AmesFormer: State-of-the-Art Mutagenicity Prediction with Graph Transformers.

Area of Science:

• Computational toxicology
• Machine learning in drug discovery

Background:

• The Ames mutagenicity test is crucial for chemical safety assessment.
• Current in silico models often use complex ensembles and molecular fingerprints, overlooking holistic molecular structure.

Purpose of the Study:

• To introduce AmesFormer, a novel graph transformer neural network for Ames mutagenicity prediction.
• To demonstrate AmesFormer's state-of-the-art performance and accessibility for regulatory and drug development applications.

Main Methods:

• Development of AmesFormer, a graph transformer neural network.
• Creation of a new Ames dataset for model training and validation.
• Benchmarking AmesFormer against 22 other Ames models on a standardized dataset.
• Evaluation of model calibration performance using temperature scaling.

Main Results:

• AmesFormer achieved state-of-the-art (SOTA) performance in Ames mutagenicity prediction.
• The model demonstrated high performance when paired with the new Ames dataset.
• Calibration performance was assessed and improved using temperature scaling.

Conclusions:

• AmesFormer offers a high-performance, accessible, and open-source computational model for Ames mutagenicity.
• The model has significant potential to aid regulatory decision-making and accelerate drug development processes.