Quantifying the Hardness of Bioactivity Prediction Tasks for Transfer Learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioactivity prediction

Background:

• Machine learning (ML) is crucial in drug discovery but hindered by data scarcity.
• Knowledge-sharing ML strategies like transfer learning, multitask learning, and meta-learning address data limitations by leveraging related tasks.
• A key challenge is understanding how source task relatedness impacts the performance of these ML models.

Purpose of the Study:

• To develop a new method for quantifying and predicting the hardness of bioactivity prediction tasks.
• To assess the relationship between task hardness and the performance of knowledge-sharing ML approaches.

Main Methods:

• Generated protein and chemical representations.
• Calculated distances between the target bioactivity prediction task and available training tasks to quantify task hardness.
• Applied the metric within a meta-learning framework on the FS-Mol dataset.

Main Results:

• Demonstrated an inverse correlation between the proposed task hardness metric and model performance (Pearson's r = -0.72) in a meta-learning scenario.
• The developed metric effectively quantifies the difficulty of a bioactivity prediction task relative to existing training data.

Conclusions:

• The new task hardness metric is a valuable tool for estimating performance improvements achievable with meta-learning in drug discovery.
• This metric can guide the application and development of knowledge-sharing ML strategies to overcome data scarcity challenges.