The Evidence Pack: How Generative AI is Designing 'De Novo' Antibodies from Scratch

For the entire history of modern medicine, drug development has been a process of discovery. Researchers either found useful molecules in nature or screened millions of chemical compounds in massive libraries to see what happened to stick to a disease target. Today, the paradigm is fundamentally shifting from discovery to generation. Driven by the same underlying mathematics that power AI image generators, biological models are now designing entirely new proteins and antibodies from scratch—molecules that have never existed in the natural world.[3][7]

This transition was the dominant theme at the 2026 BIO International Convention, where generative AI was no longer discussed as a futuristic buzzword, but as a deployed tool reshaping biopharma pipelines. Major pharmaceutical companies are aggressively partnering with specialized AI startups to overhaul their research and development processes, moving away from traditional wet-lab screening toward in-silico design.[2][5]

A prime example emerged this week when Eli Lilly announced a significant investment in Absci, an AI drug creation company. The partnership aims to develop novel medications for conditions ranging from hair loss to endometriosis. By leveraging generative models, these companies are attempting to design highly specific biologics for targets that have historically frustrated human researchers.[1]

To understand the evidence behind this shift, it is crucial to distinguish between structure prediction and generative design. A few years ago, systems like AlphaFold revolutionized biology by predicting the 3D shapes of existing proteins based on their amino acid sequences. The new wave of models, such as RFdiffusion, does the exact opposite: researchers input a desired 3D shape or binding target, and the AI generates the amino acid sequence required to build it.[3][7]

The strongest pre-clinical evidence for this technology lies in "zero-shot" antibody design. In traditional development, scientists often start with an existing antibody—perhaps extracted from an immunized mouse or a human survivor of a disease—and tweak it. Zero-shot design means the AI generates a functional, high-affinity binder on the very first try, using only the digital structure of the disease target as a prompt.[4][7]

Peer-reviewed studies in journals like Nature and Science have validated this capability in laboratory settings. Researchers have successfully used diffusion models to generate de novo proteins that bind tightly to challenging targets, including viral spike proteins and complex cancer receptors. When these AI-designed sequences are synthesized and tested in a petri dish, they fold exactly as the computer predicted and bind to their targets with remarkable precision.[3][4]

The second major claim surrounding generative biologics is a drastic reduction in pre-clinical timelines. Historically, developing a therapeutic antibody involved immunizing animals, waiting for an immune response, harvesting the antibodies, and spending years optimizing them for human use. This process is notoriously slow, expensive, and prone to failure.[5][6]

Evidence from early adopters suggests that generative AI can compress this timeline significantly. Biotech firms report cutting the time from target identification to lead candidate selection by up to 70%. Startups like Absci claim their platforms can design, synthesize, and validate a de novo antibody in under six weeks—a fraction of the years it typically takes using traditional screening methods.[1][5]

The third area of evidence centers on targeting the "undruggable." Many severe diseases are driven by proteins that lack obvious binding pockets or have highly complex structures, making them nearly impossible to target with conventional drugs. Human researchers and traditional screening libraries often fail to find a molecule that fits.[6][7]

Generative models, however, can explore the entire theoretical landscape of protein structures. They can design bespoke molecules that wedge into microscopic crevices or bind to flat surfaces on a target protein. This capability is particularly relevant for complex, multi-factorial conditions like endometriosis, where traditional drug discovery has yielded few breakthroughs over the past decade.[1][3]

Despite the overwhelming success in the digital realm and the petri dish, the clinical evidence remains the primary area of uncertainty. The translation gap between in-silico design and in-vivo success is the ultimate test for generative biologics. Biology is profoundly messy, and the human body is an incredibly hostile environment for foreign proteins.[6][7]

A major hurdle is immunogenicity. An AI-designed antibody might bind perfectly to a cancer cell in a computer simulation, but if the human immune system recognizes the de novo protein as a foreign threat, it will attack and neutralize the drug before it can work. Researchers are currently training models to "de-risk" these sequences by predicting and removing immunogenic features, but the true efficacy of this approach will only be known once large-scale human trials conclude.[4][6]

Furthermore, while generative AI drastically accelerates the discovery phase, it does not bypass the clinical trial bottleneck. Phase 1, 2, and 3 trials still require years of careful observation in human patients to ensure safety and measure long-term outcomes. The speed of AI cannot override the biological reality of how long it takes to prove a drug works in a living population.[5][6]

The next 24 to 36 months will be critical for the field. As the first wave of purely generative, zero-shot antibodies moves out of the laboratory and into human clinical trials, the biopharma industry will get its first real look at whether these engineered molecules can survive the rigors of human biology.[2][7]

If the clinical data matches the pre-clinical promise, the implications are staggering. Generative AI will have successfully transformed drug discovery from an artisanal craft of trial and error into a predictable engineering discipline, fundamentally rewriting the economics and speed of modern medicine.[5][7]