Algebraic shortcuts for leave-one-out cross-validation in supervised network inference.

Area of Science:

• Computational Biology
• Machine Learning
• Network Inference

Background:

• Supervised machine learning excels at reconstructing biological networks (e.g., protein-protein interactions, gene regulatory networks).
• Accurate model evaluation is crucial, distinguishing between predicting new interactions within a network versus predicting interactions for new biological entities.
• Hyperparameter tuning depends on the specific prediction setting, necessitating appropriate cross-validation schemes.

Purpose of the Study:

• To present an efficient kernel-based network inference technique: two-step kernel ridge regression.
• To demonstrate the computational efficiency of this model, with time complexity scaling with the number of vertices.
• To introduce cross-validation shortcuts for rapid performance estimation across various network prediction settings.

Main Methods:

• Utilizing a state-of-the-art kernel-based network inference technique: two-step kernel ridge regression.
• Analyzing time complexity to show efficiency scaling with the number of vertices.
• Developing cross-validation shortcuts for performance assessment.

Main Results:

• The two-step kernel ridge regression model trains efficiently, with time complexity dependent on the number of vertices, not edges.
• The framework provides cross-validation shortcuts for rapid performance estimation in different network prediction scenarios.
• Enables computational biologists to thoroughly assess model capabilities.

Conclusions:

• The presented method offers an efficient and versatile approach to biological network inference.
• The developed cross-validation shortcuts facilitate robust model evaluation and hyperparameter tuning.
• The RLScore software package implements these techniques for practical application.