Abstract
Explainable machine learning for molecular toxicity prediction is a promising approach for efficient drug development and chemical safety. A predictive ML model of toxicity can reduce experimental cost and time while mitigating ethical concerns by significantly reducing animal and clinical testing. Herein, we use a deep learning framework for simultaneously modeling in vitro, in vivo, and clinical toxicity data. Two different molecular input representations are used; Morgan fingerprints and pre-trained SMILES embeddings. A multi-task deep learning model accurately predicts toxicity for all endpoints, including clinical, as indicated by the area under the Receiver Operator Characteristic curve and balanced accuracy. In particular, pre-trained molecular SMILES embeddings as input to the multi-task model improved clinical toxicity predictions compared to existing models in MoleculeNet benchmark. Additionally, our multitask approach is comprehensive in the sense that it is comparable to state-of-the-art approaches for specific endpoints in in vitro, in vivo and clinical platforms. Through both the multi-task model and transfer learning, we were able to indicate the minimal need of in vivo data for clinical toxicity predictions. To provide confidence and explain the model's predictions, we adapt a post-hoc contrastive explanation method that returns pertinent positive and negative features, which correspond well to known mutagenic and reactive toxicophores, such as unsubstituted bonded heteroatoms, aromatic amines, and Michael receptors. Furthermore, toxicophore recovery by pertinent feature analysis captures more of the in vitro (53%) and in vivo (56%), rather than of the clinical (8%), endpoints, and indeed uncovers a preference in known toxicophore data towards in vitro and in vivo experimental data. To our knowledge, this is the first contrastive explanation, using both present and absent substructures, for predictions of clinical and in vivo molecular toxicity.
Introduction
Toxicity is a primary reason drug candidates fail during clinical trials, leading to billions in lost R&D costs. Traditionally, safety assessment relies on a hierarchy of testing: starting with in vitro (cell-based) assays, moving to in vivo (animal) models, and finally clinical trials in humans. However, animal models do not always translate to human biology, and there is a global ethical push to minimize animal testing. Furthermore, most existing Machine Learning (ML) models for toxicity are "black boxes." This paper aims to bridge the gap between different testing platforms while making the AI's reasoning transparent to chemists.
Methodology
The authors employed several strategies to improve performance and transparency in toxicity prediction.
Molecular Representations
The study compared standard Morgan Fingerprints against SMILES Embeddings (SE). The SEs were generated using a pre-trained neural network that learned chemistry by translating between different molecular formula formats, capturing deeper relationships between chemicals.
Multi-Task Deep Neural Networks (MTDNN)
Instead of training for a single endpoint, a multi-task architecture was used to allow "knowledge transfer" between tasks. For example, insights from in vitro assays (Tox21 dataset) can help predict clinical trial outcomes (ClinTox dataset). The model also incorporated in vivo oral toxicity data (RTECS).
Contrastive Explanations Method (CEM)
To explain predictions, the researchers used CEM to identify:
- Pertinent Positives (PP): Minimal chemical substructures present that cause a "toxic" prediction.
- Pertinent Negatives (PN): Minimal changes or missing substructures that would flip the prediction to "nontoxic."
Results and Discussion
Predictive Accuracy
On the ClinTox clinical dataset, the MTDNN-SE model achieved an AUC-ROC of 0.994, a significant improvement over the MoleculeNet baseline of 0.832. The model also achieved a balanced accuracy of ~0.96, demonstrating robustness despite skewed datasets.
The "In Vivo" Surprise
A significant finding was that adding in vivo animal data did not materially improve clinical toxicity predictions when the model already utilized in vitro data and SMILES embeddings. This suggests a potential path toward reducing reliance on animal testing without compromising safety.
Explaining the "Why"
CEM analysis showed the model learned real chemistry. Features flagged as "Pertinent Positives" matched known toxicophores, such as aromatic amines and Michael receptors.
Conclusion
This multi-task framework demonstrates that clinical toxicity can be predicted effectively with minimal animal data. While limited by current dataset sizes and a lack of 3D spatial reasoning, the model provides a transparent and accurate tool for drug safety assessment.
Summary updated