The Antibody Engineers

The therapeutic antibody business is one of the most successful franchises in the history of medicine. Humira (adalimumab), which topped global drug sales charts for a decade before its patents expired, is a monoclonal antibody. Keytruda (pembrolizumab), which has essentially redefined treatment for a dozen cancer types since 2014, is a monoclonal antibody. So is Dupixent, Ocrevus, Crizanlizumab, and dozens of other products generating billions in annual revenue across inflammation, oncology, and rare disease.

These drugs work because antibodies are extremely good at binding specific molecular targets with high affinity and high selectivity — properties that small molecules struggle to achieve for complex, large-surface-area targets. The immune system spent 500 million years evolving the antibody scaffold. It is an extraordinarily refined molecular recognition system.

AI is now redesigning that scaffold from the ground up, producing proteins with properties that 500 million years of evolution never optimized for, because the evolutionary pressures that shaped the immune system had nothing to do with pharmaceutical manufacturing, biophysical stability at room temperature, or half-life optimization for weekly subcutaneous injection.

What “AI-Designed Antibody” Means

When people say “AI-designed antibody” in 2026, they usually mean one of three things, which are quite different in their ambition and their evidence base.

The first category is AI-assisted optimization: taking an existing antibody identified through traditional discovery (hybridoma technology, phage display, transgenic mice) and using computational tools to improve specific properties — affinity, stability, half-life, reduced immunogenicity. This is the most mature and most widely deployed category. Virtually every major biologics company uses computational affinity maturation and in silico developability screening as part of their optimization workflow.

Related reading

The second category is AI-directed lead selection: using machine learning models to design or select antibody sequences with desired properties before experimental testing. Companies like AbSci, BigHat Biosciences, and Absci (yes, those are different companies with similar names — the antibody field’s naming conventions are not its strength) train models on large datasets of antibody sequences and their measured properties, then use those models to predict which untested sequences will have favorable characteristics. This dramatically reduces the experimental screen required to find a development candidate.

The third category — the most ambitious — is de novo protein design: generating entirely new protein sequences that are not variations on known antibodies but novel folds designed from scratch to bind a target. This is where David Baker’s laboratory at the University of Washington, and the commercial spinouts from that work, have been most active. RFdiffusion, a diffusion model trained on protein structures, can generate protein binders targeting essentially any surface — not just antibody-antigen interfaces but any protein-protein interaction, including “undruggable” surfaces that flat and featureless binding sites on protein faces that conventional antibodies cannot productively engage.

The De Novo Binder Frontier

The most scientifically striking work in de novo protein design has been in what Baker’s group calls “miniproteins” — small, stable protein scaffolds of 40-80 amino acids designed computationally to bind specific targets with high affinity. These are not antibodies; they’re more like designed peptides with antibody-level affinity, more stable and easier to manufacture than conventional monoclonals.

In published work through 2025, computationally designed binders against influenza hemagglutinin, SARS-CoV-2 spike protein, and several cancer-relevant receptors have achieved sub-nanomolar affinities in experimental validation. The design-to-experimental-validation cycle has compressed from years to weeks. The designs often work the first time — not every time, but often enough to make iterative experimental cycles far less burdensome than traditional approaches.

Generating Human Therapeutics (Xaira Therapeutics, which raised $1 billion in its 2024 Series A — one of the largest biotech fundraises in history) is built on this foundation. Their stated goal is to design novel proteins targeting previously inaccessible therapeutic targets, using AI to navigate the enormous space of possible protein sequences in silico before committing to expensive experimental work. Their pipeline is still preclinical, but the capital invested reflects high conviction in the technical approach.

The Developability Gap

The history of protein therapeutics is full of molecules that bind their target beautifully and then fail for reasons that have nothing to do with binding: they aggregate in the vial at room temperature; they provoke immune responses in patients; they have a two-hour half-life when a weekly injection requires a two-week half-life; they can’t be expressed at commercial scale without expensive cell culture conditions.

These “developability” properties — stability, immunogenicity, half-life, manufacturability — are what separate an interesting protein from a drug. They are also, collectively, where AI protein design is least mature. Training models to optimize for these properties requires large datasets of developability measurements for diverse protein sequences, and those datasets are harder to assemble than binding affinity data. Most experimental protein developability data lives in proprietary company databases, not in public repositories.

Several companies are building these datasets deliberately. Absci runs a platform that generates large-scale developability data (they’ve published datasets with millions of sequence-property measurements) and trains models on it. The bet is that the proprietary dataset, more than the model architecture, is the durable competitive advantage — you can copy a neural network architecture, but you can’t copy years of experimental measurements.

Related reading

The Bispecific Explosion

One area where AI design tools have had clear near-term impact is bispecific antibodies — engineered proteins that bind two different targets simultaneously. Bispecifics are more complex to engineer than conventional monoclonals because the two binding arms need to work cooperatively, not interfere with each other, and the whole molecule needs to retain acceptable stability and manufacturability.

This engineering challenge is tractable for AI. Models that predict antibody stability and aggregation propensity can guide the design of bispecific formats. Several FDA-approved bispecifics reaching the market in 2024-2026 (Tarlatamab for small cell lung cancer, Mosunetuzumab for follicular lymphoma, and others) were developed with significant computational design contributions, though calling them “AI-designed” rather than “computationally assisted” would overstate the autonomy.

The bispecific category is notable because it represents a target class that was genuinely difficult without computation. The engineering complexity of making two binding arms work together, in a stable molecule, at manufacturing scale, is high enough that trial-and-error experimental approaches are expensive and slow. Computational design makes bispecific engineering more tractable. That’s a case where AI tools expanded what’s practical, not just what’s fast.

The Access Question

A final observation that doesn’t get enough attention: therapeutic biologics, including antibodies, are among the most expensive drugs in medicine. Keytruda costs approximately $200,000 per patient per year in the United States. Dupixent, for a chronic condition like atopic dermatitis, is a lifelong treatment at similar annual cost. If AI-driven antibody engineering expands the number of treatable diseases, the implications for healthcare costs are ambiguous — more effective treatments may increase quality-adjusted life years, but the cost-per-treatment trajectory in biologics has not historically reflected manufacturing efficiency gains.

Biosimilar competition, which has somewhat reduced prices for older biologics like Humira, will eventually apply to AI-designed drugs just as to conventionally designed ones. But the window between approval and biosimilar entry is typically 8-12 years, during which the economics remain as they have been. AI making drug discovery faster doesn’t automatically make the resulting drugs more affordable. That is a policy problem, not a technology problem — and technology will not solve it on its own.

The CMC Challenge

Chemistry, Manufacturing, and Controls (CMC) — the regulatory module covering how a drug is made, tested, and assured to be consistent across batches — is a bottleneck in biologics development that AI has barely touched. Designing an antibody with desired binding properties is one challenge. Manufacturing it consistently at commercial scale, across multiple production runs, with analytical methods that confirm each batch meets specifications, is a different and deeply practical challenge.

Biologics manufacturing relies on living cell systems (Chinese hamster ovary cells are the workhorse of the industry, for historical reasons that have more to do with 1970s cell biology than any particular optimality) that are inherently variable. Cell culture conditions, media composition, dissolved oxygen levels, and dozens of other parameters affect glycosylation patterns, aggregation propensity, and product quality attributes. AI process analytical technology (PAT) tools are beginning to be applied to monitor and adjust bioreactor conditions in real time, and several large manufacturers have reported improvements in batch consistency and yield.

But the more fundamental CMC challenge — proving that a novel AI-designed protein can be manufactured consistently in the first place — requires years of process development work that AI can inform but not abbreviate. A protein that folds well in a computational model may not express efficiently in CHO cells; a designed protein with unusual structural features may aggregate during downstream purification. These are empirical problems that require experimental resolution. The path from a beautiful computer-designed protein to an approvable biological product runs through a manufacturing development process that is still fundamentally an experimental science.

Related reading

Where the Frontier Actually Is

The honest map of AI antibody engineering in 2026 shows a technology that has meaningfully improved optimization speed and expanded the practical scope of bispecific and multispecific design. It has not yet demonstrated that it can produce a fully AI-originating biologic from target-binding specification through approved manufacturing process. That complete path — design, optimize, develop, manufacture, approve — has not been demonstrated.

The companies working in this space are making credible progress on each component. The integration of those components into a reliable end-to-end pipeline is the actual frontier, and it will be defined not by the next impressive model benchmark but by the first AI-originating biologic to reach approved commercial production.

The near-term watch list includes Xaira Therapeutics’ first IND submission, expected in 2027 if their current programs stay on schedule; AbSci’s clinical entry for any of its wholly AI-generated antibody candidates; and the readout from BigHat Biosciences’ partnership programs with undisclosed pharma partners. These programs collectively represent the empirical test of whether de novo protein design translates from impressive academic demonstrations to approved therapeutics. The next 18-24 months will provide the first IND-stage data for wholly AI-designed biologics, and that data will set the narrative for AI protein engineering for years after.

The antibody is still the dominant therapeutic protein format because evolution spent 500 million years optimizing it. AI doesn’t need to beat evolution’s scaffold on every dimension — stability, immunogenicity, ease of expression — to be useful. It needs to beat it in the specific dimensions that matter for therapeutic efficacy: binding affinity for previously inaccessible surfaces, selectivity across closely related proteins, half-life tunable to the dosing schedule a patient can realistically follow. On those specific dimensions, the evidence is building that AI-designed proteins can match or exceed what natural antibody evolution produces. That’s enough of a wedge to build a generation of drugs around, even before the full end-to-end pipeline is demonstrated.