In 2020, a London-based AI lab called DeepMind solved a problem that had defeated biologists for fifty years. The protein folding problem - given the sequence of amino acids that makes a protein, predict the three-dimensional shape it will fold into - had been the field's most stubborn challenge since Anfinsen won the Nobel Prize in 1972 for showing that the shape was theoretically predictable from the sequence. AlphaFold 2 solved it. Within a year, it had predicted the shapes of 200 million proteins. It changed biology permanently. Now comes the next step - and it is bigger.
The AlphaFold Moment, Revisited
From Reading Biology to Writing It
To understand why what is happening now is significant, it helps to understand precisely what AlphaFold did - and what it did not do. Protein folding prediction takes a sequence of amino acids and outputs the three-dimensional structure of the resulting protein. It is essentially a reading problem: given the code, determine the form. AlphaFold solved this reading problem with near-experimental accuracy, at a scale and speed that transformed structural biology overnight.
What AlphaFold did not do is design new proteins. It could tell you what any known sequence would look like - but it could not tell you which sequence to design if you wanted a protein with a specific function. The step from "reading the shape of what exists" to "writing the shape of what should exist" is the step from decoding to authorship. That step is now being taken.
What a Protein Binder Is
A protein binder is a molecule - itself a protein - specifically designed to bind to the surface of another protein with high precision and affinity. Think of it as a custom-made lock pick engineered to fit one specific lock. Binding a target protein can block its function (useful for blocking disease pathways), flag it for immune system destruction (useful in cancer), or deliver a therapeutic payload to a precise location. The ability to design binders reliably and rapidly - for almost any target - is the key step toward a new generation of precision medicines.
What has changed in 2025-2026 is the arrival of generative AI models that can design binder proteins de novo - from nothing, without a starting template. You specify the target: a protein surface involved in cancer proliferation, a viral coat protein, a pathological receptor in neurodegeneration. The model generates candidate binder sequences. You synthesize them, test them, and increasingly - as experimental success rates climb - you find that they work.
The New Models
The Tools That Write Molecules
Several distinct AI systems are now capable of protein binder design, each with different architectures and strengths. Together they represent the practical toolkit of what structural biologists are calling the post-AlphaFold era.
| System | Developer | Architecture | Strength |
|---|---|---|---|
| BoltzGen | MIT / independent | Generative diffusion | First model to produce drug-pipeline-ready binders from scratch |
| Proteina-Complexa | NVIDIA | Flow-matching generative | High-affinity binders for challenging and disordered targets |
| IARA | Academic consortium | Graph neural network | Rapid structural filter - triages large candidate libraries at speed |
| Dyno Psi-Phi | Dyno Therapeutics | Agentic AI suite | End-to-end binder design from target specification to sequence |
| RFdiffusion | UW / David Baker lab | Diffusion model | Established benchmark; designed binders for cancer and toxin neutralization |
Figure 1 — Key generative AI systems in protein binder design as of May 2026
BoltzGen represents a specific milestone: it is described as the first model capable of generating novel protein binders genuinely ready to enter the drug discovery pipeline without manual redesign. Previous generative models could produce plausible structures, but experimental success rates were low enough to require considerable human optimization. BoltzGen's success rate is substantially higher, moving the technology from a promising research tool to a practical drug discovery instrument.
NVIDIA's Proteina-Complexa extends the capability further, demonstrating high-affinity binding design for targets previously considered extremely difficult - including intrinsically disordered proteins (proteins without a fixed structure) and small-molecule co-factor binding sites. These "hard targets" were the frontier cases for structure-based drug design; generative AI is beginning to solve them.
"We are entering a phase where protein binder design is becoming a scalable engineering discipline, not an artisanal research activity. The experimental success rates we are now seeing from generative models were unimaginable three years ago."
- Luciano Abriata, PhD — "Two upcoming AlphaFold moments," Medium / Advances in Biological Science, May 2026Behind all of these models lies the same fundamental transformation: the shift from searching through known protein sequence space to generating novel sequence space. Classical drug design searched for binders among existing molecules or slight variants of them. These models generate sequences that have never existed before - proteins with no evolutionary precedent, optimized mathematically for a single purpose: binding a specific surface with maximum affinity and specificity.
The Wider Context
Structural Biology's Second Revolution
Protein binder design does not exist in isolation. It sits alongside a second emerging frontier in structural biology: conformational state prediction. Proteins are not static structures. They flex, shift, and adopt different shapes depending on what they are bound to, what pH they are in, whether they have been phosphorylated. AlphaFold 2 predicted the most stable state of a protein. The next generation of models is beginning to predict the full conformational landscape - all the different shapes a protein can adopt - which matters enormously for drug design, because drugs often work by locking a protein into one specific state.
The David Baker laboratory at the University of Washington, which pioneered the RFdiffusion approach to protein design, has demonstrated binders that neutralize biological toxins, modulate immune pathway proteins, and engage receptor targets with affinities below 10 nanomolar - extraordinarily tight binding that rivals the best antibodies. A research team using evolutionary algorithms published in May 2026 showed successful design of potent, highly selective cancer biologics targeting Nectin-4, a protein overexpressed in several aggressive cancer types. The binders were designed without a starting template and validated experimentally in weeks.
Why Binding Affinity Matters
A binder's affinity is measured by its dissociation constant (Kd) - how strongly it holds to its target. An antibody typically has a Kd around 1-10 nanomolar (nM). AI-designed binders are now routinely achieving sub-10 nM affinities. Lower Kd means tighter binding means more drug effect per dose means lower required concentrations means fewer side effects. The difference between a Kd of 1 micromolar and 1 nanomolar is the difference between a drug candidate and a rejected compound. AI is moving the field toward 1 nM as a default, not an achievement.
The deep learning predictor of bindable protein surfaces, published in pre-print in April 2026, represents another capability: given any target protein, the system identifies which regions of its surface are the most "bindable" - the geometric and electrostatic properties that make a surface amenable to tight, specific binding. This acts as a map for the generative models, telling them where to aim before they design what to throw.
What This Changes
The Implications for Medicine
The implications of reliable, rapid, de novo protein binder design are difficult to overstate. Drug discovery has traditionally been measured in decades and billions of dollars. The attrition rate from initial candidate to approved drug is roughly 90% - nine out of ten promising molecules fail somewhere in the development pipeline, usually because they don't work well enough, or because they interact with something they shouldn't. A tool that designs better, more specific binders from the start changes the mathematics of that attrition.
The most immediate application is in oncology, where receptor-targeting biologics - proteins designed to bind cancer-specific surface markers and either block them or recruit immune cells to destroy the cancer - have transformed treatment for several cancer types. Every new cancer biomarker discovered is now a potential target for an AI-designed binder. The pipeline, which previously required years of structural biology and medicinal chemistry for each target, is beginning to compress to weeks.
Beyond cancer: infectious diseases, where viral surface proteins are natural binder targets; neurodegeneration, where amyloid aggregates and tau tangles present binding surfaces; autoimmune disease, where cytokine pathway proteins regulate inflammatory signaling. Almost every major disease category has proteins that, if bound correctly, could be modified toward therapeutic effect. The bottleneck was always designing binders that actually worked. That bottleneck is opening.
There is a longer-horizon implication that structural biologists are beginning to discuss openly: the possibility of designing entirely new protein architectures that evolution never produced - proteins optimized not for reproductive fitness but for specific human therapeutic needs, using secondary and tertiary structures that have no natural precedent. We are at the beginning of this. But the AlphaFold moment taught the field that "beginning" can compress very quickly when the right model arrives. The models are arriving.
Buy me a coffee