Integrated Features Based on Graph Clustering and Gene Expression

Abstract

Integrating different biological features – for example, the informativeness of topological features and gene expression – is challenging because each feature must be accounted for individually if the features are used to help forecast models. In this process, ensuring that the outcomes reflect the underlying biological structure of the network information while minimizing noise and irrelevant data is crucial. This study identifies the importance of rigorous pre-analyses in determining statistically significant correlations and joint effects among preprocess features before applying machine-learning techniques. Thus, when deploying multidimensional datasets, a systematic multi-feature methodology is presented in this paper to unify optimized graph clustering, weighted Jaccard similarity, and dimension reduction based on principal component analysis (PCA). Specifically, the objective was to identify novel uncharacterized gene associations in complex biological networks. Moreover, this study offers more refined insights into gene interactions within their networks, revealing patterns and relationships that might be hidden by broad data analysis. The method's performance was validated according to the benchmarks for a Dialogue on Reverse Engineering Assessment of Methods, fifth edition (DREAM5) challenge project, to determine its ability to analyze complex biological networks.