A neural network just designed a protein that grabs the cancer drug exatecan with 100 percent success rate on the first try. The team at Dana-Farber Cancer Institute, led by Nicholas Polizzi, did not screen thousands of candidates. They asked a closed loop of two neural networks to design a small number of high-affinity binders from scratch. The tightest binders hit nanomolar-to-picomolar affinities, surpassing the next-leading method by 70-fold for exatecan and nearly 10,000-fold for apixaban.
This is not an incremental gain in computational biology. It is a signal that the unit economics of protein design have inverted. When a neural network can outperform experimental screening on success rate, affinity, and speed simultaneously, the capital-intensive wet-lab pipelines that dominated the last two decades become a liability.
The brute-force era just ended
Directed evolution and high-throughput screening built modern protein engineering. The playbook was simple: generate a massive library of variants, test them all, pick the winners, repeat. It worked. It was also slow, expensive, and produced mountains of failure for every hit. For small-molecule drug binders, the problem is harder still because the ligand adds a third dimension of chemical complexity that most design algorithms cannot handle.
NISE replaces that loop with two neural networks that talk to each other. LASErMPNN, a graph neural network, designs protein sequences for a given backbone and docked ligand. A structure predictor, Boltz-2x, models the resulting 3D protein-ligand complex. The output feeds back into the sequence designer. The loop iterates, optimizing sequence-structure-ligand compatibility without any experimental input.
The result, published in Nature on June 24, 2026, is a method that allows researchers to start with a drug of interest and quickly generate a small number of promising proteins for laboratory testing, rather than screening thousands of possibilities by trial and error. The closed-loop neural optimization outperformed a comparable design loop using a physics-based energy function. That detail matters: neural networks are not just faster here; they are finding solutions that physics-based scoring functions miss.
Two mutations. Zero experiments. A 100-fold better binder.
The most jarring number in the paper is not the 100 percent success rate. It is what happened after the initial designs were made. LASErMPNN suggested two amino-acid substitutions that improved the affinity of the tightest exatecan binder by 100-fold without any experimental input. Two mutations. Computationally predicted. A 100-fold affinity jump.
In a traditional directed evolution campaign, a 100-fold improvement might take multiple rounds of mutagenesis and screening, consuming months and hundreds of thousands of dollars in reagents and labor. NISE delivered it in silico before anyone touched a pipette.
A shield for a fragile drug
Exatecan is a potent cancer drug with a well-known liability: its lactone ring hydrolyzes quickly in the body, limiting its therapeutic window. The NISE-designed protein did not just bind exatecan. It wrapped around the drug and protected that fragile lactone ring from hydrolysis for days.
This is functional protein design, not just binding. The protein acts as a shield, stabilizing a small molecule that would otherwise degrade. The immediate implication is a drug delivery vehicle that could extend exatecan's half-life and reduce off-target toxicity.
For apixaban, a widely prescribed blood thinner, the designed binder achieved an 83 percent success rate. The Dana-Farber team describes it as a starting point for a future antidote that captures apixaban and neutralizes its blood-thinning effect. No such antidote exists today. Apixaban-related bleeding events are managed with supportive care and, in some cases, a costly reversal agent with its own thrombotic risks.
These potential uses have not yet been tested in animals or patients. But the prototype proteins exist. They were designed computationally, produced in the lab, and characterized. The gap between a computational design and a functional binder has collapsed.
The cost structure is the real disruption
The consensus will celebrate NISE as a breakthrough for drug delivery. The real disruption is upstream, in the cost structure of the protein design industry.
NISE can design binders without experimental training data for each target. That zero-shot capability commoditizes the early-stage protein design process. The unit economics are brutal for the incumbent approach. NISE requires compute, which is cheap and getting cheaper. High-throughput screening requires robots, reagents, consumables, facilities, and skilled operators. The cost per design candidate is orders of magnitude lower in silico. When the in silico method also produces better binders, the economic argument for maintaining large experimental screening infrastructure weakens.
This is not a prediction that wet-lab work disappears. It is a prediction that the ratio of capital allocation shifts. The open-source code is already available on GitHub with 50 stars and 13 forks. The barrier to entry is a GPU cluster and some protein engineering expertise. Big pharma has both.
The first-order effect is straightforward: within 12 to 24 months, at least two major pharmaceutical companies will launch internal programs using NISE or a direct competitor. One of them will file a patent for a NISE-designed protein as a drug delivery vehicle or antidote. That filing will trigger a licensing rush for the underlying neural networks. The Dana-Farber team's LASErMPNN architecture will become a must-license asset.
The second-order effect is where the real restructuring happens. Capital currently locked into high-throughput experimental pipelines will be reallocated to in silico design cycles. R&D departments that do not restructure will find themselves spending more to get worse results than competitors who adopt the new workflow. The legacy screening infrastructure becomes a stranded asset—not overnight, but faster than most R&D leaders expect.
The third-order effect reshapes the competitive landscape. The barrier to entry for designing drug-binding proteins drops far enough that smaller biotechs and academic labs can compete with big pharma in therapeutic protein design. The NISE preprint has only 6 citations as of mid-2026. The method is not yet widely adopted. The window for first-mover advantage is narrow, but it is open.
What would falsify this prediction? If the zero-shot success rates do not generalize to a broader range of small-molecule targets—if exatecan and apixaban turn out to be cherry-picked examples—then the economic cascade stalls. But the 100 percent and 83 percent success rates on two chemically distinct drugs suggest the method is robust. The smart money bets on generalization.
What to do now
Pharma R&D leaders should audit their current small-molecule binding protein pipeline and identify candidates that could be redesigned with NISE. The method is published, the code is open source, and the proof points are exatecan and apixaban. The question is not whether the approach works. It is who moves first.
Investors should look for startups licensing or building on the LASErMPNN architecture. The neural network that can suggest two mutations for a 100-fold affinity improvement without any experimental input is the core intellectual property. Whoever commercializes that capability for a broad range of small-molecule targets will capture value that currently flows to CROs and internal screening groups.
Drug developers should start experimenting with the NISE code. The GitHub repository has 13 releases. The installation is documented. The compute requirements are not trivial, but they are a fraction of the cost of a single high-throughput screening campaign. The first teams to integrate NISE into their design workflows will have a lead that compounds.
The protein that grabbed the drug on the first try
The neural network designed a protein that grabbed exatecan on the first try. That protein is now a prototype for a new class of drug delivery vehicles and antidotes. The loop is closed. The future of protein design is not screening millions of candidates. It is asking a neural network to design the perfect binder in one shot, and then watching it work.
