AI-Designed Antibody-Drug Conjugate Platforms Enter Clinical Validation

BigHat Biosciences calls its platform Milliner, a reference to the craft of hat-making where raw materials are shaped into something precisely fitted. The name is apt. Its system integrates machine learning models with a synthetic biology wet lab that designs, builds, and tests hundreds of antibody variants per week in closed-loop cycles. The company has raised $105 million from Section 32, Andreessen Horowitz, and others. It has signed partnerships with AbbVie (up to $355 million in milestones), Eli Lilly (GI cancer antibody-drug conjugate program), and Merck. It is not a startup anymore.

BigHat is one of perhaps fifteen companies — Absci, Generate:Biomedicines, LabGenius, Antiverse, Nabla Bio — that form a new category: AI-native antibody design platforms. They share a premise: machine learning can replace brute-force screening, cutting timelines from years to months. The question, as several of these companies approach clinical data for the first time, is whether the premise holds.

The case for AI-designed antibodies

The conventional process for discovering a therapeutic antibody starts with immunizing an animal or screening a phage-display library — essentially, a lottery where you buy as many tickets as your automation budget allows. A typical campaign screens 10⁷–10⁹ candidates, then winnows through successive rounds of affinity maturation, each requiring months of manual engineering. The failure rate is high: roughly 90% of antibody candidates that enter Phase 1 never reach approval.

AI platforms flip the sequence. Instead of screening first and designing later, they model the antibody-antigen interface in silico, generate designed sequences, test them in an automated wet lab, and feed the results back to retrain the model. Its Milliner platform can iterate through this design-build-test loop at a rate of hundreds of antibodies per week. Over four years, the company has processed tens of thousands of designs across hundreds of cycles.

The economic argument is straightforward: each cycle costs less than a comparable round of manual engineering, and the model improves with every iteration. Absci's integrated drug creation platform, which combines generative AI with wet lab validation, has delivered functional antibody candidates against dermatology targets for Almirall in under 12 months — a timeline that conventional discovery would struggle to match in half again that time.

The validation gap

The counterargument is equally direct: none of these platforms has produced a drug that reached a patient.

Every approved antibody-drug conjugate today — Enhertu, Trodelvy, Kadcyla, Adcetris — was discovered through conventional methods: hybridoma technology, phage display, or transgenic mice. AI platforms have generated plenty of candidates and plenty of partnership revenue, but zero approved molecules. The closest any has come is its GI cancer ADC program, currently in IND-enabling studies and expected to enter clinical trials this year. Until data from those trials is public, the superiority of AI-designed antibodies remains a premise, not a result.

The caution is not skepticism. It is a historical pattern that applies to every platform technology in biotech. In 2024, Nature Biotechnology published a review noting that AI-designed antibodies had, at that point, produced exactly zero clinical candidates. The 2026 picture is better — multiple programs approach the IND stage — but the fundamental gap between computational promise and clinical reality has not yet closed.

The deal flow is real

Whatever the clinical uncertainty, pharma companies are voting with cash. AbbVie paid the company $30 million upfront for access to its Milliner platform. Eli Lilly made an equity investment and is backing its internal GI cancer ADC. AstraZeneca and Absci signed a $247 million collaboration. AbbVie also signed with it. Merck, Janssen, and Amgen have all done deals with platform companies in this space.

The structure of these deals matters. Pharma partners pay platform access fees, not equity stakes in a pipeline. The pharma partner pays for the right to use the AI model on its own targets, retains control of the resulting candidate, and pays milestones if the candidate advances. Its deal with Lilly bundles platform access with support for its own GI cancer ADC, but even that internal program was started using data and models developed through partnership work.

This model de-risks the pharma side: they pay for a tool, not a bet on a single molecule. But it also means the platform companies are not primarily drug developers. They are platform vendors whose customers happen to be drug developers. The distinction matters because it changes the benchmark for success. A platform company can be commercially viable, generating recurring partnership revenue, without ever producing an approved drug. Whether that is a good outcome for investors depends on the terms of each deal.

Where the convergence matters

What separates this from a standard biotech story is the AI infrastructure layer underneath it. Designing an ADC requires optimizing four variables simultaneously: the antibody's binding affinity, the linker's stability in circulation, the payload's potency, and the drug-to-antibody ratio. Each variable occupies a multidimensional space that conventional engineering explores one point at a time.

AI models, particularly graph neural networks trained on structural and biophysical data, can explore these spaces in parallel. The same transformer architectures that power large language models are being adapted to predict antibody-antigen binding, linker cleavage rates, and payload release kinetics from sequence data alone. As we wrote in July, the $347 million flowing into AI protein design platforms is funding infrastructure — models, datasets, closed-loop labs — that applies across modalities, not just antibodies.

The convergence between AI infrastructure and ADC biology is not theoretical. Generate:Biomedicines, which raised $400 million in its February 2026 IPO, uses generative diffusion models to design antibodies and other therapeutic proteins. The same class of models that produces images produces binding interfaces. Absci's platform integrates Oracle Cloud Infrastructure and AMD accelerators to run inference at scale — the compute stack behind an antibody campaign now looks more like a language model training run than a biology experiment.

What to track