Methods

Overview

The pipeline comprises the following steps:

• Data Prep: Cleaning known complex data; splitting into train and test sets; construction of train, test negative complex data
• Feature Extraction: Computation of topological characteristics of all complexes to construct train, test feature matrices.
• ML Classifier: Comparison of different ML pipelines using tpot, training and testing with the best ML model.
• Sampling: 4 sampling methods using growth strategies from seed nodes. Selection of the node to add at each growth step is done by evaluating the ML model score for the complex with this new node.

Although the pipeline broadly follows the same major steps as other supervised techniques, each step has been further optimized in an attempt to yield better solutions. Examples include train-test splits ensuring equal size distribution in the 1st step, an auto-ML pipeline performing effective feature selection and ML model selection in the 2nd and 3rd steps and exploration strategies incorporated into the sampling algorithms in the 4th step.