Don't Overweight Weights: Evaluation of Weighting Strategies for Multi-Task Bioactivity Classification Models

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Machine learning models are standard tools for predicting chemical compound bioactivity.
• Multi-task and federated learning enable privacy-preserving use of diverse data for better model generalization.

Purpose of the Study:

• Investigate strategies for averaging weighted task loss functions to train multi-task bioactivity classification models.
• Develop weighting strategies suitable for federated learning with diverse datasets.

Main Methods:

• Trained multi-task bioactivity classification models using large, real-world datasets from six pharmaceutical companies.
• Compared weighting strategies based on sub-task count, task size, and class balance.

Main Results:

• A simple sub-task weighting approach demonstrated robust model performance across all datasets.
• This method is particularly well-suited for federated learning scenarios.

Conclusions:

• Sub-task weighting is an effective strategy for training multi-task bioactivity models.
• This approach supports privacy-preserving, high-performance drug discovery using federated learning.