AI Protein Design Market (2026-2036)

The growth of this market is mainly driven by the shift of commercial value from predicting what a protein looks like to generating one that does not yet exist, since structure prediction is now available free of charge at a scale that removes it as a basis for a business, by pharmaceutical partnership structures that pay a design company upfront and research payments before any molecule reaches a clinic, and by the extension of the same generative methods from antibodies, where the market first formed, into enzymes, peptides, binders and delivery proteins where the customer sits outside the pharmaceutical industry. Design is cheaper than discovery by screening because it replaces the construction and testing of very large libraries with the construction and testing of small sets of candidates, and the cost saving accrues to whoever runs the laboratory rather than to the software vendor, which shapes how this market is priced.

However, factors such as the requirement that a designed protein be made and tested before it has any value, the small share of created value that a company selling design alone can capture, the dominance of contingent clinical milestones in partnership headline values, and the resulting gap between contracted value and revenue recognised in any year are expected to restrain the growth of this market to some extent. Most suppliers have taken on laboratory operations, partnership risk or an internal pipeline in order to capture more of what they create, and because milestones contracted today will not be reached inside the forecast period the market is far smaller than announced deal totals suggest.

The AlphaFold Protein Structure Database, developed by Google DeepMind and EMBL-EBI, which provides over 260 million protein structure predictions on open access terms and states that the data is available for academic and commercial use under a CC-BY-4.0 licence removed structure prediction as a commercial proposition. A company cannot charge for predicting the shape of a protein when over a quarter of a billion predictions are already published and licensed for commercial use at no cost, and the recognition of that achievement in the Nobel Prize in Chemistry 2024, awarded one half to David Baker for computational protein design and one quarter each to Demis Hassabis and John Jumper for protein structure prediction confirmed the problem as solved rather than as a competitive frontier. Value moved to the inverse question, which is generating a sequence that does not occur in nature and that folds into a specified shape and performs a specified function. That question is open, is not answered by any public database, and is the activity this market sells.

Design has no value until a designed protein has been made and tested. A generated sequence is a proposal, its properties are predicted rather than observed, and the proportion of designs that express, fold, bind and remain stable is established only by building them, so every commercially serious participant in this market either operates laboratory capacity or contracts for it. The data produced by that testing is the asset that compounds, because a model improves on experimental results generated against its own designs in a way it cannot improve on public sequence and structure data that every competitor also holds. Proprietary experimental data rather than model architecture is therefore the defensible position, and the industry has organised itself around that fact rather than around any algorithmic difference between suppliers.

Partnership structures rather than software licences carry most of the revenue today. The two pharmaceutical collaborations announced by Isomorphic Labs, under which the company received an upfront payment of USD 45 million from Eli Lilly and Company and USD 37.5 million from Novartis, with the partnerships stated as having the potential to be worth nearly USD 3 billion to the company in total, excluding any royalties that may result from future drug sales illustrate the economics precisely: the upfront payments total USD 82.5 million against a stated potential of nearly USD 3 billion, so slightly under three per cent of the headline value was payable at signature and the remainder depends on research, development, regulatory and commercial milestones that take years to reach. A market sized on announced deal totals would therefore be overstated by more than an order of magnitude, and the sizing in this report counts upfront payments, research funding, recognised milestone payments and software and service fees in the year they are earned, which is why the market is smaller than the deal announcements in this field suggest.

Antibodies are where the market formed and enzymes are where it is spreading. An antibody is a constrained design problem with a well understood scaffold, a large body of prior structural and sequence data and a customer able to pay pharmaceutical prices, which is why generative methods were applied there first and why most funded companies began with antibody discovery. Enzymes present a different commercial situation, since an industrial customer buys a catalyst to perform a reaction at a stated temperature, pH and substrate concentration and judges the result against a yield it can measure in a week rather than against a clinical endpoint a decade away. The feedback loop is shorter, the payment is for delivered performance rather than for optionality, and the buyer is a chemical or food company with no interest in the design method, which is why industrial enzymes are the fastest-growing application even though pharmaceutical discovery remains much the largest.

The U.S. Food and Drug Administration draft guidance Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products, published in January 2025 as a Draft Level 1 guidance carrying the notice not for implementation and non-binding recommendations, issued by the Center for Drug Evaluation and Research with other offices and centres under docket FDA-2024-D-4689, which sets out a risk-based approach for establishing and evaluating whether an artificial intelligence model is credible for a specified context of use is the relevant regulatory development, and what it does not do matters as much as what it does. There is no distinct approval pathway for a protein because it was designed computationally, so a designed antibody or enzyme intended as a medicine is reviewed as the biologic it is, on its own non-clinical and clinical evidence, and the design method is a development tool rather than a regulated article. The guidance addresses the separate question of when a model output may support a regulatory decision, and its risk-based framing ties credibility to a specified context of use rather than to the model in general, which means a company cannot qualify a design model once and rely on that qualification across submissions. The guidance remains draft and non-binding, so sponsors are building credibility documentation against recommendations that are not final.

The European Union Artificial Intelligence Act, which entered into force on 1 August 2024, applied its governance rules and obligations for general-purpose artificial intelligence models from 2 August 2025 and became generally applicable on 2 August 2026, with high-risk systems embedded in products regulated by the sectoral legislation listed in Annex I applying from 2 August 2028 after the extension of that transition period by the AI Omnibus applies on the European side, and the obligations that reach this field first are those attaching to general-purpose models rather than those attaching to high-risk systems, since a protein design foundation model is a general-purpose model while the medicine eventually developed from it is regulated under pharmaceutical law rather than as an artificial intelligence system. Documentation, training-data summaries and copyright policy obligations therefore fall on model developers on a timetable independent of any product approval, and that asymmetry has encouraged suppliers to characterise their offering as a design service rather than as a model placed on the market.