Identify, curate, and assess the quality and suitability of data from multiple pharmaceutical companies and other organisations within the consortium with a priority focus on general toxicology studies (14-day, 28-day, 90-day and chronic toxicity studies) in the form of structured toxicity data in CDISC SEND format as well as in unstructured study reports. Secondary focuses for additional data should also be considered, including in vivo toxicity study data, chemistry (e.g. structure information, where permitted), pharmacokinetic and exposure data, mechanistic in vitro systems, in silico models, multi-omics datasets, licensed content, prior consortia outputs, and public regulatory databases.

Develop a new database to host and analyse data, encompassing principles to support a common or federated data model that enables sensitive data preservation and multisource analysis that can allow NAM-enabled second species waivers. Additionally, sensitive data should undergo various levels of blinding from full blinding to structural alerts as a surrogate for the full structure and structural embeddings with noise introduction among other approaches.

Applicants are expected to consider the potential regulatory impact of the project’s results and develop a regulatory strategy and interaction plan for generating the required evidence to support regulatory decision-making, linked to the data sources used, as well as engaging with regulators in a timely manner for their input (e.g. national competent authorities, EMA, etc). The goal is to gain, by the end of the project, endorsement of the AI Foundation Toxicology Model from the regulatory authorities (e.g. EMA, FDA, including potential Innovative Science and Technology Approaches for New Drugs (ISTAND) submission) to enable second species waivers for chronic and sub chronic small molecule testing. The longer-term goal, beyond the action’s scope, of revising the regulatory guidance ICH M3(R2) in line with this project’s outcomes should also be taken into consideration when designing the regulatory strategy. The opportunity to broaden the scope of the AI Foundation Toxicology Model to encompass waiving single-species chronic testing beyond small molecules (e.g. oligonucleotides) should be part of the longer-term and sustainability planning.

Conduct a systematic review of existing AI and AI-supported foundation modelling approaches.

Select and optimise AI methods able to perform probabilistic predictions estimating the likelihood of novel safety-relevant information being identified in longer-term studies and/or in a second species, including risks of missed toxicity, organ-specific findings, and divergence in NOAEL. The potential to select the appropriate single species for biological medicines should also be considered in the AI method selection and optimisation.

Ensure transparency, interpretability, traceability, and data provenance to meet regulatory expectations.

Data harmonisation and representation learning

Feature extraction pipelines appropriate to each data modality (unstructured study reports, molecular structures, omics datasets, mechanistic assays).

Multimodal representation learning (e.g., contrastive learning) to create unified latent spaces across biological, toxicological, and metadata features.

Federated learning and privacy preserving techniques to enable secure, multi organisation data contributions without exposing proprietary information.

Predictive modelling approaches

Probabilistic supervised models (e.g., Bayesian neural networks, Gaussian processes, calibrated ensembles) to estimate the probability of missed toxicological findings under a single-species approach.

Mechanistic-informed modules integrated via knowledge graph neural networks, causal inference frameworks, and multi-task predictors for endpoints such as drug induced liver injury, cardiotoxicity, and genotoxicity.

Uncertainty quantification methods to ensure transparent confidence intervals suitable for risk-based decision-making.

Scenario simulation engines (Monte Carlo, generative models, causal models) to test counterfactuals relevant to second species value.

Transparency and explainability- Use of explainable AI techniques to identify key drivers of model predictions.- Comprehensive provenance tracking to ensure every model output is traceable.- Human interpretable decision layers to support weight-of-evidence narratives.

Long term evolution

Ability for continuous learning, single-species chronic testing waivers beyond small molecules and modular updates as new data types and methods emerge.

Design, train, and validate an AI Foundation Toxicology Model that produces probabilistic predictions across multiple dimensions: 1) study duration (sub-acute, 14–28 days; sub-chronic; and chronic), 2) species (rodent and non-rodent), and 3) outcome type including the likelihood of identifying a novel target organ, a significantly divergent toxicity profile and/or a significantly divergent NOAEL. For example, the model should predict the likelihood (0-100) that running a 9-month non-rodent study would identify a divergent toxicity profile, including significantly increased severity of pathology findings as compared to the toxicity profile identified in earlier, shorter-term studies or projected to be identified in the rodent 6-month study yet to be conducted. These specific outputs should be defined depending on the data provided and the AI methods, which will be established within the early stages of the project.

Industry, academic and SME stakeholders with consistent regulator engagement and input should define representative use cases and performance metrics, including accuracy, robustness, explainability, extensibility and trustworthiness, to validate the model. Formal benchmarking and rigorous testing are essential for ensuring consistency of results, to build trust in the model's recommendations.

Create a transparent, reproducible, weight-of-evidence decision support framework that integrates diverse data types and model outputs into standardised reasoning steps, including the direct model outputs (e.g., probability and significance of novel findings in a longer-term second species study) and model context (e.g., narrative rationale supporting the probabilistic output) from the AI Foundation Toxicology Model as well as orthogonal internal programme and external literature support. In vitro NAMs, target knowledge and literature mining would all help supplement and increase confidence in the AI Foundation Toxicology Model outputs and recommendations.

Define how uncertainty, evidence quality, consistency, severity, and human relevance including potential impact on or risk for the clinical programme should be evaluated, compatible with regulatory reasoning, to support second species waiver submissions.

Test the AI Foundation Toxicology Model, through use cases, together with the weight of evidence framework with relevant diverse stakeholder groups, including pharmaceutical and biotechnology companies, regulators and academics/SMEs working at the interface of drug safety research and regulation to ensure consistency across reviewers and jurisdictions.

Develop recommendations and practical tools for real-world implementation, including regulatory strategy and guideline revision as required, alignment with ethical and legal principles plus a governance structure for ongoing model evolution and ecosystem adoption. For instance, trustworthy AI, human oversight and verifications will follow regulatory frameworks such as the Assessment List for Trustworthy Artificial Intelligence (ALTAI).

Devise a sustainability and evolution plan for long-term hosting, maintenance, continuous data harmonisation and improvement of the AI Foundation Toxicology model and weight-of-evidence framework, enabling broad and sustainable ecosystem adoption.