Super Complex

Super Complex detects molecular complexes in protein-interaction networks using a supervised AutoML-based machine learning pipeline that learns community fitness functions to identify protein complexes.

Key Features:

• Supervised AutoML-based pipeline: Employs a distributed supervised machine learning pipeline with Automated Machine Learning to learn a community fitness function from known communities.
• Overlapping community detection: Detects overlapping communities within weighted networks to model proteins participating in multiple complexes.
• Heuristic local search algorithm: Uses a heuristic local search to identify maximally scoring communities according to the learned fitness function.
• Parallel implementation and scalability: Implements parallel execution suitable for computer clusters to handle large networks with thousands of nodes and edges.
• Novel evaluation measures: Introduces three novel evaluation measures tailored to compare learned communities against known communities.
• Generalizability and transfer learning: Supports transfer learning and generalization across datasets, enabling application when known communities are unavailable.
• Domain-specific feature integration: Allows incorporation of domain-specific features to enhance community detection.

Scientific Applications:

• Yeast network benchmarking: Demonstrated superior performance in a yeast protein-interaction network, outperforming six other supervised and four unsupervised methods.
• Human protein-interaction network analysis: Applied to a human network of approximately 8,000 nodes and 60,000 edges, identifying 1,028 protein complexes, including 234 complexes linked to SARS-CoV-2.
• Discovery of uncharacterized proteins: Identified 111 uncharacterized proteins in 103 learned complexes, supporting discovery of novel biological insights and viral interaction hypotheses.

Methodology:

Uses a distributed supervised AutoML pipeline to learn a community fitness function from known communities, applies a heuristic local search to find maximally scoring overlapping communities in weighted networks, supports parallel execution on computer clusters, employs three novel evaluation measures, and enables transfer learning and incorporation of domain-specific features.