In the ever-evolving landscape of drug discovery, understanding how a drug behaves in the body is crucial. In our latest blog, Dr Angelo Pugliese explores the pivotal role ADMETox plays in this process.

ADMETox stands for Absorption, Distribution, Metabolism, Excretion, and Toxicity, and it plays a crucial role in drug discovery. These studies help researchers understand how a drug behaves in the body, including how well it is absorbed into the bloodstream, how it is distributed throughout the body, how it is metabolised or broken down, how quickly it is eliminated, and whether it has any harmful effects. By evaluating these properties early in the development process, pharmaceutical companies, biotechs, and Contract Research Organisations (CROs), such as BioAscent, can identify potential issues with a drug candidate, leading to better decision-making and reducing the risk of failure in later stages of drug development. This methodical approach not only improves the chances of developing safe and effective drugs but also ensures compliance with regulatory requirements necessary for approval.

Given the critical nature of these ADMETox studies, the landscape of drug discovery is undergoing a significant evolution to improve how they are performed. This change is driven by the priority for faster, more accurate, and ethically sound methods, particularly in ADMETox prediction. Assessing a compound's absorption, distribution, metabolism, excretion, and toxicity is critical, but traditional experimental approaches, often involving extensive testing, are time-consuming and resource-intensive. Leveraging advances in computational chemistry, machine learning, with tools like Chemprop, the collaborative model of federated learning, and emerging technologies like Large Language Models (LLMs), offers a powerful alternative, with the potential to significantly reduce the need for time-consuming studies.

Powerful tools for molecular property prediction, guided by expertise

In recent years, machine learning (ML) has transformed molecular property prediction, a crucial aspect for chemical and ADMETox research. This field has seen the rise of specialised models like Chemprop and the emergence of a new class of AI known as foundation models, which offer significant advantages in flexibility and scalability.

Foundation models (https://en.wikipedia.org/wiki/Foundation_model), originally popularised in natural language processing (NLP), are large-scale AI systems pre-trained on vast datasets. In chemistry, foundation models such as Molformer and ChemBERTa have been developed to process molecular data, leveraging transformer-based architectures to learn general-purpose representations. These models excel because they can be fine-tuned on specific tasks, including ADMETox predictions. Fine-tuning allows foundation models to adapt to new datasets with minimal labelled data while keeping the knowledge acquired during pretraining. This makes them highly flexible and capable of generalising across diverse chemical space - an area where task-specific tools like Chemprop may struggle when faced with out-of-distribution data or limited training examples.

While foundation models provide broad applicability, tools like Chemprop remain highly effective for specific prediction tasks. Chemprop uses Directed Message Passing Neural Networks (D-MPNNs) to achieve high accuracy, particularly on small-to-medium datasets such as Tox21 (a US collaborative programme using high-throughput screening to test thousands of chemicals for toxicity) and ClinTox (a dataset used for machine learning, containing drugs labelled for toxicity based on whether they were FDA-approved or failed human clinical trials due to toxic effects). However, unlike foundation models, Chemprop requires task-specific training from scratch and does not benefit from large-scale pretraining, making it less adaptable to new or diverse datasets.

For researchers focusing on graph-based molecular representations, Graphormer combines the power of transformers with graph neural networks, making it particularly adept at capturing structural features critical for accurate property prediction. For simpler or more traditional approaches, QSAR (Quantitative Structure-Activity Relationship) models, implemented through tools like KNIME or Scikit-learn, remain effective when working with well-curated datasets.

In the end, the choice between these tools depends on, for example, dataset size, computational resources, skills/expertise (using a codeless tool like KNIME is easier than fine-tuning a foundation model), and the complexity of the prediction task. Foundation models like Molformer and ChemBERTa stand out for their ability to generalise across tasks and adapt through fine-tuning, offering a significant advantage over task-specific tools like Chemprop in many situations. However, Chemprop’s simplicity and strong performance on smaller datasets make it a reliable option for targeted predictions. Whether leveraging the adaptability of foundation models or the specialised capabilities of Chemprop, ML tools are accelerating our ability to predict molecular properties and advancing innovation in chemistry and drug discovery.

While tools like Chemprop can be highly effective, their success and indeed the success of more complex foundation models, rely heavily on the quality of the data provided and the expertise of scientists in interpreting and validating the results. This underscores a central challenge for advancing ML-driven ADMETox prediction: accessing the large, diverse datasets required for optimal model performance. Pharmaceutical companies and research institutions, including those within SULSA (Scottish Universities Life Sciences Alliance), often hold valuable datasets, but privacy regulations and concerns can prevent direct sharing. To address this critical gap, federated learning provides a vital framework, allowing models to learn collaboratively from decentralised data without exposing sensitive information.

Federated Learning and the evolution beyond MELLODDY

Federated learning is a decentralised machine learning approach that allows multiple participants to collaboratively train a model without sharing their raw data. Instead, each participant trains a local model on their own data and shares only the model updates with a central server. The server then aggregates these updates to create a global model, distributed back to the participants. This process is repeated iteratively, improving the global model over time. This approach offers several advantages: data privacy, handling diverse datasets, collaborative power, and ethical advancement. By providing robust in silico ADMETox predictions, federated learning combined with Chemprop can reduce the need for in vitro and in vivo testing.

The MELLODDY consortium (https://pmc.ncbi.nlm.nih.gov/articles/PMC11005050/) stands as a landmark example of federated learning applied to ADMETox prediction. This pioneering project, bringing together pharmaceutical companies and research institutions, demonstrated the feasibility of training robust ADMETox models without centralising sensitive data, paving the way for future applications. Building upon the foundations laid by MELLODDY, companies like Apheris are now providing the technological infrastructure to further advance federated learning in ADMETox prediction. This technology creates secure data collaboration ecosystems, allowing organisations to work together on sensitive data without direct data sharing. The know-how is being used to facilitate new ADMET and other consortia like the AI Structural Biology Consortium (AISB), for example, allowing pharmaceutical companies and research institutions to train advanced ML models on distributed datasets while maintaining data privacy. These companies provide the tech layer that enables these consortia to function, and therefore, to advance drug discovery.

While the need to protect sensitive data and intellectual property within these consortia is clear, restricting access to the developed ADMETox models solely to members represents a potential bottleneck for maximising their scientific and societal impact. For progress in drug discovery to accelerate most effectively, there is a compelling case for making these powerful predictive tools, or at least versions of them, openly accessible to the wider scientific community. Freely available models would empower academic researchers, CROs, smaller biotech companies, and scientists in resource-limited settings, nurturing broader innovation, enabling independent validation, and ensuring that the benefits derived from collaborative efforts on sensitive data ultimately contribute more widely to the development of safer, more effective drugs for the public good.

Multitask Learning and Large Language Models in ADMETox

Further advancements like Multitask Learning, which enhances predictions across multiple related endpoints, and the application of Large Language Models (LLMs) for processing diverse chemical and textual data, are also transforming ADMETox prediction, each bringing distinct advantages that, when combined, are transforming how we approach drug discovery.

Multitask learning allows us to build models that predict multiple ADMETox endpoints at once, leveraging the inherent relationships between properties like solubility, metabolic stability, and toxicity. By learning these connections together, MTL models often achieve higher accuracy, especially for endpoints with limited experimental data. This approach is already being adopted at scale in pharmaceutical R&D, where it enables more comprehensive compound profiling and reduces costly late-stage failures.

At the same time, large language models are unlocking the value of both chemical and textual data. LLMs can process and interpret everything from molecular structures to scientific literature, automating the extraction of experimental insights and helping to standardise diverse datasets. Their ability to reason across chemical and biomedical domains makes them invaluable for everything from predicting molecular properties to summarising safety data and even suggesting new research directions.

Together, these technologies are setting the stage for the next generation of ADMETox prediction. As MTL and LLMs continue to advance, we can expect even tighter integration between computational and experimental workflows, faster and more accurate compound screening, and ultimately, safer and more effective medicines reaching patients sooner.

The CRO role in the evolving ADMETox landscape

Contract Research Organisations (CROs), such as BioAscent, remain indispensable partners in drug discovery through their specialised expertise and experimental capabilities. As advanced computational methods transform ADMETox prediction, the CRO's role evolves. While typically operating on customer-specific data and thus not direct data contributors to large, federated learning consortia building foundational models, CROs are uniquely positioned to become expert users and appliers of these powerful tools, particularly if the sophisticated models generated by consortia are made openly accessible to the scientific community.

Access to a state-of-the-art predictive model (whether based on Chemprop, federated learning, or future LLM integration) would enable CROs to significantly enhance their service offerings. Their value would lie in applying these public or client-provided models to specific project compounds and, crucially, integrating in silico predictions with the rigorous experimental validation services they already provide. This synergy, combining advanced prediction with real-world testing, offers clients a more comprehensive, efficient path to understanding compound profiles and making data-driven decisions.

By expertly navigating and applying these computational advancements (when available) and grounding them with empirical data, CROs can accelerate client programs, reduce the need for extensive in vivo testing, and ultimately contribute significantly to the development of safer, more effective medicines.

To read more about integrated drug discovery services from BioAscent, click here.