A small-scale data driven and graph neural network based toxicity prediction method of compounds.

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• Toxicity prediction is vital for drug discovery, but limited data poses a challenge.
• Data-driven models offer an efficient alternative to traditional experimental methods.
• Graph Neural Networks (GNNs) show promise for chemical property prediction.

Purpose of the Study:

• To develop a small-scale, data-driven toxicity prediction method using GNNs.
• To enhance prediction accuracy through joint learning across multiple toxicity types.
• To leverage transfer learning for reliable predictions with limited data.

Main Methods:

• Proposed a joint learning strategy for multiple toxicity types.
• Constructed a graph-based model named JLGCN-MTT.
• Integrated transfer learning to utilize data from various toxicity types.

Main Results:

• JLGCN-MTT outperformed traditional machine learning and single-task GNNs on 12 toxicity prediction tasks.
• Area Under the Curve (AUC) improved by over 10% in 11 tasks.
• Significant AUC improvements (up to 11%) were observed with small training datasets (50-300 samples).

Conclusions:

• The proposed JLGCN-MTT method achieves high accuracy in small-scale toxicity prediction.
• This approach is effective even when data for specific toxicity types is scarce.
• The findings support the use of data-driven GNNs for efficient and reliable toxicity assessment in drug discovery.