Gene regulatory network inference as relaxed graph matching.

Area of Science:

• Computational Biology
• Network Science
• Genomics

Background:

• Bipartite network inference is crucial in molecular biology, particularly for gene regulatory network (GRN) inference.
• GRNs are vital for understanding disease mechanisms like cancer, but are often studied using noisy network projection data.
• Accurate inference of these networks from limited, noisy observations remains a significant challenge.

Purpose of the Study:

• To develop a robust method for estimating gene regulatory networks from noisy projected observations.
• To introduce OTTER, a novel optimization framework for bipartite network inference.
• To provide both spectral and gradient-descent based algorithms for solving the OTTER problem.

Main Methods:

• Formulated OTTER as a non-convex, analytically tractable optimization problem interpretable as relaxed graph matching.
• Derived explicit solutions for OTTER, inspiring a spectral algorithm with network recovery guarantees.
• Developed a gradient descent approach for OTTER, demonstrating enhanced noise robustness compared to the spectral method.

Main Results:

• OTTER's gradient descent approach showed similarities to the PANDA gene regulatory network inference method.
• Evaluated on three cancer-related datasets, OTTER outperformed state-of-the-art methods in predicting transcription factor binding.
• Network inference accuracy was improved by the proposed OTTER framework.

Conclusions:

• OTTER provides a powerful and flexible framework for bipartite network inference, particularly for gene regulatory networks.
• The method offers improved accuracy in predicting transcription factor binding, with implications for cancer research.
• Publicly releasing data and networks aims to foster further research in graph matching for biological network inference.