More human relevant models: how ACMs, multiomics and AI are helping change drug discovery

Historically, safety assessment of drug candidates has relied on traditional 2D cell cultures and animal models. Advanced cell models (ACMs) represent an exciting shift within the broader category of new approach methodologies (NAMs). By utilizing primary human cells or induced pluripotent stem cells grown in 3D structures like spheroids or organ-on-a-chip systems, ACMs closely mimic human tissue architecture, allow for complex multi-organ interactions, and possess the longevity needed to detect a range of toxicities. They are rapidly becoming the cornerstone of our safety assessment because they offer a faster, more sustainable and human-related model for building confidence in potential drug candidates than traditional approaches.

While ACMs offer enhanced biological relevance, their widespread adoption comes with challenges. For instance, with multiple external and in-house models available, which one is most suitable for a specific biological question? Furthermore, ACMs can introduce complexity, such as donor-to-donor variation and batch-to-batch differences in matrices. This is where multiomics makes the difference, supporting our model qualification strategy, which relies on three foundational pillars. First, we perform deep molecular characterization to confirm the model fits a particular context of use. Thanks to the massive increase in throughput and reduced costs of our omics pipelines, we can directly measure if a specific target or pathway is expressed, allowing us to short-list ACM models based on their molecular fingerprint prior to functional assessment. It was incredibly satisfying to see the US FDA (MD, USA) acknowledge the importance of omics for ACM selection and validation in their recent draft guidance on NAMs, explicitly mentioning transcriptomics for determining human biological relevance.

Second, we conduct rigorous functional benchmarking of human relevance by testing known clinical toxins. For example, when we expose liver spheroids to an established hepatotoxin, we must clearly detect the expected human mechanistic signatures, such as cholestasis and oxidative stress. Third, we execute strict technical characterization to quantify variation and guarantee highly reproducible behavior across different donors and biological batches. While we don’t currently use AI directly for this initial ACM qualification, we heavily leverage it to build highly predictive safety models once an ACM is selected and verified.

Animal studies remain necessary in certain contexts for the discovery, development and regulatory approval of new medicines, and continue to be an important part of R&D when no suitable, scientifically valid and regulatory accepted alternative exists. However, we remain committed to advancing alternatives whilst delivering new therapies to patients, balancing our responsibility to those who depend on new medicines with our commitment to the 3Rs and reducing animal use. Certain ACM assays are already replacing animal studies, and the integration of omics and artificial intelligence (AI) is rapidly accelerating this transition. Beyond using omics for informed selection and qualification, we combine these molecular profiles with machine learning (ML) approaches on models like liver spheroids to directly predict toxicity. A great example of this is a recent project where we were unable to confirm tolerability with  a conventional in vitro assays – historically this would have meant animal studies to shortlist a tolerated compound. Instead, we demonstrated that the combination of ACMs, transcriptomics and ML could successfully differentiate between tolerated and non-tolerated compounds.

To continue building confidence in this approach and find further opportunities to reduce animal usage, we consistently benchmark our multiomics readouts against in vivo studies for specific safety liabilities and drug modalities. Such approaches can also be applied to bridge the ‘translational gaps’ between ACMs and the clinic. This can be achieved by comparing drug responses in ACMs against representative clinical samples using advanced omics technologies, which boosts our confidence in the models and helps guide early drug discovery to be far more efficient. Ultimately, our human-relevant models are already driving real portfolio choices, helping us design and select new molecules and optimize dosing schedules – so trials can start sooner.

Patient safety is at the center of our efforts and the criteria we use are intentionally designed to align with global regulatory expectations, most prominently the recently released FDA draft guidance, ‘General Considerations for the Use of New Approach Methodologies in Drug Development‘. Regulators are increasingly open and excited about these technologies, but the biggest hurdles remain rooted in limited prior experience and a lack of universal standardization. Currently, the draft guidance lacks specificity regarding the handling, standardization and validation of high-dimensional molecular data. We are missing provisions or references to existing frameworks for validating bioinformatics pipelines, data normalization techniques, and the underlying algorithms used to interpret multiomics data generated from NAMs. Establishing regulatory confidence requires absolute transparency, not just in the biological model itself, but in the computational processing of its outputs. This level of rigor is equally important for AI driven predictive models, where non negotiables include defining the model scope, ensuring data quality and integrity, evaluating performance metrics, and enforcing tightly controlled stability and model governance.

It’s important to commit to data standardization, versioning and validation from day 1. We cannot solve this in a silo. By proactively partnering with academic centers of excellence, innovative technology vendors, peer pharmaceutical companies and regulatory bodies, and by openly presenting our validation frameworks at major scientific forums, we can collectively build the unified data standards required to required for regulatory purposes, ensure future benefit to patients and trust and move the industry forward.

The transition from broad omics discovery to targeted clinical endpoints is entirely driven by the purpose of the experiment. For preclinical safety evaluations, broad omics profiling is fantastic for detecting adverse outcome pathways (AOPs) and, when combined with AI, human safety prediction. However, for clinical dose predictions, a rigorously standardized and quantitative panel is essential. In that scenario, we use our broad omics data to confidently shortlist a targeted panel of biomarkers based on statistical significance, biological context and practical clinical considerations. From there, we transition into developing that specific panel by following industry best practices and rigorous fit-for-purpose (FFP) assay validation, ensuring it is robust enough for clinical trial deployment.

The reality is that no one can be a world-class expert in advanced cell models, toxicology, multiomics, data science and machine learning all at once. Instead, the future belongs to highly collaborative, matrixed teams with individuals representing each of these disciplines. While everyone maintains their core expertise, it is now absolutely critical to understand the foundational principles of the other disciplines to ensure seamless communication and prevent insights from getting lost in translation. You must be relentlessly curious about your colleagues’ work, and I am very grateful to see exactly this within our organization. As an omics expert today, you need a solid grasp of basic biology, toxicology and machine learning principles. My informal mentor, Northeastern University (MA, USA) emeritus professor Barry Karger, formulated this perfectly with his “3 Ms”: molecule, method and meaning. Beyond understanding the characteristics of the molecule you intend to measure and the method you use to measure it, the analytical bio-scientists who can deeply interpret the meaning of their data within a broader biological and computational context are the ones who will truly differentiate themselves, work effectively across teams, and deliver massive impact to the portfolio.