Ligand-Based Compound Activity Prediction via Few-Shot Learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Early-stage drug discovery frequently involves predicting new compound activities using limited data from existing compounds.
• Classifying compounds as active or inactive is a common approach, but ranking compounds by affinity is more desirable.
• Few-shot learning methods have been applied to compound activity classification.

Purpose of the Study:

• To introduce Few-Shot Compound Activity Prediction (FS-CAP), a novel neural architecture for predicting compound activities.
• To enable accurate predictions for new compounds against assays using only a few known active compounds.
• To provide a method that ranks compounds by predicted affinity, enhancing drug discovery efficiency.

Main Methods:

• FS-CAP employs a neural architecture trained on a large bioactivity dataset.
• The model aggregates encodings from known compounds and their activities to capture assay-specific information.
• A separate encoder is utilized for the new compound requiring activity prediction.

Main Results:

• FS-CAP demonstrates encouraging performance in predicting compound activities.
• The method shows superiority over traditional chemical-similarity-based techniques.
• FS-CAP also outperforms other state-of-the-art few-shot learning methods in various drug discovery settings.

Conclusions:

• FS-CAP offers a powerful approach for few-shot compound activity prediction in drug discovery.
• The model's ability to rank compounds by affinity is a significant advancement.
• FS-CAP provides a valuable, data-efficient tool for identifying promising drug candidates.