Multi-label classification and label dependence in in silico toxicity prediction

Area of Science:

• Computational toxicology
• Cheminformatics
• Machine learning

Background:

• Current predictive toxicology models often focus on single endpoints.
• These models lack the ability to learn dependencies between related toxicity endpoints.
• Understanding label dependencies is crucial for improving predictive accuracy.

Purpose of the Study:

• To compare the performance of different multi-label classification (MLC) models against independent classifiers using Tox21 challenge data.
• To develop a novel measure for quantifying label dependencies to enable data-driven label partitioning.
• To investigate the impact of label partitioning on MLC model performance.

Main Methods:

• Compared Classifier Chains (CC), Label Powerset (LP), and Stacking (SBR) MLC models against Binary Relevance (BR).
• Developed a quantitative label dependence measure and combined it with Louvain community detection for data-driven label partitioning.
• Utilized Logistic Regression and Random Forest as base classifiers with random and learned label partitioning strategies.

Main Results:

• Classifier Chains (CC) and Stacking (SBR) models showed statistically significant improvements in performance metrics (Hamming loss, multi-label accuracy).
• Model weights correlated positively with label dependencies, highlighting the importance of learning these relationships.
• MLC models with learned label partitioning were generally non-inferior to those with random or no partitioning.
• A Stacking model using Random Forest outperformed its base model in 11 out of 12 Tox21 labels.

Conclusions:

• Multi-label classification (MLC) models offer a promising approach to enhance the performance of computational toxicology predictions.
• Learning label dependencies is a key factor in improving MLC model accuracy.
• Data-driven label partitioning is a viable alternative to random partitioning for MLC model development.