Transforming drug safety with AI and predictive modeling

In this interview, News Medical-Life Sciences speaks with James McDonagh, Principal AI Scientist at ApconiX. Drawing on experience spanning computational chemistry, machine learning, pharmaceutical research, and safety science, James discusses how modern AI approaches are reshaping nonclinical safety assessment, supporting human-relevant toxicology, and helping researchers make more informed decisions across the drug development pipeline.

You have worked across the full drug development pipeline. How has that experience shaped your view of where AI can have the greatest impact in drug safety discovery?

Having worked across different parts of the discovery process, I think broad exposure changes the way you think. I have developed quantum chemistry, artificial intelligence (AI) and machine learning (ML) methods in research settings, but I have also had the chance to apply them in real discovery environments for materials and pharmaceuticals. That experience has shown me how different parts of the discovery pipeline interact, what their motivations are, and how they communicate. AI and ML have been used in early discovery for many years, but the biggest recent shift is in areas dominated by human processes and communication. Computers can now interact directly with human language and analytical images, opening new applications such as AI research assistants and AI-assisted dossier drafting. That means AI can start to work more collaboratively with scientists, although expert oversight and guardrails remain essential. In nonclinical safety, I think the biggest impact will be AI working alongside safety scientists to unearth key data, connect safety science with earlier discovery and later clinical work, and help process, predict, and summarize important outcomes. That could help safety science integrate more tightly with design, make, and test cycles in discovery, while also interfacing earlier with clinical science. Save a Copy of This Exclusive Interview by Downloading the Free PDF

Your early research focused on predicting molecular properties such as solubility using machine learning. What did that teach you about data-driven property prediction?

Data-driven models, such as Quantitative Structure Activity Relationships (QSAR), are incredibly attractive because they offer rapid methods for predicting molecular properties, including those that are difficult to access through fundamental chemical and physical theory. For new molecules, in related regions of chemical space to the training data, they can often be very accurate. The challenge is that they typically do not generalize across large diverse chemical and biological spaces. Therefore real care is needed when applying them to novel chemistry or biology. Some attempts to build more widely applicable QSAR models have been made in recent years, but most models in use today are still built for specific chemistries and properties owing to the scale and complexity of chemical space. When I started this work, especially for solubility predictions, the datasets were modest by today’s standards, often hundreds to a few thousand molecules. One of the first things that teaches you is to explore the data carefully, assess its quality and quantity, and only then move forward with modeling. The aim is to capture genuine relationships between numerical molecular descriptions and important target endpoints, but those relationships may be only locally generalizable. I also think the phrase “simple descriptors” is interesting. Some are simple, such as atom counts or bond counts, but others rely on detailed parameterization, group contributions, and graph-theoretical techniques. We still see these descriptors used today, sometimes alongside embeddings from graph neural networks or language models. The main lesson for me is that there is no single workflow. Each dataset, model, and end-use case needs careful thought.

With graph neural networks, foundation models, and diffusion models transforming the field, what can we predict reliably today that was difficult a decade ago and are there still any blind spots?

One major example is protein structure. Methods such as AlphaFold and Boltz-2 have transformed how we generate 3D structural data. Accurate protein structure prediction has been a decades-old challenge, and while there is still work to do on underrepresented protein classes, these models have opened an extraordinary pathway. We have also seen progress in multi-output models. Just over a decade ago (circa 2013), toxicology and property prediction models often focused on one endpoint at a time. Now, multi-output models, especially from methods such as graph neural networks (GNNs) and other deep learning methods, can predict multiple endpoints simultaneously. This allows for transfer learning between related properties. That creates opportunities across ADMET (absorption, distribution, metabolism, excretion, and toxicity) prediction and toxicology. Another exciting area is machine learning potentials. Molecular simulations traditionally rely on fixed mathematical functions to describe molecular interactions, but neural potentials can now replace or augment those functions. They can, in many cases, provide highly accurate simulations and structures at lower computational cost than quantum chemistry. However, its not just down to model architectures. The biggest blind spot behind all of these advances remains careful data curation, annotation, and collection. Open databases for example ChEMBL, PubChem and Chemspider together with data initiatives such as OpenADEMT and OpenBind to name but a few are making notable progress here. We still see some data for example on formulation chemistry often receiving less attention. Its is not just about data scale, but quality that is critical here to. Accessible high quality data sets are the fuel of AI methods. Architecture matters, but high-quality, accessible and abundant data is also absolutely critical.

In practical drug discovery and safety projects, where do graph-based models outperform traditional quantitative structure-activity relationship (QSAR) approaches, and where do conventional methods still offer advantages?

For molecular design, classical machine learning with traditional fingerprints is not going to generate accurate 3D structures in the way modern graph-based, diffusion, and foundation models can. Where we have large, high-quality datasets, graph neural networks, language models, and foundation models can also outperform classical methods for property prediction. We even see some attempts at generating very large and widely applicable QSAR models, although bespoke models for specific target chemistries and properties remain the main form of QSAR model in use today. However, classical machine learning models are still very useful, especially for small datasets, which are common in early discovery. Classical machine learning models often have fewer parameters and can generalize better in small chemical spaces than larger deep learning models, which may overfit in such spaces. Classical models are also often much faster to train and perform inference with using modest compute infrastructure, such as a laptop. They can also be easier to interpret because many descriptors have a chemical or physical rationale. I do not think classical machine learning belongs in the past. The method pool is broader now, and the researcher's or engineer’s understanding of the data and desired outputs should guide the modeling decision.

What does an AI-assisted drug safety workflow realistically look like today?

Toxicology is a very diverse field, spanning chemistry, biology, and clinical science. Data varies enormously in accessibility, quantity, quality, terminology, and format. That makes it a complex domain for AI, but also one where AI can be very useful when applied carefully. A realistic AI-assisted safety pipeline today is not about replacing the expert toxicologist. It is about augmentation. AI agents and Large Language Models (LLMs) can autonomously search for data, collate information, generate summaries, and produce plots. This allows safety scientists to bring their expertise to bear more efficiently, while AI helps uncover and organize the evidence. We should also remember task-specific models. DeepDILI, for example, combined classical and deep learning approaches to predict drug-induced liver injury (DILI), a major cause of post-marketing drug withdrawals. Models like this can have a real impact when focused on specific safety pain points.

ApconiX's in vitro seizure liability assay (iSLA) platform provides a human-relevant approach for assessing seizure liability. Where can AI add the most value alongside assays like this?

There are several areas where AI could add value, but one of the most exciting is helping us translate complex experimental results into actionable insights for drug discovery teams. We see computational and experimental approaches as complementary rather than competing technologies. In vitro assays such as iSLAseizure generate highly relevant biological data, while AI and machine learning can help uncover patterns within that data that might otherwise be difficult to identify. One example is work I recently presented at the BioNow BioAI Symposium, in which we explored the use of structure-activity relationships (SAR) and machine learning models alongside iSLAseizure assay data. Supported by colleagues, including Dr. Louisa Zolkiewski and Dr. Kim Rockley, ApconiX has developed an in-house seizure liability dataset that we analyzed in collaboration with MedChemica. Using machine learning and chemical informatics approaches, we identified chemical substructures that show statistically significant relationships with seizure-related outcomes. The aim is not simply to predict risk, but to understand which molecular features may contribute to that risk and provide medicinal chemists with practical guidance they can apply much earlier in the discovery process. Looking ahead, I think this is where AI can have a particularly powerful impact. Rather than acting as a standalone prediction tool, it can work alongside human-relevant assays such as iSLAseizure to connect molecular structure, experimental biology, and safety outcomes. That combination has the potential to improve decision-making earlier in the discovery process, reduce late-stage safety surprises, and ultimately support the development of safer medicines for patients.

How can AI, in vitro systems, and omics technologies support reducing animal testing while maintaining scientific confidence?

There has been a regulatory drive to reduce animal testing for some time, including through frameworks such as the European Union’s REACH directive. More recently, we have seen this strengthened with roadmaps to phase out animal testing for chemical safety. In silico and in vitro methods are likely to be central to that transition. These tools can provide efficient, human-relevant, and multiscale evidence. One could imagine an approach similar to the carcinogenicity weight-of-evidence approach, in which in vitro and in silico data are gathered to justify reduced or no animal testing at specific points in the approval process. Regulatory acceptance remains challenging because standards vary around the world. For AI, there are also questions about model approval, model updates, benchmarking, and version control. It will be very interesting to see how regulators test and define these pathways.

You’ve experienced ApconiX both from within and as a biopharma collaborator through Redx. How has your experience in both biopharma and safety science influenced the way you develop AI tools for drug discovery teams?

The tools we are building and applying at ApconiX are currently internal-facing, supporting our scientists and collaborators in delivering thoroughly researched, evidence-based, and authoritative toxicology information in a timely manner. Even though I did not directly interact with ApconiX while I was at Redx, my experience in biopharma strongly shapes how I think about usability. At ApconiX, we constantly consider how others will interpret and use the data and insights we generate. For example, our early SAR work around seizure liability has identified preliminary relationships between chemical substructural patterns and seizure risk. Although this work is in its infancy, these are the kinds of insights that discovery chemists could interpret directly and apply from day one to help mitigate risk. For target safety assessments, the ApconiX data science team have collated a large data sets including our Acuity expression atlas for expression data, enabling our data scientists and bioinformaticians to generate highly informative, data-driven insights and visualizations. By leveraging AI tools, such as LLM workflows and agents, we can efficiently support consistent and thorough analyses, making safety information more accessible and actionable.

There is considerable excitement around AI in drug development. Where do you see the largest gap between hype and reality, and where is AI perhaps underappreciated?

AI is moving so quickly that in the gap between answering this and publication, I could be proved wrong! With that caveat, there is a lot of hype around AI drug discovery campaigns. AI is certainly a substantial part of many discovery programs today and has been applied in drug discovery for many years. However, we need to be realistic about what AI is doing, where humans remain central, and how that balance may vary. AI agents are an area that has emerged relatively recently and continues to evolve quickly. They can complete tasks for example data extraction and reporting with guidance from experts, helping to generate larger, more standardized datasets over time. They can also help build closed-loop systems that connect computation and experiment directly, with strategic direction still coming from expert scientists. However, care and oversight is needed with these systems to validate their actions. I also think people often equate AI with generative AI, but there are many other AI systems which are underappreciated. For example, the GNN property prediction models and task-specific predictive tools we discussed earlier can be extremely valuable. Some of these methods receive less attention simply because they have already become familiar in drug discovery.

In safety assessment, the consequences of being wrong are significant. How do you think about uncertainty and model limitations for safety and discovery?

Uncertainty and model limitations are critical in safety assessment. There is clearly greater potential for animal or human harm from an incorrect safety assessment than from making the wrong molecule in early discovery, although uncertainty matters in both contexts. It is important to distinguish different types of uncertainty. Aleatoric uncertainty reflects irreducible randomness in a task, while epistemic uncertainty is model-related, such as hallucinations or lack of knowledge. For LLM systems, we use multiple methods to measure consistency and we have built fact-checking approaches in which a separate LLM independently verifies the primary engine's outputs. For QSAR models, including our SAR work around seizure liability, we can use statistical metrics such as confidence intervals to quantify and communicate uncertainty. Understanding that uncertainty exists and communicating it in both safety and discovery is critical. Uncertainty metrics tell us where potential blind spots may exist in our modelling allowing us to take action to improve over time. This applies to both safety and discovery. The aim is not to pretend AI removes uncertainty, but to make uncertainty visible so scientists can make better-informed decisions. Download the Full PDF to Revisit James McDonagh's Insights on AI in Drug Safety Discovery

Where can readers find more information?

About the researcher

James McDonagh is a Data Scientist at ApconiX, where he develops and applies AI, machine learning, and computational science approaches to support modern safety assessment. His work includes introducing large language model workflows, AI agents, QSAR models, and new computational infrastructure to help scientists generate and interpret toxicology data more efficiently while expanding the company's data science capabilities.

James studied Chemistry at Bangor University before completing a PhD in Computational Chemistry and Chemical Informatics at the University of St Andrews, where he combined quantum chemistry and machine learning to predict the solubility of drug-like molecules. He subsequently undertook postdoctoral research at the University of Manchester, developing novel quantum chemistry methods and machine learning models to accelerate complex molecular simulations.

Before joining ApconiX, James held roles at IBM Research Europe, Serna Bio, and Redx Pharmaceuticals, applying AI and computational methods to molecular discovery, property prediction, and drug development. His work has spanned molecular simulation, generative AI, cheminformatics, and AI-driven drug discovery, providing the multidisciplinary experience that now underpins his efforts to integrate AI into safety science and predictive toxicology.

• Anstine, D. M., Zubatyuk, R., & Isayev, O. (2025). AIMNet2: a neural network potential to meet your neutral, charged, organic, and elemental-organic needs. Chemical Science, 16(23), 10228-10244. DOI:10.1039/D5SC01873A, https://pubs.rsc.org/en/content/articlelanding/2025/sc/d5sc01873a
• Deng, J., Yang, Z., Wang, H., Ojima, I., Samaras, D., & Wang, F. (2023). A systematic study of key elements underlying molecular property prediction. Nature Communications, 14(1), 6395. DOI:10.1038/s41467-023-41927-1, https://www.nature.com/articles/s41467-023-41927-1
• Du, F., Yu, H., Zou, B., Babcock, J., Long, S., & Li, M. (2011). hERGCentral: a large database to store, retrieve, and analyze compound-human Ether-a-go-go related gene channel interactions to facilitate cardiotoxicity assessment in drug development. Assay and Drug Development Technologies, 9(6), 580-588. DOI:10.1089/adt.2011.0425, https://www.liebertpub.com/doi/10.1089/adt.2011.0425
• Heid, E., Greenman, K. P., Chung, Y., Li, S. C., Graff, D. E., Vermeire, F. H., ... & McGill, C. J. (2023). Chemprop: a machine learning package for chemical property prediction. Journal of Chemical Information and Modeling, 64(1), 9-17. DOI:10.1021/acs.jcim.3c01250, https://pubs.acs.org/doi/10.1021/acs.jcim.3c01250
• Huynh, D. L., Seal, S., Reid, D., Carpenter, A. E., Bender, A., Spjuth, O., & Kasbekar, S. (2026). AI agents in drug discovery: applications and case studies. Drug Discovery Today, 104650. DOI: 10.1016/j.drudis.2026.104650. https://www.sciencedirect.com/science/article/pii/S1359644626000553
• Jaeger, S., Fulle, S., & Turk, S. (2018). Mol2vec: unsupervised machine learning approach with chemical intuition. Journal of Chemical Information and Modeling, 58(1), 27-35. DOI:10.1021/acs.jcim.7b00616, https://pubs.acs.org/doi/10.1021/acs.jcim.7b00616
• Jiang, D., Wu, Z., Hsieh, C. Y., Chen, G., Liao, B., Wang, Z., ... & Hou, T. (2021). Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models. Journal of Cheminformatics, 13(1), 12. DOI:10.1186/s13321-020-00479-8, https://jcheminf.biomedcentral.com/articles/10.1186/s13321-020-00479-8
• Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., ... & Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583-589. DOI:10.1038/s41586-021-03819-2, https://www.nature.com/articles/s41586-021-03819-2
• Li, T., Tong, W., Roberts, R., Liu, Z., & Thakkar, S. (2020). DeepDILI: deep learning-powered drug-induced liver injury prediction using model-level representation. Chemical Research in Toxicology, 34(2), 550-565. DOI:10.1021/acs.chemrestox.0c00326, https://pubs.acs.org/doi/10.1021/acs.chemrestox.0c00326
• McDonagh, J. L., Nath, N., De Ferrari, L., Van Mourik, T., & Mitchell, J. B. (2014). Uniting cheminformatics and chemical theory to predict the intrinsic aqueous solubility of crystalline druglike molecules. Journal of Chemical Information and Modeling, 54(3), 844-856. DOI:10.1021/ci4005805, https://pubs.acs.org/doi/10.1021/ci4005805
• McDonagh, J. L., Shkurti, A., Bray, D. J., Anderson, R. L., & Pyzer-Knapp, E. O. (2019). Utilizing machine learning for efficient parameterization of coarse grained molecular force fields. Journal of Chemical Information and Modeling, 59(10), 4278-4288. DOI:10.1021/acs.jcim.9b00467, https://pubs.acs.org/doi/10.1021/acs.jcim.9b00467
• McDonagh, J. L., Silva, A. F., Vincent, M. A., & Popelier, P. L. (2017). Machine learning of dynamic electron correlation energies from topological atoms. Journal of Chemical Theory and Computation, 14(1), 216-224. DOI:10.1021/acs.jctc.7b00975, https://pubs.acs.org/doi/10.1021/acs.jctc.7b00975
• McDonagh, J. L., Wunsch, B. H., Zavitsanou, S., Harrison, A., Elmegreen, B., Gifford, S., ... & Cipcigan, F. (2023). Machine guided discovery of novel carbon capture solvents. arXiv preprint arXiv:2303.14223. DOI: https://doi.org/10.48550/arXiv.2303.14223. https://arxiv.org/abs/2303.14223
• Pallikkavaliyaveetil, N., & Chandrasekaran, S. (2025). Small data, big challenges: Machine- and deep-learning strategies for data-limited drug discovery. Advanced Drug Delivery Reviews, 115762. DOI: https://doi.org/10.1016/j.addr.2025.11576. https://www.sciencedirect.com/science/article/pii/S0169409X25002479
• Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Somnath, V. R., ... & Barzilay, R. (2025). Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv. DOI: https://doi.org/10.1101/2025.06.14.659707. https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1
• Pollack, J. (1988). Implications of recursive distributed representations. Advances in Neural Information Processing Systems, 1. (No DOI available.)
• Sato, T., Yuki, H., Ogura, K., & Honma, T. (2018). Construction of an integrated database for hERG blocking small molecules. PLoS ONE, 13(7), e0199348. DOI:10.1371/journal.pone.0199348, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0199348
• Sumer, Z., McDonagh, J. L., Fare, C., Tadikonda, R., Zolyomi, V., Bray, D., & Pyzer-Knapp, E. (2025). Providing Machine Learning Potentials with High Quality Uncertainty Estimates. arXiv preprint arXiv:2501.05250. DOI: https://doi.org/10.48550/arXiv.2501.05250. https://arxiv.org/abs/2501.05250
• U.S. Food and Drug Administration (n.d.). SafetAI Initiative. National Center for Toxicological Research – Research Focus Areas. Available at: https://www.fda.gov/about-fda/nctr-research-focus-areas/safetai-initiative
• Wiener, H. (1947). Structural determination of paraffin boiling points. Journal of the American Chemical Society, 69(1), 17-20. DOI:10.1021/ja01193a005, https://pubs.acs.org/doi/10.1021/ja01193a005