A meta-learning framework to mitigate negative transfer in transfer learning applicable to drug design

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Deep learning models struggle with sparse data, common in natural sciences like chemistry.
• Existing methods like transfer learning and meta-learning address data scarcity but are often used separately.
• Heterogeneous data distributions in fields like early-phase drug discovery pose significant challenges for machine learning.

Purpose of the Study:

• To develop a unified framework combining meta-learning and transfer learning for deep learning in cheminformatics.
• To introduce a novel meta-learning algorithm that complements transfer learning by optimizing training data selection and weight initialization.
• To mitigate the issue of negative transfer in machine learning models applied to sparse datasets.

Main Methods:

• Developed a new meta-learning algorithm to identify optimal training subsets and weight initializations.
• Integrated this meta-learning algorithm with transfer learning for a coherent framework.
• Applied the combined approach to predict protein kinase inhibitors using reduced data.

Main Results:

• Demonstrated statistically significant improvements in model performance through combined meta- and transfer learning.
• Successfully controlled for negative transfer, a common limitation of transfer learning.
• Validated the framework's effectiveness in a proof-of-concept application within cheminformatics.

Conclusions:

• The combined meta- and transfer learning framework effectively addresses data sparseness in deep learning for cheminformatics.
• This approach enhances predictive accuracy and robustly manages negative transfer.
• The method shows promise for advancing machine learning applications in data-scarce scientific domains, particularly drug discovery.