Explainable AI and the Future of In Silico Toxicology

Artificial intelligence is playing an increasing role in drug discovery, chemical research and toxicology. However, many machine learning models still work as “black boxes”, making their predictions difficult to interpret and sometimes difficult to use in industrial or regulatory contexts.

This is one of the challenges addressed by AiChemist – Explainable AI for Molecules, a Marie Skłodowska-Curie Actions Doctoral Network funded under Horizon Europe.

According to Katya Ahmad, the project focuses on two main objectives: training a new generation of researchers at the intersection of Explainable AI (XAI), chemistry and life sciences, and developing AI methods that are more accurate, interpretable and actionable.

These approaches could support several areas, including drug discovery, chemical synthesis, toxicity assessment and regulatory science.

Why Explainable AI matters in toxicology

In in silico toxicology, predictive models can help assess the potential toxicity of a chemical. But knowing whether a model predicts a substance as “toxic” or “safe” is not always sufficient.

Explainable AI goes further by helping scientists understand why a model produced a particular prediction.

For example, XAI methods may identify molecular features, reactive motifs or biological mechanisms that influence a toxicity prediction.

This additional information can make computational results easier to interpret and potentially more useful for decision-making. It can also help scientists identify structural features associated with potential risks and consider them earlier during the development of new molecules.

Supporting the development of NAMs

Explainable AI could also contribute to the development of New Approach Methodologies (NAMs).

For AI-based NAMs to support toxicological assessments, several elements are particularly important:

• a clearly defined applicability domain;

• reliable uncertainty estimates;

• and scientifically meaningful mechanistic explanations.

In practice, this means understanding which types of chemicals a model has been trained on, how confident it is in its prediction and whether the explanation behind the result makes chemical sense.

AiChemist is working on methods designed to improve model performance beyond familiar chemical space while also analysing where and why models may fail. The project is also developing open-source XAI tools that can break down toxicity predictions into chemically meaningful features.

Collaboration between science, industry and regulators

The development of AI-based NAMs also requires collaboration between academia, industry and regulatory organisations.

AiChemist brings together stakeholders from these different environments, including partners such as the FDA, NIH, ECHA and CEHTRA. This collaboration is important because the adoption of AI in toxicology will depend not only on model performance, but also on transparency, scientific interpretation and human oversight.

Rather than simply generating predictions, the aim is to develop AI tools that scientists can better understand, assess and use to support chemical safety decisions.

References

• [1] Krüger et al., 2026. SEISMO: Increasing Sample Efficiency in Molecular Optimization with a Trajectory-Aware LLM Agent.

• [2] Hunklinger & Ferruz. Towards the explainability of protein language models.

• [3] Challenges and Opportunities for Validation of AI-Based New Approach Methods. ALTEX, 2025.

• [4] Hiltscher et al., 2026. Explaining What Matters: Faithfulness in Molecular Deep Learning.

• [5] Khasanova & Tetko, 2026. Benchmarking Explainable AI Methods for Toxicophore Detection and Toxicity Prediction.

To learn more about the AiChemist project and its work on Explainable AI, molecular modelling and computational toxicology, visit aichemist.eu.