From Prediction to Design — After AlphaFold, the Science of Drawing Proteins and Building Them

Making 27 Out of 129

In protein design, hallucination is a generative technique that drives a structure-prediction neural network in reverse to produce amino acid sequences not found in nature. A neural network is a multi-layered calculation network that learns patterns on its own from many examples, and an amino acid sequence is the order in which amino acids—the parts that make up a protein—are arranged in a chain. This is a different meaning from the hallucination where a chatbot makes up facts that are not true.

The researchers synthesized genes for 129 amino acid sequences hallucinated by the neural network, put them into E. coli, and extracted the proteins produced by the E. coli. Proteins with faulty designs typically stick to each other and precipitate into clumps. The state of floating separately one by one in solution without clumping is called monodispersity, and just reaching that point passes the first hurdle. Of these, 27 showed a monodisperse state, and their circular dichroism spectra also matched the hallucinated structures. Circular dichroism is a measurement that shines light to find out whether the chains are wound into helices or spread into strands. Among them, the three-dimensional structures of three were directly solved using X-ray crystallography and NMR—techniques that experimentally read atomic positions—and matched the hallucinated model well (PMID 34853475). Making 27 out of 129. This number is the starting point for the field of drawing and making proteins from scratch that did not exist in nature.

Tracing back, the story spans half a century. Amino acid sequences bundle up spontaneously inside cells to form unique three-dimensional shapes, and this bundling work is called folding, with the resulting shape called the structure. And that shape immediately becomes the function. An enzyme catalyzing a reaction and an antibody catching a pathogen are all matters of shape. Therefore, the most famous unsolved problem in biology was "Given the sequence, can you guess the shape?", and in 2020 that problem was effectively answered. At the 14th Critical Assessment of Protein Structure Prediction (CASP14), AlphaFold2 scored a median GDT of 92.4 for 87 out of 92 evaluation domains. GDT (Global Distance Test) is a score that measures how much a predicted structure overlaps with an experimental structure, where a higher score means a better match. The competition organizers declared that this problem, which had plagued academia for over 50 years, was "solved" for single chains, and today about 214 million structures are uploaded to its database (PMID 41047692). That half-century of failures is covered in "Why It Took Fifty Years: Protein Structure Prediction and Its Long Trail of Failures", and the final leap in "From Sequence to Shape: How AlphaFold Cracked a 50-Year Problem".

Then the next question followed. If we can read, can we also write? De novo protein design is the work of deciding on a desired function first, drawing the shape that will perform that function, and finding a sequence that folds into that shape. De novo means "from the beginning," meaning it starts from a blank sheet of paper instead of modifying proteins found in nature. In proteins, information flows in the order of sequence → structure → function, and design reverses this flow. This "reverse central dogma" has been the backbone of almost all de novo design attempts (PMID 38361073).

The answer is both "Yes" and "It was not that simple." In particular, there is one prickly question. By what standard do we judge that a design has succeeded? If we grade the designed product with a predictor, won't the design and the grading share the same bias? Much of what this article covers is not a settled conclusion, but a front line in progress.

Myth: If You Can Read a Structure, Making One Is Right Around the Corner

The myth went like this. The expectation was that if a neural network learned the rules of folding so accurately, running those rules backward to spit out new proteins was just a matter of time. However, two things held this expectation back.

First, predicted structures are not verified facts. A predicted structure is a model unconfirmed by experiment, so it cannot be treated as the right answer. Here, ground truth is the measured value that serves as the benchmark to judge whether something is correct or incorrect. How proteins move as they change shape, complexes formed by multiple chains, functional interpretation, and model quality assessment remain ongoing challenges (PMID 37850517). That a prediction fits well is different from saying the predictor makes correct judgments about proteins that have never existed in the world.

Second, most proteins actually made came from physics, not neural networks. Physics-based methods are calculations that write out the forces between atoms as equations and find the shape where those equations yield the best score, with Rosetta being the representative program. The opening sentence of the ProteinMPNN paper summarizes this situation. Even though deep learning revolutionized structure prediction, nearly all experimentally characterized de novo designs were made by physics-based methods like Rosetta (PMID 36108050). Experimentally characterized means they were actually made and their properties were measured. And the success rate of that physics-based design was low. The target protein you are trying to grab onto is called the target, and a protein made to stick to the target is a binder (binding protein). The value measuring how strongly the two stick together is affinity, written as a concentration, and the smaller the number, the stronger the grip. Designing binding proteins that stick strongly using only target structure information had already become possible, but a 2023 self-assessment was that "the overall design success rate is low, leaving substantial room for improvement" (PMID 37149653).

An example showing the peak of this era is the 2020 SARS-CoV-2 minibinder. A minibinder is a very small binding protein. This virus grabs onto and enters human ACE2 receptors using the receptor-binding domain (RBD), which is a piece of its spike protein. The researchers designed two routes: a backbone mimicking the helix of ACE2 that contacts the RBD, and a backbone that docked to the RBD to find a new binding mode, and 10 of them bound to the RBD. The affinities were 100 picomolar to 10 nanomolar, and the concentration inhibiting cell infection by half (IC50) was 24 picomolar to 35 nanomolar. The cryo-EM structure was nearly identical to the computational model (PMID 32907861). Cryo-EM is an instrument that freezes proteins and shines electrons on them to figure out their shapes. It is a remarkable achievement. Yet this achievement presupposed large-scale computational backbone generation, experimental screening, and the tinkering of structural biologists. It worked, but it was not something anyone could do for any target.

The Plot Twist: Running Predictors in Reverse Opened Up Design

The turning point started in an unexpected place. Instead of modifying the predictor into a designer, they used it completely in reverse.

The 2021 hallucination approach works like this. You create a completely random amino acid sequence and put it into a structure prediction network. The prediction result is a blurry distance map with no distinct features. A distance map is a table listing how far apart two positions are for every pair in a folded protein. Next comes the core step. The researchers changed the sequence little by little to maximize how distinctly the predicted inter-position distance distribution deviated from the "average for all proteins." A featureless map means there is no shape, so moving away from the average was chosen as the good direction. Put simply, they pushed the neural network to "spit out the sequence that looks most like a protein to you." Starting from different random starting points poured out various sequences and structures, and the result is the 129 and 27 mentioned earlier (PMID 34853475).

Hallucination soon evolved toward adding conditions. A few positions where a protein actually does its work are functional sites, and the rest of the chain is a scaffold supporting those sites. Pharmaceutical hallucination optimizes a sequence so that the predicted structure includes the desired functional site, and inpainting starts from a functional site and fills in the remaining sequence and structure by passing them through the neural network once. Using these two methods, immunogen candidates, receptor traps, metalloproteins, enzymes, and binding proteins were designed (PMID 35862514). It was also applied to symmetrical assemblies. A mass where multiple identical chains are attached around like a ring is a symmetrical homooligomer, and a single individual chain is a protomer. As soon as just the number and length of protomers were specified, symmetrical homooligomers poured out, and the crystal structures of 7 designs almost overlapped with the computational models at a median RMSD of 0.6 angstroms. RMSD is the average distance between the same atoms when two structures are overlaid, and an angstrom is a unit of length about the size of a single atom. Up to 1,550 residues (a residue is a single amino acid strung into a chain), three 10-nanometer giant rings of C33 symmetry were confirmed by cryo-EM, and they were significantly different from any previously solved structures (PMID 36108048).

A symbolic case where hallucination reached even function is the artificial luciferase. Luciferase is an enzyme that helps reactions emit light like fireflies. Family-wide hallucination is a method that mass-produces idealized structures with various pocket shapes, as if imagining an entire family at once. The researchers designed the active site in the structures made that way to catalyze the oxidative chemiluminescence of a synthetic substrate. A substrate is the counterpart molecule that an enzyme catches and changes, and the active site is the pocket where that work happens. The one with the best activity was a small enzyme of 13.9 kilodaltons, with a melting point exceeding 95°C and a catalytic efficiency ($k_{cat}/K_m$) of $10^6\text{ M}^{-1}\text{s}^{-1}$—a level similar to natural luciferases—while having much higher substrate specificity (PMID 36813896). Catalytic efficiency is a value measuring how quickly an enzyme processes a substrate, and the larger it is, the better.

Mechanism: Models that Draw Shapes, Models that Find Sequences, and Models that Ask Back

Hallucination was powerful, but it was slow and hard to control. The real leap came from splitting one hard problem into two easier problems.

The first piece: drawing shapes. The idea of diffusion models is simple. You train a network to gradually add noise to a normal object until it becomes complete noise, and then make it run that process in reverse. You can then start from pure noise and strip away the noise step by step to create a new object. Image generation models work this way. In 2023, RFdiffusion fine-tuned the RoseTTAFold structure prediction network precisely on this task of "stripping away noise mixed into a structure," turning it into a generative model for protein backbones. Fine-tuning is the process of teaching an already trained neural network a new task to change its use, and the backbone is the skeleton of the amino acid chain, the line-like main chain left after removing the side chains. RFdiffusion accomplished unconditional generation, monomer design specifying the overall shape where the chain winds (topology), binding protein design, symmetric oligomer design, and enzyme active site scaffolding all within a single framework. A monomer is a protein made of a single chain, and scaffolding is the work of building a support to hold functional parts in place. The researchers experimentally characterized hundreds of symmetric assemblies, metal-binding proteins, and binding proteins (PMID 37433327).

The second piece: finding sequences that fold into those shapes. Having just a shape is useless; you need the amino acid sequence that will actually produce that shape. The problem of deciding on a shape first and finding the sequence that will fold into that shape in reverse is inverse folding, and ProteinMPNN took over this job. When measuring how well it restores the original sequence on natural backbones, ProteinMPNN scored 52.4%, significantly leading Rosetta's 32.9%. What is more impressive is what comes next. The researchers revived designs made with Rosetta or AlphaFold that had previously failed using X-ray, cryo-EM, and functional assays—across monomers, cyclic homooligomers, tetrahedral nanoparticles, and target-binding proteins alike (PMID 36108050). This means the cause of failure for the revived designs was not that the shape was wrong, but that the sequence failed to support that shape.

The third piece: asking back. After building the backbone and placing the sequence, researchers feed that sequence back into a structure prediction tool to ask, "Will it really fold into this shape?" When AlphaFold2 or RoseTTAFold were used to evaluate and filter the designed sequences for the probability of adopting the designed monomer structure and binding to the target as intended, the success rate in designing binding proteins increased nearly 10-fold (PMID 37149653). The ability to select, rather than the ability to make, is what raised the success rate. This asking back is the most underestimated aspect in this field and the seed of the debates to come.

The assembly of these three pieces is today's standard pipeline. You make shapes with a diffusion model, place sequences with the ProteinMPNN family, and verify them with a structure prediction tool (PMID 41562285). And this standardization itself changed the field. As open-source RFdiffusion and ProteinMPNN combined with structure prediction tools, biochemists and molecular biologists became able to explore applications broadly (PMID 42056544). The methodology also kept pushing forward. RFdiffusion2 removed the constraint of having to pre-specify the sequence positions and backbone coordinates of catalytic residues, designing enzymes using only functional group geometry. A functional group is a bundle of a few atoms that actually drives a reaction. In a benchmark consisting of 41 diverse active sites, RFdiffusion2 generated scaffolds for all 41, compared to previous methods handling 16 (PMID 41339749). A benchmark is a comparative test that makes multiple methods compete on the same problem.

Scorecard: Things That Never Existed in Nature Have Actually Been Built

Before getting into the debate, we need to look at what has actually been verified at the workbench to be fair. The values cited below come from measurements independent of the models—such as crystal structures, cryo-EM, patch-clamp, and catalytic constants. A crystal structure is a structure obtained by solidifying a protein into a crystal and reading the positions of its atoms with X-rays, and a patch-clamp is an experiment that directly measures the current flowing through a tiny pore in a cell membrane.

Enzymes. A serine hydrolase—a multi-step reactive enzyme with a catalytic triad and an oxyanion hole—was designed using minimal active-site descriptions, achieving a catalytic efficiency of up to 2.2×10⁵ M⁻¹s⁻¹, and its crystal structure matched the designed model with a Cα RMSD of under 1 angstrom. New catalysts emerged from five folds different from natural serine hydrolases (PMID 39946508). Here, a fold is the overall pattern of how the chain wraps.

Antibodies. The narrow spot on a target where an antibody actually sticks is called an epitope. When a neural network fine-tuned with RFdiffusion2 was combined with screening that displays candidates on yeast surfaces to pick out only the ones that stick, it became possible to fully compute and design antibody variable heavy chains (VHHs), single-chain variable fragments (scFvs), and even full antibodies that bind to user-specified epitopes. Cryo-EM confirmed the binding posture of the designs, and the high-resolution structure showed that the shapes of all six CDR loops matched the design. A CDR loop is the loop at the end of an antibody that grabs the target. However, the affinity of the initial computational designs was unremarkable, with a Kd of tens to hundreds of nanomolars, and separate affinity maturation was required to push it down to single-digit nanomolars (PMID 41193805). Affinity maturation is the extra work of slightly changing the sequence to pick out the ones that stick harder. A subsequent approach obtained functional antibodies with nanomolar affinity across all four targets after testing only 43 to 101 per antigen, and the designs expressed well in mammalian cells while being novel in sequence and structure (PMID 42337361).

Devices That Do Not Exist in Nature. An ion channel is a protein that pierces through a cell membrane like a hole to let only specific ions pass through, and the selectivity filter is the narrow neck inside that hole that screens which ions to let through. A Ca²⁺ channel was designed from scratch by specifying only the geometry of the selectivity filter residues, and both tetrameric and hexameric assemblies showed higher conductivity for Ca²⁺ than for Na⁺ or other divalent ions in patch-clamps. Tetramers and hexamers are configurations where four and six of the same chains come together, respectively. Mutating the filter residues made that selectivity disappear. The cryo-EM structure of the hexameric channel was nearly identical to the designed model (PMID 41125887).

Regarding this trend, a 2026 review summarizes it as follows: while there is still room for improvement in success rates and activity, the long-standing challenge of designing new structures, assemblies, and binding proteins is "on the verge of being solved," and the core question in this field is now no longer "how to design" but "what to design" (PMID 42056544).

But How Do We Judge Success — How Reliable Is Double-Checking?

The scorecard above is real. However, there is a serious pitfall in how that scorecard is read. The final step of the design pipeline is to double-check using a predictor, but how reliable that double-checking is remains a question that needs to be tested separately.

That test exists. The researchers built a benchmark from 614 de novo-designed monomers experimentally characterized across 11 studies from 2012 to 2021. Success was clearly defined—expressed, soluble in water, monomeric, and folding into the correct local shape where the chain winds into helices or flattens into strands here and there, in other words, the correct secondary structure. The evaluated items are all designs from the pre-RFdiffusion era, and they are monomers rather than binders. Then they measured how well AlphaFold, ProteinMPNN, and ESMFold distinguished success from failure.

Here are the results. All predictive models showed only "moderate" discriminatory power. The more painful part comes next. Many of the failed designs had better confidence scores than the successful ones, and the confidence scores varied depending on the protein topology. A confidence score is a score that the predictor assigns itself, saying, "My prediction is reliable to this extent." The conclusion is restrained but clear—these models are useful as a primary filter before experiments, but without task-specific training, their ability to accurately predict experimental success is limited (PMID 41556605).

The case that illustrates this problem most vividly is a study on miniproteins targeting the SARS-CoV-2 nucleocapsid protein. 538 candidates were designed with RFdiffusion, and 19 of them were selected and produced in E. coli. All of them were soluble and purified at 8 to 10 mg/L, confirmed to be properly folded by circular dichroism, and returned to their original shape when heated to unfold and then cooled down. Nine of them had a melting temperature exceeding 75°C. Up to this point, it is stellar. Yet 19 out of 19 unexpectedly formed oligomers, and binding was confirmed in only 3 of them—with affinities of 115 nanomolar, 4.8 micromolar, and 7.32 micromolar, respectively. An oligomer is a clump formed by multiple copies of the same protein sticking to each other, and if a design that is supposed to float alone sticks to its own kind, its ability to stick to the target can be crucially compromised. The authors point to this unexpected oligomerization as a major barrier to discovering miniprotein binders. The 1.8-angstrom crystal structure of one of those binders matched the design precisely. So the structure was right, but the function did not work out. And a decisive single line. Predictive metrics like AlphaFold3's ipTM and computational simulations did not match the experimental data (PMID 41665261).

A 2026 study that bibliometrically reviewed two decades of literature diagnoses the limits of this field not in the analysis results, but in its closing remarks. Models trained on stable, crystallizable structures struggle to generalize outside the training distribution, static predictions are blind to the structural dynamics governing allostery and catalysis, and in silico confidence scores have repeatedly proved poor as predictors of wet-lab results (PMID 42470596). Outside the training distribution refers to targets whose properties differ from those used in training. In silico means doing it solely inside a computer, and wet-lab experiments are experiments actually done at the bench handling reagents. However, this diagnosis is the authors' synthesized judgment rather than a value measured by the bibliometric analysis.

This loop does not mean that "design achievements are fake." The achievements on the scorecard above are ones confirmed by experiments unrelated to the models. The problem lies elsewhere. As long as the step of choosing what to build before experiments relies on predictors, we are systematically drawn to the kinds of designs that predictors like. What is confirmed here, however, is not proof that "the generator and the scorer share the same bias." What was observed is the fact that the discriminatory power of confidence scores is moderate and fluctuates depending on protein topology, and in the miniprotein case, the authors' diagnosis is more specific—RFdiffusion reliably predicted structural integrity and stability, and the axis it missed was oligomerization (PMID 41665261). And if you cross-validate with multiple computational techniques judged by different principles, there is room to catch flaws before experiments. Indeed, when 518 monomers and 2,112 botulinum toxin binders were analyzed together with multiple predictors and fragment quality metrics, a machine learning model was created that could quite accurately predict well-folding monomers and binders using just a few items summarizing properties numerically (PMID 36275629). This means that even if a universal confidence score falls short, judgments can be made if it is trained separately tailored to the task.

This does not mean that predictor double-checking is useless. It was clearly effective for culling candidates at the population level (PMID 37149653). What is lacking is the accuracy for ranking the success or failure of individual designs.

The Bottleneck Is Still Wet Lab Experiments

The narrative that "AI builds proteins" focuses on the making side. However, the center of gravity of the actual work lies elsewhere.

Generating binding proteins can have low in-silico success rates, so getting enough candidates to pass on to experiments often takes thousands of designs and hundreds of GPU hours (PMID 41562285). Let us look at the mini-protein case from earlier again — 538 were designed, 19 were produced, and 3 bound. Computation churned out 538, but the verdict came from the lab bench.

There is also broader validation of the expectation that AI predictions can replace experiments. The entire network of how proteins inside a cell bind to and detach from one another is called the interactome. A study that systematically compares AI-predicted yeast and human interactomes with experimentally mapped ones concludes this. The quality of high-confidence predictions is on par with established experimental methods. However, in screening across a whole organism's set of proteins (the proteome), the tested AI approaches fell short of experiment-based reference maps in discovering entirely new interactions — the yeast interactome map produced by this study found over 40 times more new interactions than its AI counterpart. The authors' judgment is this — at the current stage, the main contribution of AI predictions is attaching quaternary structure models, which depict how multiple chains assemble and interlock, to experimentally discovered interactions (PMID 41935050).

The Word "Success Rate" and Preclinical Studies Are Not Therapeutics

The most common misreading when interpreting numbers in the design field is missing the fact that the denominator of the success rate varies from paper to paper. The denominator is the latter number in "out of how many," meaning what was chosen as the population. The same word "success" refers to entirely different things. For early hallucination research, 27/129 is the ratio that was monodisperse after expression and purification and whose CD spectrum matched the model (PMID 34853475). For miniprotein research, 3/19 is the ratio for which binding was confirmed out of those picked and produced, and the original design count was 538 (PMID 41665261). For antibody research, "43 to 101 experiments" is the number of designs submitted to experiments, not a success rate (PMID 42337361). For RFdiffusion2, 41/41 is the passing rate for the in silico scaffolding benchmark (PMID 41339749).

There are also values that give a sense of the overall level. A 2023 review summarized that the experimental success rate of the latest design protocols at the time was approaching 20% (arXiv:2310.09685, preprint). A preprint is a manuscript made public first without undergoing review (peer review) by researchers in the same field.

However, we must honestly distinguish between types of evidence here. The most notable success rate figures recently — such as values where testing only 30 to 100 per target yields a macrocyclic peptide hit rate exceeding 90% (arXiv:2507.19375, preprint) — are largely technical reports produced directly by the model development entities themselves and are preprints that have not undergone peer review. Each hit rate was calculated based on a different set of targets, a different experimental format, and a different definition of a "hit," and has not been independently reproduced. For example, depending on whether a "hit rate of 90%" is the ratio of designs where binding was detected or the ratio that fell within the desired affinity range, the weight of the same number changes completely. This does not mean these numbers are wrong, but it does mean they should not be treated as evidence of the same tier as peer-reviewed crystal structures and catalytic constants.

One more thing. Papers reporting only in silico metrics are also mixed into the literature. One design study reports high prediction confidence and favorable interfacial energy, yet explicitly states in the conclusion, "The designs reported here are computational predictions, and biochemical and cellular experimental validation is required to determine functional activity" (PMID 42222300). Interfacial energy is a value calculated and scored to show how well the surfaces where two proteins touch each other fit together. Such self-limitation is exemplary, but if a summary or news report strips away this sentence, it is easily read as "AI pierced an impenetrable target."

And there is one more dangerous leap when covering this field. Most of the achievements of designed binding proteins are characterization at the preclinical stage. Preclinical is the stage tested only in cells and animals, before trying it in humans. IL-23 and IL-17 are well-validated therapeutic targets in autoinflammatory diseases. To block this pathway, minibinders targeting the IL-23 receptor (IL-23R) and IL-17 showed antibody-level low picomolar affinity, were extremely stable enabling oral administration, and orally administered IL-23R minibinders showed better efficacy than clinical anti-IL-23 antibodies in a mouse colitis model and favorable pharmacokinetics in rats (PMID 38936360). Pharmacokinetics refers to how a administered drug spreads through and is eliminated from the body. This is proof of concept in animal models. It is not data speaking to human efficacy and safety, nor is it a story about approved therapeutics. Reviews summarizing this field maintain the same temperature — that de novo binding protein design has drastically reduced development time and resources and raised hit rates, and that characterization in preclinical models is paving the way toward therapeutic development (PMID 40897171). The expression "is paving" accurately hits the current location.

The remaining challenges are also clear. A 2026 review cited earlier summarizes that structures, assemblies, and binding proteins are on the verge of being solved, but what remains is the catalytic design of reactions with high energy barriers, and the design of switches and nanomachines that integrate binding, conformational change, and catalysis into one (PMID 42056544). The axis of dynamics is also major. Nature's signaling proteins work by changing shape, and an approach to designing dynamic transitions between intra-domain (a unit that folds independently inside a protein) geometries at the atomic level has only recently emerged, validating a form designed with four structures (PMID 40403060). The criticism that current models optimized for static structures fundamentally fail to represent conformational ensembles, post-translational modifications, and intracellular context remains valid (PMID 42470596).

Significance: The Question Shifts from "How to Make It" to "What to Make"

First, the common belief that "once prediction is solved, design is automatically solved" was both wrong and half right. It was wrong because simply using a predictor as-is does not result in design. What actually worked was a three-part division of labor: retrofitting the prediction network into a generative model by retraining it on a noise-removal task, separately solving a distinct problem called inverse folding, and filtering the results back through the predictor. It was half right because two of those three pieces were ultimately byproducts of structure prediction research. Neural networks that learned folding could indeed be run in reverse.

Second, the next bottleneck in this field is not making things better, but judging them better. Generative capability has improved rapidly. However, deciding what to make before doing experiments still relies on the reliability scores of predictors, and in retrospective benchmarks targeting monomer design, those scores sometimes gave better values to failed designs than to successful ones (PMID 41556605). Only 3 out of 19 tested designs bound, and the original design pool was 538. A report that prediction metrics mismatched experiments in that case (PMID 41665261) reveals the magnitude of this gap. Until the judgment problem is solved, design will remain tethered to the throughput of wet-lab experiments.

Third, the point where the game has changed is nevertheless clear. For a long time, protein engineering relied on random selection—a method of making and filtering massive numbers of variants—but the paradigm is now shifting toward rational computational design (PMID 42056544). And rather than replacing evolution, design is also being used to move the starting point of evolution. There is a study where three botulinum neurotoxin protease types were redesigned using ProteinMPNN to increase stability, and then the redesigned versions and the wild type were placed side-by-side to run phage-assisted continuous evolution. A protease is an enzyme that cuts proteins, and the wild type is the original form as it exists in nature. Phage-assisted continuous evolution is a technique that runs evolution nonstop in the laboratory by borrowing the rapid reproduction of viruses. When started from three different redesigned enzymes, proteases with consistently higher activity than the wild type under the same selection conditions emerged. Across four evolution campaigns, the redesigns opened up sequence regions that the wild type could not reach, and the mutations obtained through redesign did not work on the wild-type background. When evolved to cut therapeutically relevant targets, the protease evolved from the redesigned starting point reached a selective specificity more than 79 times that of the best variant evolved from the wild type (PMID 42486980).

This final picture neatly summarizes the present state of the field. A shape drawn by humans, a sequence found by a neural network, and laboratory evolution running over it once more. Design has not defeated evolution; design has moved the starting point of evolution to a place nature has never reached.

What opened up after the answer to a 50-year-old problem was an even bigger problem. Only the nature of that problem has changed. From an era of asking how to make it, to an era of asking what to make and how to know whether it really worked. The first question had a scorecard. There was a competition called CASP, and there were answer keys in the form of structures already solved by experiment. The second question has no such answer keys. Whether a protein that does not exist in nature works as intended is something no one knows until they actually make that protein. That is why this field has not yet left the wet lab bench, and why design remains controversial even after prediction has been solved.

Source Papers

The nature of each paper is noted alongside. Figures already present in the main text are not repeated.

Broad Overview

• "The past, present and future of de novo protein design." (2026) — PMID 42056544 — The overview that served as the backbone for this article. The paradigm shift from random selection to intentional computational design, and the list of remaining challenges.
• "From CASP13 to the Nobel Prize: DeepMind's AlphaFold Journey." (2026) — PMID 41047692 — CASP14 performance and the declaration of "solved," along with the scale of the AlphaFold DB.
• "Sparks of function by de novo protein design." (2024) — PMID 38361073 — An overview formulating design in the reverse direction of sequence-to-structure-to-function.
• "Challenges in bridging the gap between protein structure prediction and functional interpretation." (2025) — PMID 37850517 — A warning that predicted structures are unvalidated models and cannot be treated as definitive answers.

Physics-Based Era

• "De novo design of picomolar SARS-CoV-2 miniprotein inhibitors." (2020) — PMID 32907861 — The pinnacle of physics-based design. An achievement predicated on large-scale backbone generation and experimental screening.
• "Robust deep learning-based protein sequence design using ProteinMPNN." (2022) — PMID 36108050 — A diagnosis that nearly all experimentally characterized designs were physics-based, along with inverse-folding sequence recovery and rescuing failed designs.
• "Improving de novo protein binder design with deep learning." (2023) — PMID 37149653 — Self-diagnosis regarding low success rates, and the results of raising success rates by cross-examining predictors.

Hallucination Lineage

• "De novo protein design by deep network hallucination." (2021) — PMID 34853475 — A method starting from random sequences and pushing them toward the "most protein-like" direction. The first numbers in this article come from here.
• "Scaffolding protein functional sites using deep learning." (2022) — PMID 35862514 — Scaffolding functional sites using constraint hallucination and inpainting.
• "Hallucinating symmetric protein assemblies." (2022) — PMID 36108048 — Generating symmetric assemblies by specifying only the number and length of protomers, confirmed via crystal structures and cryo-EM.
• "De novo design of luciferases using deep learning." (2023) — PMID 36813896 — An instance where hallucination reached function. Artificial luciferases.

Standard Pipeline

• "De novo design of protein structure and function with RFdiffusion." (2023) — PMID 37433327 — Transforming a structure prediction network into a backbone generative model by fine-tuning it on a denoising task.
• "Atom-level enzyme active site scaffolding using RFdiffusion2." (2026) — PMID 41339749 — Designing enzymes using only functional group geometry without specifying residue positions. In silico benchmark pass rates.
• "ProteinDJ: A high-performance and modular protein design pipeline." (2026) — PMID 41562285 — The recognition that computational volume is a bottleneck because the in silico success rate of binder generation is low.

Report Card — Measurements Independent of Models

• "Computational design of serine hydrolases." (2025) — PMID 39946508 — Multi-step reaction enzyme design confirmed by catalytic constants and crystal structures.
• "Atomically accurate de novo design of antibodies with RFdiffusion." (2026) — PMID 41193805 — Confirming even CDR loop conformations with cryo-EM. Also reporting the affinity limitations of initial designs.
• "Efficient generation of epitope-targeted antibodies with Germinal." (2026) — PMID 42337361 — Functional antibodies across all four targets after testing only a small number per antigen.
• "Bottom-up design of Ca2+ channels from defined selectivity filter geometry." (2025) — PMID 41125887 — Ion selectivity confirmed by patch-clamp, abolished by filter mutation.

The Adjudication Problem

• "Evaluating zero-shot prediction of monomeric protein design success by AlphaFold, ESMFold, and ProteinMPNN." (2026) — PMID 41556605 — Experimentally characterized monomer benchmarks. Confidence metrics showed intermediate discriminative power, and failed designs sometimes received better values.
• "Experimental analyses of RFdiffusion designed Miniproteins for binding to SARS-CoV-2 nucleocapsid protein." (2025) — PMID 41665261 — A case where the structure was correct but function failed. Prediction metrics conflicted with experiments, and the missed axis was oligomerization.
• "Identifying well-folded de novo proteins in the new era of accurate structure prediction." (2022) — PMID 36275629 — A task-specific adjudication model utilizing multiple predictors together. An alternative to general-purpose scores.
• "Artificial Intelligence and Protein Design: A retrospective study on 20-year emerging trends." (2026) — PMID 42470596 — Bibliometric analysis. The cited limitation diagnosis is the authors' comprehensive judgment rather than an analytical result.
• "Experimental assessment of AI-based interactome mapping." (2026) — PMID 41935050 — High-confidence predictions were on par with experiments, but fell far short of experimental maps in discovering novel interactions.

Places Where Evidence Grades Diverge

• "Generative artificial intelligence for de novo protein design" (2023) — arXiv:2310.09685 (preprint) — An overview gauging the overall experimental success rate of design protocols at the time.
• "Latent-X: An Atom-level Frontier Model for De Novo Protein Binder Design" (2025) — arXiv:2507.19375 (preprint) — A technical report by the model development team. Cited as a representative example of high hit-rate figures.
• "De Novo Protein Design Enables Targeting of Intractable Oncogenic Protein-Protein Interfaces." (2026) — PMID 42222300 — A case that looked only at in silico metrics while explicitly stating the need for experimental validation in the conclusion.

Preclinical and Remaining Challenges

• "Preclinical proof of principle for orally delivered Th17 antagonist miniproteins." (2024) — PMID 38936360 — Proof of principle in mice and rats. Not human data.
• "Code to complex: AI-driven de novo binder design." (2025) — PMID 40897171 — An overview with the temperature that preclinical characterization is "paving the way."
• "Deep learning-guided design of dynamic proteins." (2025) — PMID 40403060 — The starting point of designing conformational transitions at the atomic level.
• "AI-redesigned starting points and outcomes enhance protein evolution." (2026) — PMID 42486980 — A case where redesign shifted the starting point of laboratory evolution.