Task-Similarity is a Crucial Factor for Few-Shot Meta-Learning of Structure-Activity Relationships.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Machine learning models are crucial for molecular design but require substantial training data.
• Drug discovery often faces limited data in early stages.
• Meta-learning offers a solution by leveraging existing data for new targets.

Purpose of the Study:

• To evaluate two meta-learning methods, model-agnostic meta-learning (MAML) and adaptive deep kernel fitting (ADKF), for regression tasks in drug discovery.
• To investigate the impact of dataset size and task similarity on model predictability.

Main Methods:

• Assessed MAML and ADKF in a regression setting.
• Analyzed performance based on varying dataset sizes and training task similarities.
• Compared meta-learning approaches against a single-task baseline model.

Main Results:

• ADKF significantly outperformed MAML and the single-task baseline on inhibition data.
• Model performance, particularly for ADKF, showed variability across different test tasks.
• Predictability improvements were most substantial when the target task resembled meta-learning tasks.

Conclusions:

• Meta-learning, especially ADKF, can enhance few-shot learning for molecular design.
• Task similarity is a critical factor for successful meta-learning application in drug discovery.
• Further research is needed to optimize meta-learning strategies for diverse drug discovery challenges.