Antibody design with AI: Foundation models, generative approaches, and the biologics pipeline

Few areas of biologics research have drawn more attention from computational biology than AI antibody design, and for good reason: antibodies are among the most important and fastest-growing classes of therapeutics, and their design poses problems that artificial intelligence (AI) is, in principle, well suited to address. But antibody design is not simply small-molecule design applied to bigger molecules. The sequence space is astronomically large, the developability constraints that determine whether an antibody can actually become a drug are severe, and the biology of how an antibody recognizes its target is genuinely complex. This article covers which AI approaches are actually working for biologics, and which remain closer to promise than practice.

It sits within DDN's coverage of generative AI in molecular design, complements the small-molecule discussions of diffusion models and AI-driven lead optimization, and connects to the guide to AI across drug discovery. Because antibody design leans heavily on protein structure, it also builds on the discussion of AlphaFold and structure prediction.

What makes antibody design computationally different

Antibodies are large Y-shaped proteins, and their therapeutic function depends on a small set of loops at the tips of the Y. As a comprehensive overview of antibody generative models describes, an antibody has two antigen-binding Fab regions and one Fc region, and within the Fab, the complementarity-determining regions (CDRs) are responsible for antigen binding, with the third heavy-chain loop, CDRH3, especially crucial to the antibody-antigen interaction. Designing an antibody means, above all, designing those loops, and that is a fundamentally different problem from designing a small molecule.

Continue reading below...

EbooksA new era of therapeutic antibodies

Explore the scientific advances driving more specific, potent, and adaptable antibody-based therapies.

Read More

Why antibody design resists the small-molecule playbook:

• Astronomical sequence space. An antibody is a protein of hundreds of amino acids, and even restricting design to the CDR loops leaves a combinatorial space of possible sequences far larger than the chemical space of small molecules, too large to screen exhaustively by any means.
• Severe developability constraints. A sequence that binds is not enough. To become a drug, an antibody must also be expressible, stable, soluble, non-aggregating, and non-immunogenic, a set of manufacturing and behavioral requirements collectively called developability that many binding sequences fail.
• Complex, structure-dependent binding. Antigen recognition depends on the three-dimensional shape of the CDR loops and how they contact the target, so a purely sequence-level view misses much of what determines binding, making structure central.
• Sparse, biased data. The experimental data available to train models is limited and skewed toward antibodies that already worked, which shapes and constrains what models can reliably learn.

These differences explain why antibody AI has developed its own distinct set of methods rather than borrowing wholesale from small-molecule design. The tools that matter here—protein language models, developability predictors, and structure-based generative methods—are built around the specific structure of the antibody problem.

Antibody design is not small-molecule design applied to bigger molecules. The sequence space is vastly larger, developability constraints are severe, and binding depends on loop structures, which is why biologics AI evolved its own toolkit.

Protein language models for antibody sequences

The foundation of much current antibody AI is the protein language model (PLM), which applies the same principles as a text language model to protein sequences, learning the statistical patterns of amino acid sequences from large databases. Trained on millions of natural protein or antibody sequences, a PLM develops an implicit understanding of what makes a sequence plausible and functional, which can then guide design. A landmark demonstration showed that a structure-informed protein language model could guide the evolution of antibody complexes without supervision, improving binding and other properties by suggesting mutations the model considered favorable.

How protein language models are used in antibody work:

• Guided affinity maturation. Starting from an antibody with at least weak binding, a PLM can suggest sequence changes likely to improve affinity, guiding the optimization more efficiently than random mutagenesis.
• Sequence scoring and screening. A model's assessment of how natural or plausible a sequence is, often measured by perplexity, can screen candidate antibodies, flagging sequences likely to behave poorly before they are made.
• Humanization and liability removal. PLMs can help make an antibody more human-like to reduce immunogenicity and can flag sequence liabilities associated with instability or aggregation, supporting the engineering that turns a binder into a drug candidate.
• Foundation for generative design. The same models that score and optimize sequences also underpin generative approaches that propose new antibody sequences, making PLMs the shared substrate across much of the field.

A crucial nuance, and one the honest literature is careful about, is that most successful PLM applications so far start from an antibody that already has at least weak function and improve it. That is genuinely useful; affinity maturation is a real and common task, but it is a different and easier problem than designing a binder from scratch, a distinction that matters when reading claims about AI antibody design.

Developability prediction

If any antibody AI application has clear, near-term practical value, it is developability prediction. The reason is economic and painful: an antibody can bind its target beautifully and still fail as a drug because it aggregates, is unstable, expresses poorly, or provokes an immune response, and discovering those failures late is expensive. A 2026 study applying protein language models to antibody developability prediction showed that PLMs can support developability assessment in real industrial discovery pipelines, both through supervised prediction of specific properties and through unsupervised screening using sequence perplexity.

What developability prediction covers:

• Aggregation propensity. Antibodies that aggregate are difficult to manufacture and can be immunogenic, so predicting aggregation risk from sequence and structure lets teams deprioritize risky candidates early.
• Stability and expression. Models predict whether an antibody will fold stably and express at usable levels, properties essential to manufacturability that are otherwise discovered only through experiment.
• Solubility and viscosity. High-concentration formulation, needed for many antibody drugs, demands good solubility and manageable viscosity, which predictive models can flag from sequence.
• Immunogenicity risk. Predicting whether an antibody sequence is likely to trigger an unwanted immune response supports humanization and helps avoid a failure mode that can end a program.

The value of developability prediction is that it front-loads failure. Instead of discovering a manufacturing or stability problem after months of work, a team can weigh developability alongside binding from the start, choosing candidates that are more likely to survive the whole journey to a drug. That is a concrete, unglamorous and genuinely useful contribution.

Epitope-directed design

A more ambitious goal is epitope-directed design: creating an antibody that binds a specific, chosen site on a target antigen. This matters therapeutically because where an antibody binds often determines whether it works; blocking a particular interaction, for example, requires binding a particular region. Recent methods have begun to tackle this directly. The Germinal method for de novo epitope-targeted antibody design generates antibody CDR loops against a chosen epitope without any starting binder, a genuine step toward designing antibodies from scratch rather than only optimizing existing ones.

Continue reading below...

WebinarsDecoding immune–tumor interactions with functional genomics

Discover how coculture models and CRISPR tools reveal new insights into tumour microenvironments.

Read More

What epitope-directed and de novo design involve:

• Targeting a specific site. Rather than finding any binder, epitope-directed design aims at a defined region of the antigen, which is what gives an antibody its specific mechanism of action.
• De novo CDR generation. The most ambitious methods generate the binding loops from scratch against a target, conditioned on the antigen structure, rather than starting from a known antibody, expanding what is designable in principle.
• Structure-conditioned generation. These approaches lean on predicted antigen and complex structures, connecting antibody design directly to the structure-prediction advances that made reliable modeling of protein complexes possible.
• Honest success rates. The field reports its binding rates candidly, and they show both progress and difficulty; one benchmark noted generative models achieving roughly a 10.6% binding rate for heavy-chain CDR3 designs and about 1.8% for full heavy-chain CDR designs, meaningful given the difficulty but far from routine success.

Those single-digit to low-double-digit binding rates are the honest measure of where de novo antibody design stands: real, improving and clearly better than random, but not yet a reliable, push-button capability. A method that produces a binder for roughly one in ten designs against a favorable target is scientifically impressive and practically useful as a source of leads, while still being far from replacing established discovery approaches.

First AI-assisted antibodies in the clinic

The question everyone asks is whether any of this has produced actual drugs, and the honest answer is that AI-assisted antibodies are beginning to enter clinical development, though the field is early and claims deserve careful reading. The broader significance of computational protein design was underscored when the 2024 Nobel Prize in Chemistry recognized computational protein design and structure prediction, awarded to David Baker, John Jumper and Demis Hassabis, a signal of how far the underlying science has come.

How to read the clinical picture accurately:

• AI-assisted, not fully AI-designed. Most antibodies described as AI-derived were optimized or partly designed with AI within a broader discovery effort, rather than generated entirely from scratch by a model, an important distinction often blurred in coverage.
• Early clinical stage. The AI-assisted antibodies reaching the clinic are in early trials, and as with AI-discovered small molecules, early-stage entry is encouraging but does not yet demonstrate that AI produces more successful drugs.
• Affinity maturation is the mature use. The clearest clinical contributions so far come from AI-guided optimization of existing antibodies, the affinity-maturation task where the methods are most reliable, rather than from fully de novo design.
• The translational question remains open. As across AI in drug discovery, whether AI-assisted antibodies ultimately succeed in patients at higher rates than conventionally discovered ones is not yet answerable, because too few have progressed far enough.

The balanced reading is that AI is contributing to antibodies now entering clinical development, most credibly through optimization rather than pure de novo design, and that the ultimate test, better clinical success, lies ahead. The trajectory is real; the verdict is pending.

Combining generative AI with experimental affinity maturation

The theme that unifies successful antibody AI is that computation and experiment work together rather than one replacing the other. The most productive programs use AI to propose and prioritize, then use high-throughput experimental methods to test, measure, and refine, feeding results back to improve the models.

How the integrated loop works in practice:

• Generate and prioritize computationally. AI proposes candidate sequences or designs and ranks them by predicted binding and developability, focusing experimental effort on the most promising, exactly the prioritization benefit seen across AI in drug discovery.
• Test at high throughput. Experimental display and screening technologies measure binding and properties for many candidates in parallel, providing the ground truth that computational prediction cannot fully supply.
• Refine with the data. Experimental results feed back to improve the models and guide the next round, an active-learning loop that tightens with each iteration and is where the combination outperforms either approach alone.
• Reserve de novo for where it fits. Fully de novo design is used where it adds most, opening targets or epitopes that are hard to address otherwise, while affinity maturation and optimization carry much of the routine load.

The balanced conclusion is that AI has become a genuine and growing part of antibody discovery, strongest today in developability prediction and AI-guided optimization, advancing but still early in fully de novo epitope-directed design, and most powerful when tightly coupled to experiment. For how antibody design fits the wider molecular design landscape, DDN's generative AI in molecular design hub and its guide to AI across drug discovery provide the surrounding view.

This article was produced in accordance with Drug Discovery News’ Editorial Policies.