AI is remaking biologics discovery

Biologics discovery used to be a wet-lab slog. Researchers would screen huge libraries of antibodies or proteins, hunt for a molecule that bound the right target, and then spend months or years trying to fix everything that was wrong with it. Artificial intelligence is changing that rhythm. A review highlighted this week describes a field moving from trial and error toward design, where models read biological sequences, predict structures, and generate candidate molecules before a pipette ever moves. That shift starts with a simple idea. Proteins and antibodies are sequences, and sequences have patterns. Protein language models are trained on enormous biological datasets and learn the statistical rules of that patterning, much as large language models learn grammar from text. In biologics, that means a model can score whether an antibody sequence looks natural, suggest mutations, and help researchers search a much larger design space than older screening methods could reach. Structure prediction made the next leap possible. AlphaFold turned protein folding from a stubborn experimental bottleneck into a computational routine. AlphaFold 3 pushed further by predicting how proteins interact with other molecules, including DNA, RNA, and ligands. That matters because biologics do not work as isolated shapes. They work by binding, avoiding, blocking, or carrying something else. Once those interactions became easier to model, design stopped being guesswork and started looking more like engineering. Now generative models are pushing beyond prediction into invention. Diffusion models, autoregressive systems, and related tools can propose entirely new proteins, peptides, and antibody variants with chosen properties. Researchers have already shown that AI systems can design antibodies from scratch, a result Nature called a landmark moment in 2024, even if those molecules are still far from routine clinical use. The important change is not that AI has solved biology. It is that candidate generation is no longer the slowest step. As that front end speeds up, the bottleneck slides downstream. In biologics, finding a binder is only the beginning. The candidate also has to stay stable, avoid aggregation, be manufacturable at scale, circulate in the body in a useful way, and not trigger the wrong immune response. That cluster of traits is often called developability, and it is where many attractive early molecules fall apart. AI is now being used there too, especially in antibody work, to predict stability, optimize paired heavy and light chains, and flag liabilities earlier. This is why faster design could actually raise the stakes for preclinical judgment. If teams can generate promising candidates in days instead of months, they can also move weak candidates forward faster. A model that is excellent at structure may still miss immunogenicity, pharmacokinetics, or the messy effects of cellular context. The review behind this week’s story makes that point plainly: in silico success still does not reliably predict in vivo success. Speed does not remove risk. It compresses the time available to notice it. That compression is pushing drug discovery toward closed loops, where AI proposes molecules, automated experiments test them, and the new data flows back into the model. The field is also reorganizing around access to proprietary datasets, because the best models need more than public protein structures. They need assay results, failed candidates, manufacturing data, and safety signals that drug companies have historically kept to themselves. In antibody discovery, that private data may matter as much as the model architecture. The result is a different kind of race. It is not just about who can generate the most molecules. It is about who can kill the wrong ones earliest, with evidence strong enough to survive the jump from code to animals to humans. In biologics, the molecules are getting easier to imagine. The hard part is still deciding which one deserves to enter a body.