A Multitask Approach to Learn Molecular Properties.

Area of Science:

• Computational chemistry
• Machine learning
• cheminformatics

Background:

• Multitask learning aims to improve model performance by leveraging shared information across related tasks.
• Deep neural networks are a common multitask framework but face challenges in consistent performance gains over single-task models.
• Resolving intertask correlations remains a key challenge in developing robust multitask models.

Purpose of the Study:

• To introduce an alternative multitask learning framework using problem transformation methods.
• To develop multitask models that improve upon the performance of independent single-target models.
• To enhance the prediction accuracy for a wide range of molecular properties.

Main Methods:

• Constructed multitask models by stacking a base regressor and classifier.
• Utilized an additional training stage on an expanded molecular feature space for multitarget predictions.
• Implemented and evaluated the model architecture on QM9, Alchemy, and Tox21 datasets using various machine learning techniques.

Main Results:

• Achieved a 1-10% enhancement in forecasting precision for molecular properties.
• Demonstrated consistent improvement in task prediction accuracy compared to independent single-target models.
• Showcased notable superiority in addressing intertarget dependence.

Conclusions:

• The proposed problem transformation framework offers a superior alternative for multitask learning in cheminformatics.
• The method effectively tackles intertarget dependence, leading to improved predictive performance.
• The framework shows significant potential for simulating diverse molecular properties.