3MTox: A motif-level graph-based multi-view chemical language model for toxicity identification with deep...

Area of Science:

• Computational toxicology
• cheminformatics
• machine learning

Background:

• Toxicity identification is crucial for human health, protecting against chemical hazards.
• Experimental toxicity testing is slow and expensive.
• Computational methods, including machine learning (ML) and deep learning (DL), offer faster alternatives but face challenges like feature reliance and overfitting.

Purpose of the Study:

• To propose a novel, high-performance computational model for toxicity identification.
• To address limitations of existing ML/DL methods in toxicity prediction.

Main Methods:

• Development of a motifs-level graph-based multi-view pretraining language model named 3MTox.
• Utilizing Bidirectional Encoder Representations from Transformers (BERT) as the core framework.
• Employing motif graphs as input for the model.

Main Results:

• 3MTox achieved state-of-the-art performance on benchmark toxicity datasets.
• The model outperformed existing baseline methods in toxicity prediction.
• Demonstrated interpretability by accurately identifying specific toxicity sites within molecules.

Conclusions:

• 3MTox presents a promising advancement in computational toxicity identification.
• The model's performance and interpretability contribute to better toxicity assessment and analysis.
• Offers a powerful tool for early hazard detection from chemical compounds.