Related Experiment Video
Updated: Apr 4, 2026
Constructing and Visualizing Models using Mime-based Machine-learning Framework
Published on: July 22, 2025
A pipeline for developing AI-driven models to predict molecular initiating events: a case study on neural tube
Job H Berkhout1,2, Merel Florian1,3, Domenico Gadaleta4
1Centre for Health Protection, National Institute for Public Health and the Environment, Bilthoven, The Netherlands.
Abstract:
Adverse Outcome Pathways (AOPs) describe the sequence of molecular and cellular events that lead to toxicity. Each pathway begins with a Molecular Initiating Event (MIE) and ends in an Adverse Outcome. Early identification of chemical activity on MIE-relevant protein targets supports first-line toxicity assessment and helps researchers prioritize mechanisms for subsequent experimental investigation. Here we present an automated AI pipeline that converts raw ChEMBL bioactivity data into optimized deep learning models for MIE prediction. The pipeline builds on the Knowledge-Guided Pre-training of Graph Transformer (KPGT) framework, which represents chemical structures as knowledge-enriched molecular graphs. It integrates data curation, molecular graph generation, and model training and tuning. This integration enables users to construct target-specific prediction models in a seamless and reproducible way, starting from initial data and ending with deployable AI. We demonstrate its use in a neural tube defect (NTD) case study, where fine-tuned KPGT models outperformed traditional Support Vector Machine models with a radial basis function kernel (SVM-RBF) when predicting MIEs linked to developmental toxicity. The results highlight the potential of AI-driven toxicity modeling to accelerate AOP development, improve endpoint prioritization, and prioritize chemicals for experimental follow-up. By providing an end-to-end, data-to-model workflow, the pipeline lowers the technical barrier to using modern graph-based neural architectures in toxicology. It offers a reproducible route to deployable MIE prediction models that support AOP development, compound prioritization, and early-stage chemical safety evaluation.
Insights
An AI pipeline automates the creation of deep learning models for predicting Molecular Initiating Events (MIEs), accelerating toxicity assessments and Adverse Outcome Pathway (AOP) development.
Area of Science:
- Toxicology
- Computational Chemistry
- Artificial Intelligence
Background:
- Adverse Outcome Pathways (AOPs) link molecular events to toxicity.
- Early identification of chemical interactions with Molecular Initiating Events (MIEs) is crucial for toxicity assessment.
- Current methods for MIE prediction can be resource-intensive.
Purpose of the Study:
- To develop an automated AI pipeline for generating deep learning models for MIE prediction.
- To streamline the process of converting raw bioactivity data into deployable predictive models.
- To facilitate the acceleration of AOP development and chemical safety evaluations.
Main Methods:
- An automated pipeline integrating data curation, molecular graph generation, and model training was developed.
- The pipeline utilizes the Knowledge-Guided Pre-training of Graph Transformer (KPGT) framework for knowledge-enriched molecular graph representation.
- Deep learning models were trained and tuned for MIE prediction using ChEMBL bioactivity data.
Main Results:
- The AI pipeline successfully converted raw bioactivity data into optimized deep learning models for MIE prediction.
- Fine-tuned KPGT models demonstrated superior performance compared to traditional Support Vector Machine models in a neural tube defect case study.
- The pipeline enables seamless and reproducible construction of target-specific prediction models.
Conclusions:
- The AI pipeline offers an end-to-end workflow for MIE prediction, lowering the barrier to using advanced graph neural networks in toxicology.
- This approach accelerates AOP development, improves endpoint prioritization, and aids in prioritizing chemicals for experimental follow-up.
- The developed models support early-stage chemical safety evaluation and compound prioritization.
