Multimodal Multi-Granularity Fusion Model with Mamba Architecture for Ames Mutagenicity Prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Artificial intelligence in drug discovery

Background:

• Traditional Ames tests for chemical mutagenicity are time-consuming, expensive, and exhibit variability.
• Limitations in current methods hinder efficient high-throughput safety screening of chemicals.

Purpose of the Study:

• To develop a novel computational model, AMPred-LWN, for accurate and efficient Ames mutagenicity prediction.
• To overcome the limitations of traditional assays by integrating diverse molecular representations.

Main Methods:

• AMPred-LWN employs a multimodal approach, fusing atomic-level graphs, functional group sequences, and molecular fingerprints.
• The model integrates enhanced graph neural networks (GIN, GAT) with the Mamba-2 sequence architecture and a novel bidirectional ConBiMamba module.
• This architecture efficiently captures multiscale and long-range chemical features, mitigating unidirectional biases.

Main Results:

• AMPred-LWN achieved state-of-the-art performance on Ames mutagenicity prediction, with an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.922 and Accuracy (ACC) of 0.852.
• The model outperformed existing baseline methods and demonstrated strong generalization on external datasets.
• Inference time was reduced by over 30% compared to traditional methods.

Conclusions:

• AMPred-LWN provides a significant advancement in computational toxicology for Ames mutagenicity prediction.
• The model's interpretability highlights key mutagenic substructures and protective features, offering valuable structure-activity relationship insights.
• This approach accelerates safety screening and aids in the design of safer chemicals.