Penerapan Overlapping Community Detection pada Heterogeneous Network Terkait Kanker Berbasis Machine Learning

Abstract

Jaringan biologis merepresentasikan interaksi molekuler yang kompleks pada sistem kanker dan umumnya bersifat heterogen karena melibatkan berbagai jenis entitas biologis. Akibatnya, sebagian besar metode deteksi komunitas yang dirancang untuk jaringan homogen kurang optimal dalam menangkap hubungan lintas entitas. Penelitian ini bertujuan membangun representasi jaringan heterogen yang mengintegrasikan interaksi senyawa–protein dan protein–protein terkait kanker serta menerapkannya untuk mendeteksi komunitas yang saling tumpang tindih. Data interaksi biologis diperoleh dari Stanford Network Analysis Project dan data gen kanker dari Kyoto Encyclopedia of Genes and Genomes pada Human Cancer Pathway. Representasi jaringan dipelajari menggunakan embedding berbasis random walk, yaitu metapath2vec. Vektor hasil embedding kemudian digunakan untuk membangun jaringan homogen berbobot berdasarkan cosine similarity melalui pendekatan Top-K similarity. Jaringan hasil transformasi dianalisis menggunakan algoritma BigCLAM untuk mengidentifikasi komunitas tumpang tindih. Konfigurasi terbaik diperoleh pada K=12 komunitas dengan overlapping modularity 0,2366, internal density 0,4218, dan conductance 0,2578, menunjukkan komunitas yang kompak dan terpisah dengan baik. Sebanyak 33 gen teridentifikasi tumpang tindih antar komunitas, dengan PTEN dan CASP3 berperan sebagai hub multifungsional pada tiga komunitas. Pipeline yang dikembangkan juga berpotensi diadaptasi pada domain nonbiologis dengan struktur data serupa. Selain itu, pendekatan ini dapat mendukung penelitian lanjutan, seperti identifikasi biomarker dan penargetan obat pada kanker.

Biological networks represent complex molecular interactions in cancer systems and are typically heterogeneous, involving diverse biological entities. As a result, community detection methods designed for homogeneous networks often fail to capture cross-entity relationships. This study aims to construct a heterogeneous network representation integrating compound–protein and cancer-related protein–protein interactions, and apply it to detect overlapping communities. Interaction data were obtained from the Stanford Network Analysis Project, and cancer gene data from the KEGG Human Cancer Pathway. The network representation was learned using the random walk-based embedding method metapath2vec. The resulting embeddings were used to build a weighted homogeneous network based on cosine similarity via a Top-K similarity approach. This transformed network was then analyzed using the BigCLAM algorithm to identify overlapping communities. The optimal configuration was achieved at K=12 communities, with overlapping modularity of 0.2366, internal density of 0.4218, and conductance of 0.2578, indicating compact and well-separated communities. A total of 33 genes were identified as overlapping across communities, with PTEN and CASP3 acting as multifunctional hubs in three communities. The pipeline developed here has potential for adaptation to non-biological domains with similar data structures, and supports further research such as biomarker identification and drug target discovery in cancer.