Multitask Pretraining Framework for Improving Predictivity of Machine Learning Chemical Bioactivity Models for...

Area of Science:

• Computational toxicology
• Machine learning in drug discovery
• Chemical risk assessment

Background:

• Quantitative structure-activity relationship (QSAR) models are vital for chemical hazard screening.
• Single-task machine learning (ML) models lack transferability and require retraining for new endpoints.
• Existing models struggle with small, noisy datasets common in toxicology.

Purpose of the Study:

• To develop an ML approach for predicting chemical bioactivity on low-data toxicological endpoints.
• To create a multitask ML model that can be adapted for novel prediction tasks.
• To improve the utility of ML QSAR methods for hazard screening.

Main Methods:

• Trained a multitask ML model on multiple ToxCast datasets (moderate size).
• Combined the pretrained multitask model with task-specific predictors (random forest, neural network) for novel predictions on small datasets.
• Utilized molecular representations generated by the multitask model.

Main Results:

• The novel ML pipeline demonstrated statistically superior performance on most tasks compared to standard ML approaches.
• The multitask model's molecular representations captured generalizable information by integrating data from multiple endpoints.
• The approach effectively handled multiple small, noisy datasets.

Conclusions:

• The developed ML pipeline enhances the prediction of chemical bioactivity for under-resourced toxicological endpoints.
• Multitask learning provides a more robust and transferable approach for QSAR modeling in toxicology.
• This work advances the application of ML for efficient and reliable chemical hazard screening.