Few-Shot Learning for Low-Data Drug Discovery.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Ligand-based virtual screening (LBVS) in drug discovery is a low-data problem due to expensive data acquisition.
• Conventional machine learning methods require large datasets, limiting their application in hit discovery and lead optimization.

Purpose of the Study:

• To explore few-shot machine learning techniques for hit discovery and lead optimization.
• To introduce Prototypical and Relation Networks as novel metric-based meta-learning approaches.
• To evaluate different input embeddings, including extended-connectivity fingerprints (ECFP) and graph convolutional networks (GCN).

Main Methods:

• Implemented metric-based meta-learning with Prototypical and Relation Networks.
• Utilized extended-connectivity fingerprints (ECFP) and graph convolutional network (GCN) embeddings as input features.
• Assessed model performance on MUV, DUD-E, and Tox21 datasets for toxicity and LBVS tasks.

Main Results:

• Graph convolutional network embeddings consistently outperformed ECFP for toxicity and LBVS.
• Few-shot models showed variable performance on MUV and DUD-E due to structurally distinct active compounds.
• Prototypical Networks achieved state-of-the-art results on Tox21 data, outperforming Matching Networks.
• Training times were significantly reduced (up to 190%) with few-shot models.

Conclusions:

• Few-shot learning effectiveness is data-dependent; performance varies with the structural diversity of active compounds.
• Prototypical Networks offer a faster and effective approach for hit discovery on specific datasets like Tox21.
• GCN-generated embeddings show promise for improving LBVS and toxicity prediction in low-data scenarios.