Benchmarking AI to predict mutagenicity by combing chemical structure and genome-wide mutation data

Project Information

MUTAPREDICT

Grant agreement ID: 101189234

DOI

10.3030/101189234

Project closed

EC signature date 26 September 2024

Start date 1 October 2024

End date 31 March 2026

Funded under

• European Research Council (ERC)

Total cost

No data

EU contribution

€ 150 000,00

Coordinated by

PRINSES MAXIMA CENTRUM VOOR KINDERONCOLOGIE BV
Netherlands

Periodic Reporting for period 1 - MUTAPREDICT (Benchmarking AI to predict mutagenicity by combing chemical structure and genome-wide mutation data)

Reporting period: 2024-10-01 to 2026-03-31

Summary of the context and overall objectives of the project

Genotoxicity and mutagenicity assessment are essential components of drug development and chemical safety evaluation. Mutagenic compounds can induce DNA damage and genomic instability, potentially leading to cancer, developmental disorders and other severe adverse health effects. Regulatory authorities therefore require extensive genetic toxicology testing before pharmaceutical compounds and other chemicals can progress toward clinical application or market approval.
Current mutagenicity assessment strategies rely heavily on animal experiments and conventional in vitro assays, including the Ames test, micronucleus assay and mouse lymphoma assay. Although widely used, these methods have important limitations. Many assays show limited predictive power for human-relevant mutagenicity, are resource-intensive and time-consuming, and often fail to capture the complexity of mutational processes occurring in human tissues. At the same time, scientific, societal and regulatory pressure is increasing to reduce reliance on animal testing and develop more predictive human-relevant alternatives.
Recent advances in whole-genome sequencing, stem-cell technologies and artificial intelligence (AI) create new opportunities for next-generation predictive toxicology. Large-scale mutational datasets generated in human cells now enable direct characterization of mutation patterns induced by chemical exposure. However, these datasets remain underutilized in existing computational mutagenicity prediction tools, which are largely based on conventional quantitative structure–activity relationship approaches and often lack sufficient predictive performance across structurally diverse compounds.
The ERC Proof of Concept project MUTAPREDICT was initiated to address this unmet need by translating previous ERC-funded discoveries into a practical and commercially relevant mutagenicity prediction platform. The overall objective was to develop and validate an AI-based, non-animal mutagenicity prediction platform integrating human-relevant genomic mutational data with advanced machine-learning methodologies.
The project aimed to optimize AI-based mutagenicity prediction methodologies, identify informative molecular representations, explore chemical space across diverse pharmaceutical and chemical compounds, and develop a user-friendly software interface for practical implementation. In addition, the project focused on positioning the technology within the pharmaceutical and biotechnology market, establishing a long-term intellectual-property strategy, and validating the commercial feasibility of the platform.
MUTAPREDICT contributes to the emerging field of AI-supported human-relevant predictive toxicology by demonstrating the value of integrating genomic mutational signatures into mutagenicity prediction workflows. The platform has the potential to reduce dependence on animal testing and support safer and more efficient development of pharmaceuticals and chemicals. The project also established a commercialization roadmap targeting pharmaceutical companies, biotechnology companies and contract research organizations involved in preclinical safety assessment and compound development.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

Activity 1 – Optimize prediction software (technical activity)
After benchmarking machine learning approaches, a Neural Network was selected as optimal for mutagenicity prediction. Due to high structural diversity across public WGS data and in-house single-cell WGS data, a single unified model was developed for all compound classes and demonstrated improved prospective prediction of novel compounds.

Activity 2 – Explore chemical space (technical activity)
A training dataset was built from Ames, micronucleus (MN), and mouse lymphoma assay (MLA) compounds. Data were deduplicated and SMILES structures validated. Compounds were clustered to account for structural classes. This produced specialized models optimized for structurally related compound groups.
To improve human relevance, transfer learning was applied using public and in-house WGS mutagenicity data from hematopoietic stem cells exposed to chemicals.

Activity 3 – Market positioning
A structured analysis defined target users, competitive landscape, and key use cases. Focus was placed on pharmaceutical companies, biotech firms, and CROs, identifying value in preclinical safety decision support. MUTAPREDICT was positioned relative to existing in silico and genetic toxicology tools, highlighting its differentiation as an AI-driven, non-animal mutagenicity prediction platform. Primary applications include early drug development, compound triage, and design optimization.

Activity 4 – IP strategy development
An IP audit and patent landscape analysis were conducted to identify background IP and protectable innovations. Key assets include algorithms, data-processing workflows, and software architecture. The resulting strategy prioritizes building a strong, commercially relevant IP portfolio supporting future patent filings, international protection, and exploitation via spin-out, licensing, or partnerships.

Activity 5 – Business strategy development
A business strategy assessment evaluated market demand, customer needs, commercial models, and entry routes. Insights from market and IP analyses were integrated into a coherent business plan. The outcome is a validated commercialization pathway for MUTAPREDICT, with defined target users, value proposition, and a roadmap for market entry, growth, and scaling.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

MUTAPREDICT advances beyond the state of the art through the integration of human whole-genome sequencing (WGS)-based mutational datasets with advanced artificial-intelligence methodologies, particularly Graph Neural Networks (GNNs). Rather than relying solely on conventional molecular descriptors, the developed platform directly interprets molecular structural information and combines this with transfer-learning strategies using human mutational data. This approach substantially improves predictive performance and human relevance compared to currently available QSAR-based tools.
A major innovation of the project is the demonstration that transfer learning using human stem-cell mutational datasets improves classification of both mutagenic and non-mutagenic compounds. The project further demonstrated that NN-based approaches can successfully generalize across chemically diverse compound classes while maintaining robust predictive performance on previously unseen compounds.

The project also moves beyond the state of the art by combining:
*AI-driven mutagenicity prediction,
*human-relevant genomic mutational data,
*large-scale curated chemical databases,
*high-throughput prediction capabilities,
*and interpretable visualization of mutagenic molecular substructures within a single integrated platform.

The development of a user-oriented interface would further increases accessibility of advanced predictive toxicology methodologies to researchers and industrial users without extensive computational expertise, facilitating broader adoption in pharmaceutical and biotechnology environments.