network in a second separate step. In this work, we present TiCoNE 2: An extension that can now seamlessly incorporate multiple data types within its composite clustering model. Systematic evaluation on 50 random data sets, as well as on 2,400 data sets containing enriched cluster structure and varying levels of noise, shows that our approach is able to successfully recover cluster patterns embedded in random data and that it is more robust towards noise than non-composite models using only one data type, when applied to two data types simultaneously. Herein, each data set was clustered using five different similarity functions into k=10/30 clusters, resulting to ~5,000 clusterings in total. We evaluated the quality of each derived clustering with the Jaccard index and an internal validity score. We used TiCoNE to calculate empirical p-values for all generated clusters with different permutation functions, resulting in ~80,000 cluster p-values. We show, that derived p-values can be used to reliably distinguish between foreground and background clusters. TiCoNE 2 allows researchers to seamlessly analyze time-series data together with biological interaction networks in an intuitive way and thereby provides more robust results than single data type cluster analyses.