Language of Toxicity: An eXplainable Artificial Intelligence Approach
The study introduces a descriptor‑free, language‑inspired framework for predicting toxicity in small molecules, using canonical SMILES as words. It combines a multiscale CNN, a GRU, and an attention mechanism to capture chemical patterns and focus on relevant substructures. Across eight endpoints, the model achieved an average AUC of 0.83 on datasets of roughly 1,000 samples, many of which were imbalanced.
The study introduces a descriptor‑free, language‑inspired framework for predicting toxicity in small molecules, using canonical SMILES as words. It combines a multiscale CNN, a GRU, and an attention mechanism to capture chemical patterns and focus on relevant substructures. Across eight endpoints, the model achieved an average AUC of 0.83 on datasets of roughly 1,000 samples, many of which were imbalanced.
The explainable AI approach and focus on chemical safety align with Swiss AI’s interest in transparent, domain‑specific AI solutions.
Original source record. Open the original record to verify the underlying announcement.