3Br-MGD: few-shot toxicity prediction with a three-branch deep encoder and meta-learning framework

Area of Science:

• Computational toxicology
• Drug discovery
• Machine learning in pharmacology

Background:

• Accurate prediction of pharmaceutical compound toxicity is crucial for drug development.
• Existing computational methods face challenges with complex molecular data and limited training samples.
• Early toxicity assessment enhances patient safety and reduces drug development costs.

Purpose of the Study:

• To develop a novel deep learning framework for enhanced molecular toxicity prediction.
• To address limitations in current methods regarding complex molecular characteristics and low-data scenarios.
• To improve the accuracy, robustness, and generalization of toxicity predictions.

Main Methods:

• Proposed a three-branch deep learning framework (3Br-MGD) integrating meta-learning.
• Utilized complementary molecular representations: Morgan fingerprints (FingerprintMLP), molecular graphs (GCNs), and SMILES strings (1D-CNNs).
• Employed a Prototypical Network-based few-shot learning approach for rapid adaptation to new tasks with limited data.

Main Results:

• 3Br-MGD consistently outperformed conventional baselines on benchmark toxicity datasets.
• Demonstrated superior predictive accuracy, robustness, and generalization capabilities.
• Showcased reduced dependence on large datasets and enhanced interpretability through integrated molecular views.

Conclusions:

• The 3Br-MGD framework offers a significant advancement in computational toxicology for drug discovery.
• Its ability to leverage diverse molecular representations and few-shot learning improves toxicity prediction efficiency and reliability.
• This approach accelerates the identification of safer drug candidates and optimizes the drug development pipeline.