Mulaqua: An interpretable multimodal deep learning framework for identifying PMT/vPvM substances in drinking water.

Area of Science:

• Environmental Chemistry
• Computational Toxicology
• Data Science

Background:

• Drinking water is threatened by chemical pollutants, including persistent, mobile, and toxic (PMT) and very persistent and very mobile (vPvM) substances.
• These emerging contaminants pose significant risks to human health, necessitating improved monitoring and management.
• Traditional detection methods are slow and resource-intensive, highlighting the need for efficient computational solutions.

Purpose of the Study:

• To develop a novel deep learning (DL) approach for rapid and economical identification of PMT/vPvM substances.
• To introduce Mulaqua, the first DL model specifically designed for this purpose.
• To provide insights into molecular characteristics influencing PMT/vPvM classification.

Main Methods:

• Mulaqua employs a multimodal DL strategy, integrating molecular string and image representations.
• Data imbalance was addressed using Simplified Molecular Input Line Entry System (SMILES) enumeration for data augmentation.
• Interpretability analyses were conducted to understand the influence of molecular structures.

Main Results:

• Mulaqua achieved high performance metrics: accuracy (ACC) of 0.920, F1-score (F1) of 0.590, and Matthews correlation coefficient (MCC) of 0.548.
• The model demonstrated excellent transferability and improved performance on external datasets compared to baselines.
• Interpretability analyses provided insights into structure-activity relationships for PMT/vPvM characterization.

Conclusions:

• Mulaqua is a pioneering, publicly available DL tool for identifying PMT/vPvM substances.
• It offers a proactive and efficient method for early hazard identification and regulatory prioritization.
• This approach can significantly enhance environmental risk management strategies for chemical contaminants.