Multi-Label Models for Toxicity Prediction

In silico predictive models of chemical toxicity often utilizes single-label machine learning models. However, these models lack the ability to learn possible dependencies between toxicity labels. This project explores the use of multi-label strategies (image) as a wrapper for machine learning models for toxicity prediction. Additionally, we explore methods such as Bayesian networks and Louvain community detection for learning label partitions from pair-wise label dependencies, in order to circumvent the random label partitioning method commonly used with Classifier Chains and Label Powersets.