A graph neural network approach for molecule carcinogenicity prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Identifying molecular carcinogenicity is crucial for cancer prevention but hindered by costly, low-throughput animal testing.
• Limited carcinogenicity data presents a significant challenge for developing accurate predictive models.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for predicting molecular carcinogenicity from chemical structures.
• To address data scarcity issues in carcinogenicity prediction through advanced machine learning techniques.

Main Methods:

• Proposed CONCERTO, a deep learning model integrating graph transformers and molecular fingerprints for structure-based carcinogenicity prediction.
• Employed multi-round pre-training on mutagenicity data and transfer learning from self-supervised models to overcome data size constraints.

Main Results:

• CONCERTO demonstrates strong performance and generalizability on external validation datasets.
• The model effectively predicts carcinogenicity from molecular structure, offering a viable alternative to traditional methods.

Conclusions:

• CONCERTO provides a powerful tool for predicting molecular carcinogenicity, potentially guiding experimental design and reducing costs.
• The model offers insights into the molecular mechanisms underlying carcinogenicity, aiding in the development of safer chemicals.