Teaching biology to AI makes protein models dramatically better at predicting antibodies
AI models that study proteins have a blind spot: they learn from data without understanding how evolution actually works.
Protein language models work much like the AI systems that generate text: they identify patterns across vast biological datasets and use those patterns to predict which protein sequences will have which properties. The technology has already shaken up pharmaceutical research. But there is a fundamental limitation: these models learn from what exists, not from how it came to exist. For antibodies that distinction matters enormously.
Antibodies are not born optimized. They undergo a process called affinity maturation: as the immune system fights an infection, B cells mutate their antibody genes and the body selects the variants that bind most effectively to the target. It is evolution at microscopic scale, happening in real time. A model that truly understands those evolutionary trajectories — which mutations are likely, which binding properties improve — could in principle design antibodies that outperform anything the immune system would naturally produce.
Biology as a training rule, not just training data
The study, published in eLife, introduces a training paradigm that explicitly incorporates that biological logic. Rather than training purely on statistical patterns in known antibody sequences, the model is trained on the evolutionary paths antibodies take during maturation. It learns not just which sequences exist, but why they emerged and what the next step in that evolution might look like.
In benchmark tests, the biologically informed model outperformed existing approaches at predicting affinity maturation trajectories. That means it was better at forecasting which antibody variants would bind more strongly to their targets.
From research tool to drug pipeline
The practical stakes are high. Antibody therapies are one of the fastest-growing classes of medicines, used across cancer, autoimmune diseases, and infections. But designing an antibody that binds well, remains stable, and does not trigger an immune reaction in the patient is expensive and slow. If AI models can accelerate that process by identifying better candidates earlier, the impact on drug development timelines could be substantial.
One unresolved question hangs over the approach: how much biology is enough? Models can be enriched with ever more biological constraints, but at some point they become so complex that they are difficult to interpret or generalize. Finding the right balance between biological fidelity and practical usability is a methodological challenge the field has not yet solved.