Interpretable online network dictionary learning for inferring long-range chromatin interactions.

Area of Science:

• Computational Biology
• Genomics
• Network Science

Background:

• Dictionary learning (DL) and matrix factorization (MF) are common in computational biology for clustering.
• Existing DL methods lack interpretability and scalability for graph-structured data.

Purpose of the Study:

• Introduce online convex network dictionary learning (online cvxNDL), a novel DL algorithm.
• Address limitations of traditional DL for large, network-structured biological data.

Main Methods:

• Online convex network dictionary learning (online cvxNDL) implemented via matrix factorization.
• Application to 3D-genome RNAPII ChIA-Drop data, modeling interactions as hypergraphs.
• Incorporation of subnetwork sampling for network data.

Main Results:

• Online cvxNDL achieves accuracy comparable to classical DL methods while offering superior interpretability and scalability.
• Identified distinct proximal and distal chromatin interaction patterns across and within chromosomes.
• Successfully associated interaction patterns with RNA transcripts and inferred cellular functions using Gene Ontology enrichment analysis.

Conclusions:

• Online cvxNDL is an effective and interpretable tool for analyzing large-scale, network-structured biological data, such as 3D genome interactions.
• The method reveals novel insights into chromatin organization and its relation to gene regulation and cellular functions.