AlphaProteo: De Novo Protein Binder Design with AI

Last Updated: October 2026

For most of the history of protein engineering, getting a new protein to grab a chosen target meant screening enormous libraries and hoping. AlphaProteo, the generative system Google DeepMind described in September 2024, tried a different bet: give a model a target structure and, optionally, the patch of surface you care about, and let it propose a few dozen to a few hundred complete binders that mostly work on the first try. In the technical report, the best designs against some targets reached sub-nanomolar affinity without any laboratory affinity maturation, and success rates ranged from 9 percent to 88 percent depending on the target.

Two years later the claim needs recalibrating. The headline “3 to 300 times better” is about binding affinity, not hit rate, and one of eight targets failed completely. Meanwhile open tools such as BindCraft, BoltzGen and RFdiffusion3 have narrowed the gap, and AlphaProteo itself remains a closed system with no public download. This post corrects the numbers older explainers get wrong, explains what the published architecture does and does not disclose, and places AlphaProteo in the 2026 field.

What this covers: the binder-design problem, the generator-plus-filter architecture, the per-target wet-lab results, how a design campaign runs, the open competitors, biosecurity, honest limits, and a practical checklist for choosing a tool.

What Changed for October 2026

This is a full rewrite of the April 2026 post. The earlier version contained several figures that cannot be traced to DeepMind’s materials, so the corrections matter more than the additions.

Target list corrected. The seven successful targets are BHRF1, SARS-CoV-2 spike receptor-binding domain (SC2RBD), IL-7 receptor alpha (IL-7RA), PD-L1, TrkA, IL-17A and VEGF-A. The earlier post listed an incorrect set and omitted BHRF1, IL-17A and VEGF-A. TNF-alpha, an eighth target, produced no binders.
Hit-rate claims corrected. The earlier post stated a 62 percent hit rate on PD-L1 and “3 to 300 times” higher hit rates. The technical report gives 15 percent on PD-L1, and the 3 to 300 times figure refers to binding affinity versus the best prior methods.
Invented operational details removed. Per-run compute costs, partner names for AlphaProteo specifically, and a 24-hour pipeline timeline were not sourced and have been deleted.
Competitors updated. BindCraft appeared in Nature in August 2025, Chai-2 reported antibody design in mid-2025, BoltzGen was released as open source in late 2025, and RFdiffusion3 was released in December 2025.
Biosecurity section added. A 2025 Science paper on screening DNA synthesis orders against AI-redesigned toxins changed how access to design tools is being discussed.
Access status. As far as public sources show, AlphaProteo is still not released as open software. This is stated with the caveat that private licensing arrangements would not be visible to outsiders.

Context and Background: The Binder Problem and Why It Is Hard

A protein binder is a protein engineered to attach to a specific surface on another molecule. The surface might be a viral protein you want to neutralize, a signalling receptor you want to block, or a biomarker you want to detect. Antibodies are the natural example, but binders can also be small, hyper-stable scaffolds of 50 to 140 amino acids that are cheap to produce in bacteria and tolerate heat that would destroy an antibody.

Three properties define a good binder. Affinity, usually quoted as the dissociation constant Kd, measures how tightly it holds on: nanomolar is good, picomolar is exceptional. Specificity measures whether it ignores everything else. Developability covers expression yield, solubility and thermal stability, the unglamorous properties that decide whether a molecule can be manufactured.

Before generative models, there were three routes. Rational design used a known interface and tried to rebuild it, which worked rarely. Directed evolution and display technologies screened libraries of millions to billions of variants, which worked but cost months and required an existing starting point. Computational design using physics-based tools such as Rosetta delivered occasional successes, but the literature shows very low experimental success rates on hard targets, often well under one percent per design tested.

The first deep-learning break came from the Baker lab. RFdiffusion, published in 2023, adapted the diffusion-model idea from image generation to protein backbones, and paired with ProteinMPNN for sequence design it made de novo binders practical for many research groups. Our overview of AI protein design with RFdiffusion covers that lineage, and the follow-up on RFdiffusion2 and deep-learning protein design covers the enzyme-oriented successor.

Structure prediction also matters, because every modern binder pipeline asks a folding model whether a design will really adopt its intended shape and dock where intended. That is why AlphaFold 3 appears inside AlphaProteo’s filter. For background on the predictor, read our guides to AlphaFold 3 architecture and diffusion and AlphaFold 3 protein-ligand co-folding.

The primary source for everything in the next sections is DeepMind’s own announcement, AlphaProteo generates novel proteins for biology and health research, together with the accompanying technical report, “De novo design of high-affinity protein binders with AlphaProteo”. Where this article quotes per-target numbers, they come from that report.

The framing question for the rest of this article is what a generative binder tool actually changes. It does not replace the wet lab, and it does not make therapeutics. It changes the economics of the first step, from a library of billions screened blindly to a few dozen to a few hundred designs tested deliberately. That shift is worth understanding precisely rather than in slogans.

Reference Architecture: A Generator, a Filter, and a Wet-Lab Loop

AlphaProteo has two published components: a generative model that proposes binder structures and sequences conditioned on a target and optional hotspot residues, and a filter model that scores designs by predicted chance of experimental success using AlphaFold 3 and AlphaFold 2 confidence metrics. The machine learning details are deliberately withheld, so the architecture is known at the interface level, not the layer level.

Figure 1: The AlphaProteo design pipeline as described in DeepMind’s technical report. The generator and filter internals are not published.

The diagram shows inputs on the left of the flow and measured affinities at the end. The important structural fact is that selection is split in two. The generator is optimized to produce plausible binders, and a separate filter decides which ones deserve the expense of synthesis. That split is the main reason hit rates are high enough to avoid large screening libraries.

What the report says about the generative model

The report states that the generative model was trained on structure and sequence data from the Protein Data Bank (PDB), plus a distillation set of structures predicted by AlphaFold. DeepMind’s blog describes the distillation data as more than 100 million predicted structures from AlphaFold. Distillation here means using a strong predictor’s outputs as additional training examples, which widens coverage far beyond the roughly experimental structures in the PDB.

The model takes a target structure and, optionally, hotspot residues that mark the epitope, and it emits a candidate binder of 50 to 140 amino acids with a structure and a sequence. Crucially, no known binders for the target were used. The report calls the designs zero-shot, meaning the model generates them without target-specific examples or experimental feedback. That is a strong claim, though with a caveat covered later: three targets were used during model development.

What the report does not say is how the generator works internally. The older version of this post described a “graph neural network structure encoder” and a “diffusion-based generative module” as if confirmed. They are not confirmed. DeepMind says the machine-learning methods are omitted for “biosecurity and commercial considerations”. It is reasonable to guess a diffusion-style or other iterative generative process given the field, but that is inference, and you should treat any blog that describes layer counts or loss functions for AlphaProteo as speculation.

What the filter does and why it matters

The filter is arguably the more portable idea. It takes a candidate complex, runs it through structure predictors, and uses their confidence outputs to estimate whether the binder will actually bind. The report describes the in silico success criteria used for benchmarking. For AlphaFold 2 they were interchain predicted aligned error (pAE) below 10, binder-aligned RMSD below 1 angstrom against the predicted monomer, and pLDDT above 80. For the AlphaFold 3 benchmark they tightened to interchain pAE below 1.5, predicted TM-score above 0.8, and complex RMSD below 2.5.

Why does this matter? Because the predictors are cheap relative to a wet-lab experiment. If a structure predictor can say “this binder, placed here, fits and is confidently folded”, it removes most designs before they cost money. The key point is not that the metric is perfect. It is that a metric correlated with experimental success lets you spend your laboratory budget on the right 60 to 170 sequences.

The hotspot is the user’s interface to the model

Hotspot residues are the control surface. The report says hotspots were chosen by hand using known binding-partner interfaces where available, structural properties such as hydrophobic patches and polar pockets, and prior literature. Choosing them badly is a failure mode no model can fix. A hotspot on a region that is disordered, buried, or shielded by glycans will yield designs that look good in silico and bind nothing.

This is also why the target list is not random. BHRF1 is a viral protein with a hydrophobic groove, a well-formed pocket ideal for a designed helix to insert into. The hardest target that succeeded, IL-17A, has a polar pocket on a dimer. The one that failed, TNF-alpha, has a flat and highly polar site between subunits of a homotrimer. Binding is easier where the surface offers shape complementarity and hydrophobic burial, and harder where it offers only scattered polar contacts.

Two model versions

The report distinguishes an early model, v1, used across all seven wet-lab targets, and an improved v2 evaluated in silico. On an AlphaFold 2 based benchmark over five targets, v2 outperformed RFdiffusion on four. On an AlphaFold 3 based benchmark over nine targets, v2 achieved higher in silico success on all nine, while v1 beat RFdiffusion on six of the nine. This is an important caveat: the v2 numbers are predicted success, not experimental results, so they say the model improved against its own filter, which is not the same as improved lab outcomes.

Where this sits in a pipeline

In practice, a campaign using AlphaProteo-class tools has four layers. A target definition layer prepares the structure and chooses epitopes. A generation layer produces thousands of candidates. A scoring layer reduces them with structure predictors and physics-based checks. A wet-lab layer tests tens to low hundreds. Each layer can be swapped, and open pipelines have done exactly that, as the comparison section shows. Our companion piece on protein conformational state prediction explains why an input structure in a single conformation can itself be a limiting assumption.

Deeper Analysis: What the Wet-Lab Data Actually Show

The experimental claims are strong but frequently misquoted. Reading the per-target table is the best defence against the headline.

Per-target results from the technical report

For each target, DeepMind selected designs by automated filtering and tested them. Success rate is the fraction of tested designs that bound the target in yeast surface display. Best Kd is the tightest measured affinity among them.

Target	Role	Designs tested	Success rate	Best Kd
BHRF1	Viral protein, hydrophobic groove	94	88%	8.5 nM
SC2RBD	SARS-CoV-2 spike domain	172	12%	26 nM
IL-7RA	Cytokine receptor	94	25%	0.082 nM
PD-L1	Immune checkpoint, flat epitope	159	15%	0.18 nM
TrkA	Neurotrophin receptor	131	9%	0.96 nM
IL-17A	Cytokine, polar pocket on a dimer	63	14%	8.4 nM
VEGF-A	Angiogenesis factor	94	33%	0.48 nM
TNF-alpha	Homotrimer, flat polar site	54	0%	none

Converting rates to counts is simple arithmetic and gives a feel for scale: roughly 83 binders among 94 BHRF1 designs, about 31 for VEGF-A, about 23 for IL-7RA, about 24 for PD-L1, about 21 for SC2RBD, about 12 for TrkA and about 9 for IL-17A. These counts are my own calculation from the rounded rates, not figures the report states.

Three observations follow. First, the median success rate across the seven targets is about 15 percent by my calculation, and BHRF1 at 88 percent is the outlier that inflates the headline range. A realistic expectation for a new, non-ideal target is therefore one in six or seven designs, not nine in ten. Second, four targets yielded sub-nanomolar binders, with nine sub-nanomolar designs in total. Third, no target needed affinity maturation to reach nanomolar range, which is the genuine advance over display-library workflows.

The “3 to 300 times” claim, decoded

DeepMind’s blog states that AlphaProteo binders showed 3 to 300 times better binding affinities than the best existing methods, and on average bound 10 times more strongly. This is an affinity comparison. Hit rate comparisons appear separately and are less dramatic against RFdiffusion, as the next table shows.

Target	AlphaProteo success	RFdiffusion success	AlphaProteo best Kd	RFdiffusion best Kd
IL-7RA	25% (94 tested)	17% (95 tested)	0.082 nM	14 nM
PD-L1	15% (159 tested)	13% (95 tested)	0.18 nM	1.6 nM
TrkA	9% (131 tested)	0% (95 tested)	0.96 nM	370 nM

The head-to-head on three overlapping targets is the most defensible comparison in the report, because the same assay was applied to both methods. Success rates are modestly higher on IL-7RA and PD-L1 and decisively higher on TrkA, where RFdiffusion found no binders among 95 tested. The affinity gap is large everywhere, from roughly 9 times on PD-L1 to about 385 times on TrkA. So the honest summary is: a modest advantage in hit rate on easy targets, a large advantage in hit rate on hard targets, and a consistent large advantage in affinity.

Figure 2: The validation loop. Designs pass an in silico filter, then binding assays, then structural and functional confirmation.

The figure traces the path of one design. Only candidates passing the filter are expressed. Binding is measured by yeast display and then quantitative assays, and a small subset is taken forward to structures and functional tests. The loop is the real product: the model’s value is that it makes the first pass of this loop succeed often.

Comparison with earlier published methods

The report also compares against literature methods, using each paper’s unoptimized best designs. On BHRF1, an earlier method reported 18 percent success and 58 nM best Kd, against AlphaProteo’s 88 percent and 8.5 nM. On IL-7RA, an earlier reported success rate of 0.15 percent compares to 25 percent. On TrkA, 0.07 percent compares to 9 percent. On IL-17A, 0.02 percent compares to 14 percent. The report summarizes affinity improvements over the best unoptimized prior method as 7 times for BHRF1, 4 times for SC2RBD, 37 times for IL-7RA, 5 times for PD-L1, 380 times for TrkA and 5 times for IL-17A.

A caution applies. Cross-paper comparisons differ in assay, threshold, library format and effort. They are useful for orders of magnitude and unreliable for fine ranking, which is why the direct RFdiffusion experiment above carries more weight.

Do the structures match the designs?

For SC2RBD, four binders were solved by cryo-electron microscopy at resolutions between 4.5 and 6.0 angstroms, with binder backbone deviations from the design between 0.84 and 3.14 angstroms. For VEGF-A, an X-ray structure at 2.65 angstroms showed the binder backbone within 0.78 angstroms of the design and a designed hydrogen bond between a binder histidine and a VEGF tyrosine recapitulated nearly perfectly. Such structure-level agreement is much more convincing than a binding curve, because it shows the model placed atoms, not just stickiness.

Do the binders do anything useful?

The functional tests are modest but real. Against live SARS-CoV-2, binders were tested on ancestral, BA.1, JN.1 and XBB.1.5 variants, and every variant was neutralized by at least one designed binder. Reported neutralization EC50 values were around 89 nM and up, two to ten times weaker than the binding Kd of 26 to 30 nM, a gap the report says is typical of clinical antibodies such as sotrovimab. For VEGF-A, a binder with 0.48 nM Kd reduced phosphorylation of VEGFR2, ERK and AKT in endothelial cells at 1 micromolar, more strongly than equimolar bevacizumab and comparably to a small-molecule inhibitor in that assay.

These are cell-culture results at one concentration. They show that binding translates into pathway blocking. They do not show pharmacokinetics, immunogenicity or efficacy in animals, and the report explicitly states the work is for research use, not clinical use.

Developability: the quiet strength

About 93 percent of designs expressed successfully in E. coli, most were monodisperse by size-exclusion chromatography, and thermal melts showed no unfolding up to 95 degrees Celsius. Small hyper-stable scaffolds are a genuine manufacturing advantage compared with full antibodies. That said, stability in a test tube is not stability in plasma, and a small protein is cleared rapidly by the kidneys unless engineered otherwise.

The 2026 Field: How AlphaProteo Compares with Open Alternatives

When AlphaProteo appeared, its main rival was RFdiffusion. By October 2026 there are at least four serious families, three of them open. The comparison below uses only figures from the respective primary or near-primary sources, and the numbers are not directly comparable because targets, assays and success definitions differ.

Figure 3: The 2026 binder-design landscape, grouped by approach and access model.

The figure groups tools by how they generate designs and who can use them. The practical dividing line for most teams is not accuracy but access: AlphaProteo and Chai-2 are closed, while the other three families can be run on your own hardware.

BindCraft: the hallucination route

BindCraft, from Bruno Correia’s group at EPFL, was published in Nature in August 2025. It backpropagates through AlphaFold 2 weights to hallucinate a binder and its structure simultaneously, then redesigns sequences with ProteinMPNN and filters with AlphaFold 2 monomer metrics and physics-based scoring. Across 12 targets, the paper reports experimental success rates of 10 to 100 percent, averaging 46.3 percent, with predominantly nanomolar affinities and no high-throughput screening or affinity maturation. Targets included PD-1, PD-L1, IFNAR2, CD45, allergens, and CRISPR nucleases.

The average looks higher than AlphaProteo’s median, but the targets and success definitions differ, so do not rank them on those two numbers. The important point is that an open pipeline built on AlphaFold 2 now reports success rates in the same range as a closed generative system, with one-shot design and a modest number of tested designs.

RFdiffusion and RFdiffusion3: the open diffusion line

RFdiffusion remains the reference baseline. RFdiffusion3 was released on 3 December 2025 by the Baker lab and collaborators through the RosettaCommons Foundry. Coverage describes it as an all-atom model that can design against proteins, DNA and small molecules and handle enzyme design. A third-party technical summary reports about 168 million parameters, a roughly tenfold speed improvement over RFdiffusion2, and a 14-atom-per-residue representation. I did not independently verify those technical details against the paper, so treat them as reported. The predecessor is covered in our RFdiffusion2 article.

BoltzGen and Boltz-2: open and unified

BoltzGen, released by the MIT-linked Boltz team in November 2025 as a preprint, is an all-atom generative model that unifies design with structure prediction. The authors report nanomolar-affinity binders for 6 of 9 novel targets, 66 percent, with additional validation on five benchmark targets, and they released training code, inference code, weights and designs under the MIT License. The 9 targets were selected to have no similar proteins in bound contexts above 30 percent sequence identity, which is a harder generalization test than testing on familiar targets. It is still a preprint, and the numbers should be read that way.

Boltz-2, released in June 2025 with Recursion under an MIT license, is a scoring tool more than a generator. It predicts binding affinity, and its authors report FEP-class accuracy in about 18 seconds on one GPU, with a Pearson correlation of 0.62 on a hit-to-lead benchmark against 0.72 for full free-energy perturbation. Used downstream of any generator, it adds a ranking signal. Our broader guide to scientific foundation models places these systems in context.

Chai-2: closed, antibody-focused

Chai-2 from Chai Discovery targets antibodies. Its July 2025 preprint reports testing up to 20 designs per target on 52 antigens with no known antibodies in the PDB, finding binders for about half, with an average hit rate of 15.5 percent across designs, 20.0 percent for single-domain VHH and 13.7 percent for scFv, and a two-week target-to-validated-binder timeline. It overlaps AlphaProteo in spirit, but its output is an antibody format, not a small scaffold, which matters for downstream therapeutic development.

Decision matrix

Need	AlphaProteo	BindCraft	RFdiffusion family	BoltzGen	Chai-2
Can I run it myself today	No public release	Yes, open source	Yes, open	Yes, MIT license	No, closed
Small hyper-stable scaffold	Yes, 50 to 140 residues	Yes	Yes	Yes	Not primary
Antibody format output	No	No	Separate antibody work	Includes nanobody designs	Yes, VHH and scFv
Evidence base	Technical report, 8 targets	Nature, 12 targets	Many papers	Preprint, 9 novel targets	Preprint, 52 targets
Best for	Reading the benchmark	Fast academic projects	Customization	Broad molecule types	Antibody discovery

The matrix is a snapshot of public claims, not a bake-off. No neutral, blinded head-to-head across all five exists as far as I know, and vendor and preprint numbers selectively report favourable targets. A third-party blog comparison I found made sweeping rankings but I could not retrieve it for verification, so I have not used it.

What the open tools changed

Two years ago, a binder design capability with sub-nanomolar outputs looked like an exclusive advantage. Now the gap is closing in capability and has largely closed in access, while the closed systems retain advantages that are harder to see in a table: integrated filtering trained on proprietary data, wet-lab feedback loops, and engineering for reliability. The meaningful competition is shifting from “who can design a binder” to “who has the best filter and the fastest experimental loop”.

Biosecurity: The Reason the Methods Are Withheld

DeepMind’s report omits the machine-learning methods “due to biosecurity and commercial considerations”, and the blog mentions a phased approach to sharing, developed in conversation with the Nuclear Threat Initiative’s AI Bio Forum. That stance contrasts with the open release of competing tools, and the tension is real.

Figure 4: Where controls sit. Whatever the access tier of the design tool, the DNA synthesis order is the common chokepoint.

The diagram makes the key point: a design model, open or gated, produces a sequence, and a sequence has to be turned into DNA before it matters physically. Synthesis screening is therefore the shared control point.

What the 2025 Science study showed

In October 2025, Microsoft researchers led by Eric Horvitz published “Strengthening nucleic acid biosecurity screening against generative protein design tools” in Science. After a two-year red-teaming effort, they reported that AI-redesigned protein sequences could evade existing hazard-detection systems used by synthesis providers, and that patches developed with providers made screening significantly more AI-resilient. I could not retrieve the exact detection percentages, so I leave them out. The study also set up tiered access to the sensitive data through the International Biosecurity and Biosafety Initiative for Science (IBBIS).

What this means for binder tools specifically

Binder design is not toxin design, and the AlphaProteo targets are therapeutic or research proteins. But the underlying capability, generating functional proteins whose sequences differ from natural ones, is general. Sequence-similarity screening depends on resemblance to known hazards, and generative tools can erode resemblance. The defensible position is layered: screening that considers function and structure, not only sequence homology; customer verification at synthesis companies; and access policies proportionate to a tool’s capability.

My opinion, labelled as such: withholding implementation details is a reasonable short-term decision for a capability at the frontier, but it is not a durable control when open tools reach similar performance. The weight of protection will fall on the synthesis chokepoint, so teams using any binder tool should order DNA only from providers that screen, and should document intended use.

Trade-offs, Gotchas, and What Goes Wrong

The headline numbers hide failure modes that every practitioner should plan for.

Targets that are flat and polar fail. TNF-alpha produced zero binders in 54 designs, and the authors attribute it to a flat, highly polar interface between subunits of a homotrimer. They also report that about 80 percent of 200 sampled PDB targets showed higher in silico success than IL-17A, the hardest success, which suggests the method generalizes broadly in silico. In silico success is a prediction, though, and TNF-alpha shows that biology can still defeat it.

Model development overlapped with test targets. The report says SC2RBD, PD-L1 and TrkA were used during model development. Performance on genuinely unseen targets could be lower than those three suggest. The other four targets, notably IL-17A and VEGF-A, give a cleaner picture, and they include the lower end of the range.

Static input structures. All designs used target crystal structures as inputs. The report demonstrates nothing on intrinsically disordered targets or NMR ensembles, and dynamic or cryptic epitopes are a known weakness of the whole field.

Limited specificity testing. Binders were counter-screened against only seven other targets. Proteome-wide off-target binding is untested, and that is the property that decides whether a therapeutic is safe.

Yeast display is not therapeutic efficacy. A “hit” in yeast display means detectable binding. It says nothing about half-life, immunogenicity, aggregation at high concentration or manufacturability at scale. The step from binder to drug remains long and uncertain.

Hotspot errors propagate silently. A wrong epitope yields confident-looking nonsense. Always cross-check hotspot choices against known biology, and design to more than one epitope.

Filter and metrics drift. Success criteria tuned to AlphaFold 2 or 3 confidence can reward designs the predictor likes, not designs that bind. Independent orthogonal scoring, such as a second structure predictor or an affinity model, reduces this risk.

Reproducibility and access. Because AlphaProteo is not distributed, results cannot be independently reproduced. The numbers come from the developer’s report, performed in collaboration with academic labs including the Francis Crick Institute, and have not been replicated by an external group using the software. That is a limit on the evidence, not an accusation.

Practical Recommendations

If you are choosing a binder-design approach in late 2026, start from what you can run and what you need to produce, not from which model has the best headline.

For an academic or small-biotech team with a defined target, use an open pipeline. BindCraft is the lowest-friction option for small scaffold binders, BoltzGen is attractive when the target type is unusual, and RFdiffusion3 with ProteinMPNN is the choice when you want to customize constraints. Run at least two generators on the same target and compare which designs overlap, because agreement between independent methods is a useful confidence signal.

For antibody-format needs, the relevant tools are the antibody-specific ones, and Chai-2 is the most prominent closed example. For anyone treating AlphaProteo as a benchmark, use it as an evidence base: its report is the clearest public statement of what a well-tuned generator-plus-filter pipeline can do.

Plan the experiments as carefully as the computation. Test 50 to 200 designs per epitope rather than a handful, include at least one negative control, and measure affinity with an orthogonal assay before believing any hit.

Checklist before committing lab budget:

Confirm the target structure is high quality and in the relevant conformation.
Choose two or three distinct hotspot patches, not one.
Prefer hydrophobic, concave epitopes for a first campaign; budget for failure on flat polar ones.
Generate with two tools and filter with a third scoring method.
Express in a system you can scale, and test stability early.
Counter-screen against related proteins before declaring specificity.
Order DNA only from providers that screen sequences, and record intended use.
Treat any binder as a research reagent until pharmacokinetics and immunogenicity data exist.

Frequently Asked Questions

Is AlphaProteo open source or publicly available?

No public release exists as far as I can verify. DeepMind published a blog post and a technical report describing results, but the machine-learning methods and code are withheld for biosecurity and commercial reasons. Private collaborations, such as with the Francis Crick Institute for validation, are visible, while other licensing arrangements would not be. If you need a runnable tool, use open alternatives such as BindCraft, BoltzGen or RFdiffusion3, and treat AlphaProteo’s report as a benchmark reference.

What hit rates did AlphaProteo actually achieve?

Experimental success rates ranged from 9 percent on TrkA to 88 percent on BHRF1 across seven targets, with 12 percent on SC2RBD, 14 percent on IL-17A, 15 percent on PD-L1, 25 percent on IL-7RA and 33 percent on VEGF-A. The median is about 15 percent. TNF-alpha, an eighth target, yielded no binders in 54 designs. The often-quoted “3 to 300 times” refers to affinity improvement over prior methods, not hit rate.

How does AlphaProteo compare with RFdiffusion?

In the report’s direct test on three targets, AlphaProteo had higher success rates on IL-7RA (25 versus 17 percent), PD-L1 (15 versus 13 percent) and TrkA (9 versus 0 percent), and much tighter best affinities, for example 0.96 nM versus 370 nM on TrkA. Those experiments were run by the developers, with sample sizes of 94 to 159 designs per target. RFdiffusion3 has since been released, and was not part of that comparison.

Can AI-designed binders become drugs?

Possibly, but binding is only the first requirement. A therapeutic also needs suitable half-life, low immunogenicity, specificity across the proteome, manufacturability and efficacy in animals and people. The AlphaProteo report is explicitly for research use, not clinical use, and I found no confirmed clinical program built on AlphaProteo binders. Isomorphic Labs has said first clinical trials are expected by the end of 2026, but public reporting ties those programs to AlphaFold 3-based drug candidates, and I found no link to AlphaProteo.

What is the difference between a de novo binder and an antibody?

An antibody is a large, roughly 150 kilodalton protein from the immune system, optimized by evolution for binding. A de novo binder is a small, computationally designed protein, in AlphaProteo’s case 50 to 140 residues, that can be produced in bacteria and is often very thermostable. Antibodies have mature manufacturing and clinical track records. De novo binders offer speed and stability but less clinical precedent.

Is protein binder design a biosecurity risk?

The specific targets in AlphaProteo’s report are therapeutic or research proteins, but generative protein tools are dual-use in principle. A 2025 Science study from Microsoft found AI-redesigned sequences could evade existing DNA synthesis screening until patches were applied. The practical controls are screening at synthesis providers, customer verification and proportionate access to powerful tools. Researchers should order DNA only from screened providers.