Artificial intelligence is steadily reshaping biologics discovery, moving the field from empirical experimentation toward predictive design. Nowhere is this shift more visible than in the development of antibodies and protein therapeutics.
Professor Charlotte Deane, Professor of Structural Bioinformatics at the University of Oxford, has been at the forefront of this transformation. Her work has helped make antibody structure prediction scalable, enabling researchers to model thousands of antibodies rapidly and early in the drug discovery process.
According to Professor Deane, the most significant change is not that AI has replaced experimentation—but that it has fundamentally altered how experiments are chosen.
From Structure at the End to Structure at the Beginning
Historically, structural biology played a relatively late-stage role in biologics development. Today, computational tools allow researchers to generate structural models of antibodies at the very start of a project.
Tools such as ImmuneBuilder, developed by Professor Deane’s group, allow large-scale structural prediction that would previously have required years of experimental effort. This scalability is transforming discovery pipelines.
Rather than screening broadly and refining slowly, researchers can now use computational models to propose promising variants, narrowing experimental work to the most informative candidates. The result is fewer iterations and more focused laboratory campaigns.
The pipeline itself is evolving—from experiment-first toward computation-guided experimentation.
A Field at a Tipping Point
Despite rapid progress, Professor Deane describes the field as being at a “tipping point.”
Most biologics programs are still driven primarily by experiments. However, AI tools are increasingly influencing which experiments are performed and in what order.
Where computational methods are already proving powerful is in optimization. Once an initial binder is identified, AI can suggest alternative sequences predicted to improve affinity or developability. In these cases, the computer is beginning to guide decision-making.
Fully computation-first design—where AI proposes initial libraries or directs early-stage strategies—remains less common. But the trajectory is clear: prediction is steadily moving earlier in the process.
The Limits of Data
For all the excitement surrounding AI in drug discovery, Professor Deane emphasizes a crucial reality: AI is only as strong as the data it learns from.
Antibodies present particular challenges. They are flexible, diverse, and capable of subtle conformational changes that influence binding. Yet experimental data describing antibody flexibility is limited. Without sufficient data, predictive accuracy suffers.
Even in areas that appear data-rich, the scale of biological diversity dwarfs available datasets. Public databases contain roughly 1,500–2,000 independent antibody–antigen complex structures. In contrast, the human immune system can generate billions of distinct antibodies.
This mismatch between biological complexity and available training data remains one of the field’s most significant constraints.
Studies increasingly show that improvements in prediction often depend more on data quality and representativeness than on new model architectures. Better datasets—not just better algorithms—are essential to progress.
Where AI Delivers—and Where It Doesn’t
Today’s AI tools can reliably predict individual antibody structures and are making progress in modeling developability properties such as “humanness.” There are also promising advances in predicting antibody–antigen complexes—particularly when binding is already known.
However, important challenges remain.
Accurate binding affinity prediction is still difficult. Modeling flexibility and poly-specificity remains limited. And predicting entirely new binders—rather than modeling known interactions—is still a major hurdle.
Separating meaningful progress from hype requires transparency in model training and validation. Open implementations allow independent evaluation, but not all reported advances offer that level of clarity.
The Enduring Role of Human Expertise
As AI becomes more integrated into drug discovery, human expertise remains indispensable.
Predictive models generate plausible solutions—but they do not inherently provide definitive answers. For example, a structure prediction tool will produce a model even if an antibody and antigen do not bind. Interpretation and biological judgment remain essential.
Scientists still define therapeutic targets, frame hypotheses, and determine whether computational outputs are biologically meaningful. AI augments expertise—it does not replace it.
The Next Five to Ten Years
Looking ahead, Professor Deane envisions a future where computation increasingly drives iteration cycles.
Instead of alternating between broad experimental campaigns and occasional modeling, discovery workflows may begin with computational generation of candidate solutions. A small subset would then be experimentally validated, and results fed back into models for refinement.
The goal is not perfect prediction, but faster convergence—reducing the number of experimental cycles needed to arrive at an effective therapeutic candidate.
While fully automated therapeutic design remains a distant vision, incremental improvements in data integration, scalable modeling, and multi-parameter optimization are steadily pushing the field forward.
Understanding the Hard Problems
One of the most pressing unresolved challenges is antibody flexibility—understanding how dynamic structural changes influence binding and function. Addressing this will require better data, improved modeling strategies, and deeper integration of structural biology with AI.
These themes will form part of Professor Deane’s presentation at NextGen Biomed 2026 in London, where she will explore scalable prediction tools and the next frontier in biologics design.
Artificial intelligence is not replacing the laboratory. Instead, it is redefining how questions are asked—and how quickly answers can be found.
As computation becomes more deeply embedded in biologics discovery, the balance between prediction and experimentation will continue to shift. The future of antibody and protein therapeutics may well be shaped not only by what we can observe, but by what we can predict.
