Can artificial intelligence truly transform drug discovery without equally transforming the biology and data that underpin it?

Jeroen Verheyen at Semarion

Artificial intelligence (AI) is often presented as the force that will transform drug discovery. In many ways, this is true. AI is already reshaping how researchers design molecules, interrogate biological networks, prioritise targets and interpret complex data sets. But there is a risk that the current excitement around AI obscures a more fundamental point: AI does not remove the need for experimental biology. It raises the standard for it.

Drug discovery remains an experimental science. Computational models can propose hypotheses, but those hypotheses still need to be tested in biological systems that are reproducible, scalable and relevant to human disease. The limiting factor is therefore not only algorithmic capability. It is also the quality, comparability and structure of the biological data that feeds these systems.

This is especially important for cell-based workflows. Cells remain one of the most important interfaces between molecular design and biological understanding. They allow researchers to observe target engagement, pathway modulation, toxicity, phenotypic change and disease-relevant response in living systems. Yet many cell-based workflows are still difficult to standardise, hard to automate and challenging to compare across experiments, sites and model systems.

The future of drug screening will not be driven by AI alone. It will depend on a golden triangle of high-quality biological data, flexible lab automation and AI-guided decision-making. Biology provides the disease-relevant signal. Automation converts that signal into reproducible, scalable and well-annotated data. AI interprets the data, proposes better experiments and accelerates decisions. Remove any corner and the system becomes weaker.

For decades, drug screening has relied on a powerful operating model: simplify the assay; miniaturise the format; increase throughput and look for robust signals. This model has delivered enormous value. High-throughput screening, biochemical assays, engineered cell lines and plate-based workflows remain essential tools across discovery.

However, the demands on screening are changing. The industry is no longer asking only whether a compound modulates a single target or pathway. Discovery teams increasingly need to understand how compounds behave across cell types, genetic backgrounds, disease states and phenotypic contexts. They need to detect subtle mechanism-of-action signatures, identify early safety liabilities, understand selectivity and make better translational decisions before expensive downstream studies.

This shift creates tension. Simpler assays are often easier to scale, but they can miss important biology. More complex models may be more physiologically relevant, but they are often harder to run reproducibly, automate and analyse. As a result, discovery teams face a persistent compromise between throughput, robustness and biological richness.

This compromise matters because R&D productivity remains one of the central challenges in pharmaceutical innovation. High attrition, long cycle times and the cost of late-stage failure continue to place pressure on discovery organisations to make better decisions earlier.1,2 The next generation of screening must therefore do more than generate larger data sets. It must generate better data sets: more relevant; more reproducible; more comparable and more useful for decision-making.

In an AI-enabled discovery environment, a cell-based assay is no longer just an experiment. It is part of the data infrastructure that determines what a model can learn. This requires a shift in mindset. Cells are often treated as living materials that are prepared for each experiment, with day-to-day variation accepted as part of biological work. But for data-driven discovery, this variability becomes a limiting factor. Differences in passage number, culture history, differentiation state, plating density, attachment time, media conditions and handling can all influence assay behaviour. If these factors are not controlled or captured, they become hidden sources of noise in the data set.

Riss and colleagues have argued for the importance of ‘treating cells as reagents’ when designing reproducible cell-based assays.3 This does not mean ignoring the complexity of living systems. It means recognising that reproducible biology requires consistent cell sources, validated identity, controlled culture conditions, standard operating procedures and careful attention to phenotypic drift. In many workflows, it also means thinking more seriously about banking, qualification and standardised preparation of cellular materials.

This point becomes more important as discovery becomes more computational. AI systems do not simply need large volumes of data. They need data with context. Experimental conditions, cell state, assay timing, imaging settings, perturbation history and quality-control metrics all become part of the information layer. Without this context, data sets may be large but difficult to interpret, compare or reuse. The FAIR data principles, which emphasise that data should be findable, accessible, interoperable and reusable, are highly relevant here.4 For cell-based screening, FAIRness cannot begin at the data repository. It must begin at the bench. The physical workflow must be designed so that the resulting data is well annotated, consistent and machine-actionable.

The industry is moving towards more human-relevant biology. Patient-derived cells, induced pluripotent stem cell-derived models, organoids, spheroids, co-cultures, organ-on-chip systems and other new approach methodologies are gaining attention because they offer the potential to improve translational relevance. Regulatory and scientific momentum behind non-animal and human-relevant methods is also increasing, with agencies such as the US Food and Drug Administration highlighting the role of in vitro human-based systems, microphysiological systems and computational approaches in modernising drug development.5

This is a positive shift. Traditional immortalised cell lines and simplified assays have limitations, particularly when disease biology depends on tissue architecture, cellular heterogeneity, differentiation state or multicellular interactions. Advanced models can capture biology that simpler systems cannot.

Yet increased physiological relevance often comes at the cost of increased workflow complexity. Organoids can vary in size, morphology and maturation state. Induced pluripotent stem cells-derived cells can be affected by differentiation efficiency and batch-to-batch variability. Primary and patient-derived cells may be scarce, expensive or difficult to expand. Organ-on-chip systems can reproduce dynamic tissue-like environments, but may require specialised handling, perfusion control and analysis methods.

The key challenge is therefore not whether advanced models are valuable. It is how to make them reproducible, scalable and compatible with modern screening. A disease-relevant model that cannot be standardised or compared across experiments may generate insight in a specialised laboratory, but it will struggle to become a routine discovery engine. Conversely, an automated workflow that scales a biologically weak model may produce large volumes of data with limited translational value.

The golden triangle matters because it prevents the industry from over-optimising one dimension. Better biology must be paired with better workflow engineering. Better automation must preserve biological relevance. Better AI must be grounded in experimental systems that produce reliable feedback.

High-content imaging and phenotypic profiling provide a useful example of where the field is heading. Unlike single-endpoint assays, image-based approaches can capture rich information about cell morphology, organelle structure, cell cycle state, stress response, spatial organisation and population heterogeneity. Instead of asking only whether a compound changes one measurement, high-content phenomics asks how a compound changes a biological system. Cell painting is a prominent example. By using a standardised set of fluorescent markers to capture multiple cellular compartments, cell painting generates morphological profiles that can be used for mechanism-of-action analysis, compound clustering, toxicity prediction and biological discovery.6,7 More broadly, high-content imaging has become increasingly important because it produces data that is naturally suited to machine learning. Images contain far more information than most traditional assay readouts and AI methods are increasingly capable of extracting meaningful patterns from these complex data sets.8

However, high-content phenomics also illustrates the central challenge of AI-enabled screening. The image analysis may be computational, but the quality of the output depends heavily on the quality of the experiment. Cell preparation, assay timing, staining quality, imaging conditions, plate effects, batch effects and metadata all influence what the model learns. If the biological and experimental layer is inconsistent, AI may detect artefacts rather than mechanisms.

This is why the future of high-content screening is not simply ‘more images’ or ‘more AI’. It is better integration of biological models, experimental design, automation, imaging, quality control, metadata and computational analysis. In this sense, high-content phenomics is becoming a bridge between wet-lab biology and machine intelligence. It converts complex cellular response into structured, analysable data, but only when the workflow is engineered with reproducibility and comparability in mind.

The most important shift in drug screening may not be from manual workflows to automated workflows. It may be from isolated experiments to learning systems. In the traditional model, an assay is designed, a screen is run, hits are analysed and the programme moves forward. In the emerging model, experiments become part of a feedback loop. AI proposes hypotheses or prioritises compounds. Automated systems execute experiments. Cell-based models generate rich biological data. Computational methods interpret the results. The next experiment is then designed based on what has been learned.

This is the logic behind design-make-test-analyse cycles, active learning and self-driving laboratory concepts. Autonomous laboratories remain an emerging field, and fully closed-loop biological discovery is not yet routine across pharmaceutical R&D. But the direction of travel is clear. The value of automation is not only that it reduces manual labour. It allows experiments to become more repeatable, schedulable and connected to data systems. The value of AI is not only that it predicts outcomes. It can help decide what experiment should be run next. The value of cell-based biology is not only that it provides a readout. It grounds computational learning in biological reality. Self-driving laboratory reviews have highlighted the importance of combining automated experimental workflows with machine learning, optimisation algorithms and data infrastructure.9 In drug discovery, this vision will require particular care because biological systems are variable, context-dependent and difficult to compress into simple objective functions. The goal should not be automation for its own sake, but higher-quality learning loops.

For cell-based workflows, this implies several practical priorities. First, biological inputs need to be more consistent and better characterised. Second, workflows need to be compatible with automation without losing biological relevance. Third, assay outputs need to be rich enough to support meaningful decision-making. Fourth, metadata and provenance need to be captured as part of the experiment, not reconstructed afterwards. Finally, experimental systems need to support comparison across models, conditions and timepoints. These are not merely operational details. They are strategic capabilities. In a world where AI can rapidly generate hypotheses, the competitive advantage shifts to organisations that can test those hypotheses quickly, reproducibly and meaningfully.

The next phase of drug screening will be won by organisations that treat cell-based workflows as engineered, data-producing systems. This does not mean reducing biology to an industrial process or pretending that living systems are simple. It means acknowledging that biological relevance and workflow robustness must be developed together.

AI will not replace the lab. It will make the lab more important. The companies and research organisations that succeed will not be purely computational, nor purely experimental. They will integrate model quality, automation, data infrastructure and computational learning into a single operating model. The industry has already invested heavily in AI. It has also invested heavily in automation, imaging and advanced cell models. The next challenge is integration. Better models, better automation and better AI are not separate trends. They only become transformative when they reinforce one another.

Cell-based workflows sit at the centre of this transition. They are where disease biology is translated into experimental data, where automation must meet biological complexity and where AI can learn from the real behaviour of living systems. If the industry wants AI-enabled drug discovery to fulfil its promise, it must engineer the biology layer with the same ambition that it is now applying to algorithms.

The future of drug discovery will not be defined by AI alone. It will be defined by the golden triangle of biology, automation and AI, and by how well we connect them into faster, more reliable and more human-relevant learning systems.

1. Visit: pubmed.ncbi.nlm.nih.gov/20168317/

2. Visit: pubmed.ncbi.nlm.nih.gov/26928437/

3. Visit: pubmed.ncbi.nlm.nih.gov/34530643/

4. Visit: nature.com/articles/sdata201618

5. Visit: fda.gov/science-research/science-and-research-special-topics/new-approach-methodologies-nams

6. Visit: pubmed.ncbi.nlm.nih.gov/27560178/

7. Visit: pubmed.ncbi.nlm.nih.gov/39639168/

8. Visit: pubmed.ncbi.nlm.nih.gov/38797109/

9. Visit: pubs.acs.org/doi/10.1021/acs.chemrev.4c00055

Jeroen Verheyen is co-founder and chief executive officer of Semarion. He has a multidisciplinary background spanning biomedical sciences, nanoscience, bionanotechnology, cell-based assay development and life sciences commercialisation. Before co-founding Semarion, Jeroen worked as a research associate at the University of Cambridge, UK, and held commercial and leadership roles across life sciences innovation, consulting and early-stage technology ventures. His work focuses on translating interdisciplinary technologies into practical drug discovery tools, including assay-ready cells, cell multiplexing, high-content imaging and automation-compatible adherent cell workflows.