Paper: Language Modeling Materializes a World Model of Protein Biology "If protein language models keep scaling, what actually happens?"

The AI world has been taught the scaling lesson many times over: more parameters, more data, more compute, and the loss falls along some predictable curve. Then, past a certain scale, abilities that seemed like they should not exist yet suddenly become usable.

So is there an equivalent law inside protein language models? Starting from predicting amino-acid tokens, does a model eventually emerge with an internal understanding of protein structure, function, interactions and even the design space? That is the question this new paper from Biohub and EvolutionaryScale most wants to answer.

The title is ambitious: Language Modeling Materializes a World Model of Protein Biology. It is a claim that deserves suspicion, but reading through, the authors are not trying to show that a model gained a few points. They are trying to connect a whole system: from the protein language model ESMC to the structure predictor ESMFold2, from experimental validation of therapeutic binder design to sparse autoencoders that dissect the model's internal features, and finally out to an ESM Atlas covering 6.8 billion protein sequences and 1.1 billion predicted structures.

In other words, the paper is not really about a single model. It is about a chain of capabilities, and about whether the scaling law of protein language models is pushing them through some emergence point, from sequence statistics learners into a unified stage of structure prediction, complex modelling, binder design and world-scale mapping.

So let us go from ESMFold1 to ESMFold2, and put it on the same page as AlphaFold3, Boltz, Chai, RFdiffusion and BindCraft to see what is actually distinctive.

01 — From ESMFold1 to ESMFold2: the language-model route catches up on complexes

ESMFold1's selling point was clear: no MSA, just the ESM2 protein language model representation, folding a structure quickly from a single sequence.

That route differs fundamentally from the AlphaFold family. AlphaFold2's strength comes largely from the evolutionary covariation in MSAs; ESMFold1 compresses those evolutionary statistics into language model parameters ahead of time, so inference needs no homology search, only folding on top of ESM2 representations.

The advantages are speed, simplicity and better behaviour on orphan proteins and designed sequences. The weakness was equally clear: ESMFold1 was good at single-chain prediction and was not the main tool for complex biomolecular interactions, especially antibody–antigen, protein–protein interfaces and protein–ligand tasks.

ESMFold2 is positioned as systematic homework on exactly that weakness.

It no longer only predicts monomer structure from a language model representation; it wires ESMC representations into an all-atom folding architecture aimed at complex interactions. Three changes matter most.

First, the backbone moves from ESM2 to ESMC, with 300M, 600M and 6B parameter versions and training data expanded from roughly 50 million sequences in the ESM2 era to about 2.8 billion, including large amounts of metagenomic sequence. It has read not only the familiar faces in classic databases but something closer to the real distribution of proteins in nature.

Second, ESMFold2 explicitly builds and iteratively updates a pair representation. It aggregates the layer representations of a frozen ESMC-6B, projects them into two-dimensional pair space, and updates them repeatedly with recurrent folding layers. Pair representations matter because a structure is not a property of individual residues; it is the organization of residue–residue, chain–chain and atom–atom relationships.

Third, ESMFold2 generates atomic coordinates with a diffusion transformer, which brings it closer to the current all-atom structure prediction paradigm and lets it handle more complex molecular interactions.

Figure 1 is really the logical starting point of the paper. As pretraining compute increases, the structural and functional information in ESMC representations keeps improving, following a near log-linear scaling trend.

That sentence matters. It means scale does not just make the model better at guessing amino acids; it makes the model better at capturing folding, function and evolutionary constraints. A falling MLM loss starts to cash out as a stronger biological representation.

Reading the ESM line as one evolutionary ladder makes this clearer. In the ESM1/ESM2 stage, contact, fold and function information begins to appear in the representations. With ESMFold1, those representations are strong enough to support single-chain structure prediction. With ESMC and ESMFold2, the authors try to show they are strong enough for complex prediction, antibody–antigen folding and binder search.

I prefer to read ESMFold2 as a downstream cashing-in of the protein language model scaling law. It is not a structure prediction miracle trained from scratch; it is a set of new capabilities released by plugging a stronger ESMC representation into pair representations, recurrent folding layers and a diffusion head.

That is also the generational difference from ESMFold1. ESMFold1 proved that language model representations can fold proteins; ESMFold2 tries to prove that they can support complex interaction prediction and enter a design search loop. In a sense, that is a waypoint in moving protein foundation models from representation emergence toward capability emergence.

02 — Against AlphaFold3, Boltz and Chai: what is different about ESMFold2?

On tasks alone, ESMFold2 sits easily inside the competition between all-atom complex predictors such as AlphaFold3, Boltz-1/2, Chai-1 and Protenix. They all answer the same question: given biomolecular sequences and context, predict the complex structure.

But ESMFold2 takes a different route.

AlphaFold3, Boltz and Chai are closer to an explicit structural modelling route: they are organized around MSAs, templates, pair representations and diffusion modules, learning a mapping from inputs to complex coordinates directly. Their strength is a complete structural modelling system, especially suited to multi-molecular complexes, ligands and nucleic acids.

ESMFold2 is closer to a language-model-representation-first route: learn a general protein representation from massive sequence data with ESMC, then attach that representation to a structure prediction trunk and a diffusion head. It gives up the earlier "no MSA" posture and the paper offers an MSA-conditioned mode, but the selling point is that even in a single-sequence setting it reaches strong complex prediction performance.

The numbers are direct. With single-sequence input, ESMFold2 reaches a 50% ± 2% DockQ pass rate on FoldBench antibody–antigen, 70% ± 1% on protein–protein interfaces and 57% ± 1% on protein–ligand tasks. With MSA, antibody–antigen and protein–protein pass rates rise to 53% ± 2% and 76% ± 1%.

Speed is more interesting. With 10 loops and 200 diffusion steps, ESMFold2 predicts a 1024-residue structure in 15.8 seconds; ESMFold2-Fast needs 9.4 seconds. The paper stresses that both sit at the upper-right of the accuracy–throughput Pareto frontier.

That matters for AIDD. If structure prediction is only accurate, it is an analysis tool. If it is accurate and fast, it becomes a primitive inside the design loop. ESMFold2's distinctiveness is exactly there: rather than chasing the highest accuracy alone, it combines single-sequence representation, complex folding and high-throughput inference to serve binder search.

03 — Against RFdiffusion and BindCraft: less a generator, more a scorer and prior inside a search engine

Talk about protein design and RFdiffusion and BindCraft come to mind.

RFdiffusion generates a backbone first, then inverse-folds a sequence with tools such as ProteinMPNN; it is essentially a structure generator, good at generating new backbones under constraints, especially motif scaffolding and binder backbone generation.

BindCraft takes another route: treat AlphaFold-Multimer as a differentiable scoring function and back-propagate through sequence space by hallucination. It does not generate backbones and filter; it lets the structure predictor take part in sequence–structure co-design.

The ESMC + ESMFold2 binder design route sits between the two but is not identical to either. It is closer to a system of language model prior + structure prediction scoring + large-scale search: ESMC makes candidate sequences look like proteins nature could fold and express, ESMFold2 judges whether a candidate forms a sensible interface with the target, and metrics such as ipTM and distograms rank the flood of candidates so that a few reach the bench.

So ESMFold2 here is not RFdiffusion generating structures from nothing, nor a fully differentiable hallucination optimisation of the BindCraft kind. It turns a folding model into a discriminator and structural prior for high-throughput design search.

The authors designed minibinders and single-chain antibody fragments (scFvs) against five targets — PDGFRβ, EGFR, PD-L1, CD45 and CTLA-4 — and validated them with BLI, cell staining, competition ELISA, PD-L1/PD-1 functional blocking assays and cryo-EM.

A few results deserve pulling out.

First, adding inference-time compute does raise experimental hit rates. In the harder antibody design setting the hit rate went from 12.1% to 21.0%. The scale and quality of digital experiments can be converted into wet-lab hit rate.

Second, several designs reach nanomolar affinity: an EGFR minibinder around 0.29 nM, a CTLA-4 minibinder around 0.068 nM, a PD-L1 minibinder around 1.7 nM. For early de novo binder design, that is not weak binding.

Third, the cryo-EM structure of the EGFR–minibinder complex is about 1.204 Å RMSD from the computational model, meaning that at least in this case the designed binding conformation was realized experimentally.

The most useful takeaway is not that the model succeeds in one shot. It is that ESMFold2's efficiency makes large-scale structural screening practical, compressing expensive experimental exploration into a loop of many in-silico searches and a few experimental checks.

04 — Sparse autoencoders: does the model really hold a dictionary of protein concepts?

The usual objection to protein foundation models is that they are just very large correlation machines, and that talk of a world model is a nice narrative. This paper does not answer only with benchmarks; it dissects ESMC's latent space with a sparse autoencoder (SAE). Roughly, an SAE breaks a high-dimensional neural representation into sparser, more interpretable features, looking for the "concept directions" the model reuses internally.

Figure 4 shows the hierarchical organization of those features, from local amino-acid properties and secondary structure through tertiary motifs to domain and fold level, low-complexity regions, cellular localization and functional sites.

The paper shows some representative features: catalytic motifs, disulfide bonds, beta-barrel scaffolds, DNA polymerase clamp domains, membrane helix transitions, glycoside hydrolase active sites. The model's interior is not just fuzzy sequence similarity; directions corresponding to biochemistry, structural biology and protein function appear inside it.

The implication is that a foundation model may be more than an embedding extractor. If SAE features correspond stably to structural and functional concepts, a black-box scorer can become a tool for mining mechanistic hints, retrieving remote functional homologues, or helping annotate targets.

Of course, these "concepts" are still statistical structure inside model space, not a complete biophysical mechanism. It is more of an entry point: a chance to observe how the model organizes the protein world internally.

05 — ESM Atlas: putting protein dark matter into one coordinate system

The paper ends by pulling the view out to a very large scale.

Using the ESMC and ESMFold2 representation system, the authors built the ESM Atlas, covering 6.8 billion protein sequences and predicting 1.1 billion atomic-resolution structures. The scale is striking: not annotation for a handful of model organisms, but an attempt to place a large part of the protein space that life on Earth has explored into a single coordinate system.

The most striking visual is the large UMAP map. Each point is not just a sequence but a projection of protein family, structure, function and environmental source into the model's representation space. The authors go on to show the RNase H region, the Cas12/TnpB region, and connections between uncharacterized metagenomic proteins and known functional features.

The point is that many sequences have no clear annotation in traditional databases, and some are barely similar to anything known. But if they activate function-related features in model space, they can be pulled back into an interpretable biological context.

ESM Atlas is, in other words, a map of protein dark matter. It cannot replace experiments, but it can tell you which unknown proteins might be worth a look, where remote links between families might exist, and which structural or functional themes recur across environments.

This is also where the ESM family differs from many pure structure predictors. The dominant narrative of AlphaFold-class models is: given an object, predict its structure. The dominant narrative of ESM Atlas is closer to: given the whole natural protein space, organize structure–function relationships.

06 — Where to stay calm

The results are impressive, but I would keep at least three boundaries in mind.

First, ESMFold2's benchmarks are strong, but success at structure prediction does not mean every downstream function is reliable. Dynamics, conformational change, membrane environments, post-translational modifications and intracellular concentration effects are not fully covered by a single static structure model.

Second, binder design here has experimental validation, which is one of the paper's weightiest contributions. But those results come from the authors' chosen targets, pipeline and selection strategy. Transferred to harder targets, tighter developability constraints or a stricter in vivo setting, the hit rates may not carry over.

Third, "world model" is an attractive phrase, but I read it as: the model learned a representation system usable for predicting, designing and organizing protein biology. It is not a digital life scientist holding complete causal mechanisms. It can suggest leads and accelerate search, but experiments and mechanistic work still close the loop.

Seen this way, ESMFold2's value is not replacing AlphaFold3, Boltz, Chai, RFdiffusion or BindCraft. It completes a distinctive piece of the puzzle: using large-scale protein language model representations to connect high-throughput complex folding, therapeutic binder search, model interpretability and a life-scale map.

07 — Closing

Looking back, the most interesting thing about ESMFold2 is not that it is more accurate than ESMFold1, nor that it beats someone on a benchmark. It pushes the ESM language-model route from single-chain structure prediction into complex interaction prediction and a closed experimental design loop.

If ESMFold1 showed that MSAs are not the only door into structure prediction, ESMFold2 goes further: when the language model representation is strong enough, the pair representation efficient enough and diffusion folding fast enough, it can become a primitive inside the AIDD design loop.

Next to RFdiffusion and BindCraft, the distinctiveness of ESMC + ESMFold2 is clear. It neither generates alone nor optimises by hallucination alone; it tries to turn the protein prior learned from natural sequence space into a design platform that is searchable, interpretable and experimentally testable.

Perhaps in a few years this paper will matter less for how fast or accurate ESMFold2 is, and more for pushing the protein language model scaling law into something more concrete: scale produces representations, and representations can be converted into structure, design, explanation and discovery.

The end of protein language models may not be reading sequences. It may be learning to read what the vast space of life is hiding.

It trained no new model and added no complex physical energy terms. Instead it used a clever trick: treat the structure prediction module of AlphaFold2.3 Multimer as a differentiable scoring function, back-propagate through sequence space by hallucination, and hallucinate a reasonable binder step by step. With high success rates across many targets, the idea proved workable, and binder design was redefined almost overnight.

But if you have actually run BindCraft, or read its loss design carefully, you quickly notice a fact: BindCraft is clever at binder design but far from perfect. It proves the road is passable rather than being a complete system. It leaves real gaps in stability optimisation, sequence naturalness, structural controllability, and adaptation to special proteins such as antibodies.

That is exactly where the next batch of models came in. They do not reject BindCraft. Almost all of them stand on its shoulders to patch, constrain and strengthen the paradigm. Today let us look at four models of the post-BindCraft era: BindEnergyCraft, FoldCraft, Germinal and mBER.

01 — BindEnergyCraft: an energy-based objective

If BindCraft was the first to bring AlphaFold into binder design, BindEnergyCraft is the first to seriously ask: what are we actually optimising?

BindEnergyCraft does not reject BindCraft's framework. It still uses AlphaFold-Multimer, still searches sequence space by back-propagation, and can still attach ProteinMPNN as post-processing. It changes exactly one thing: the objective.

The authors noticed that AlphaFold does not only output a single most-probable conformation when predicting structure; it also provides a full statistical description of structural uncertainty, especially the predicted aligned error (pAE) distribution at the level of residue pairs. In BindCraft these signals were used mostly for post-hoc filtering rather than in optimisation.

BindEnergyCraft's key innovation is to reconstruct these scattered confidence signals into a unified, differentiable global energy function, pTMEnergy. The function comes from AlphaFold's implicit modelling of structure prediction likelihood: the logit distribution from the pAE head is interpreted as residue-pair energy terms and jointly modelled over the whole complex. Unlike the heuristic ipTM, which leans toward "did it bind", pTMEnergy gives a dense and continuous gradient signal across interface residue pairs.

The results are not surprising. With the pipeline essentially unchanged, BindEnergyCraft has smoother gradient descent, more stable convergence, visibly fewer interface clashes and higher in-silico success across several targets. It is not a disruptive new model but a textbook example of taking BindCraft from "it runs" to "it runs well".

02 — FoldCraft: fold-conditioned design

FoldCraft asks a very direct question: why must a binder fold freely?

In many practical applications we do not actually need a brand-new fold topology. A scaffold that has already been validated, is structurally stable and expresses reliably is often more valuable to an engineer. FoldCraft tightens the free structural search space of BindCraft accordingly.

The core idea is to introduce an explicit fold condition into hallucination. FoldCraft picks a reference scaffold template in advance and computes the contact map of its internal residues. During design, the binder–target complex predicted by AlphaFold-Multimer is projected into a contact map and compared residue by residue against the template contact map, and the similarity between them (defined as mean squared error) is used directly as the optimisation objective.

The neat part is that a contact map encodes fold topology and local geometry at once without over-constraining the amino acid composition. More importantly, once contact-map similarity is optimised successfully, the traditional AlphaFold confidence metrics such as ipTM, pLDDT and pAE tend to improve on their own. Several loss terms that needed manual balancing in BindCraft collapse into one geometrically intuitive objective.

In practice FoldCraft does not chase structural novelty; it emphasises controllability and reusability. It suits tasks where the binder fold type must be specified, such as Ig-like or VHH topologies. In that sense FoldCraft is a clear structural guardrail on BindCraft's free-form design.

03 — Germinal: an antibody language model as a cross-entropy constraint

The work that systematically brought the BindCraft idea into antibody design is Germinal.

In antibodies, binding is nowhere near enough. A binder that looks good in structure prediction but violates basic antibody biology will often fail at expression, stability or immunogenicity. Germinal starts precisely from these repeatedly confirmed problems.

Unlike the methods above, Germinal admits from the start that antibody design is a highly constrained optimisation problem. It fixes the antibody framework sequence and allows search only in the CDR regions, guaranteeing overall folding and expressibility. At the structural level it introduces a dedicated paratope loss that explicitly suppresses framework residues from taking part in binding, forcing the binding interface to be CDR-driven. It also penalises secondary structure in the CDRs so they do not form overly stable α-helices or β-sheets, keeping the flexible loop conformation antibodies normally use.

Most importantly, Germinal introduces the antibody language model IgLM at the sequence level, making the implicit standard of "does this look like a natural antibody" an explicit, optimisable cross-entropy term. During optimisation, the structural confidence losses from AlphaFold (ipTM, pAE) act together with the IgLM log-likelihood term, forming a multi-objective problem. The authors note that the two objectives usually sit on a Pareto front, so weight scheduling is needed to balance structural quality against sequence naturalness.

These constraints are not meant to make the model conservative; they make the hallucination happen inside the subspace of what is feasible for an antibody. As a result, Germinal's nanobodies are not only structurally sound but closer to natural antibody sequence distributions, with stable experimental hit rates across several targets. That makes Germinal one of the first works to achieve systematic antibody "formatting" in the BindCraft family.

04 — mBER: template and language priors for VHH design

Like Germinal, mBER targets antibody optimisation. The method is uncomplicated but very pragmatic. It starts from a mature nanobody framework, fixes the overall backbone conformation, and only varies the CDR regions.

At the sequence level, mBER uses a protein language model (such as ESM2) to provide a prior distribution that guides sampling toward sequences matching natural antibody statistics. At the structural level it relies directly on AlphaFold-Multimer to predict the binder–target complex, using confidence metrics such as ipTM as the main filter.

Unlike Germinal, mBER does not build elaborate multi-term losses. It emphasises a prior plus filter strategy: the language model restricts the search space, AlphaFold decides which designs are worth keeping. That choice lets mBER generate and evaluate millions of candidate nanobody sequences without fine-tuning any base model.

This scale marks an important shift for BindCraft-family methods: the focus moves from the success rate of a single binder to the throughput and usability of a platform. mBER shows that with a well-chosen structural template and sequence prior, even a relatively simple scoring strategy can deliver decent hit rates in real experiments.

05 — Closing

Looking back at the work that followed BindCraft, one judgement is hard to avoid: the competitive focus in binder design is moving from how big the model is to whether the constraints are placed correctly.

BindEnergyCraft shows that heuristic metrics such as ipTM are no longer sufficient, and that the probability and energy information inside a structure predictor deserves to be taken seriously. FoldCraft shows that in many practical settings complete freedom is not an advantage and controllability matters more. Germinal and mBER show a trend clearly: as design moves toward antibodies and industrial application, biological priors and constraints are not a burden but a source of success rate.

Seen this way, the post-BindCraft era is not a race for the next generative model. It is a collective exploration of how to properly restrict a model's freedom. The valuable binder design system of the future is probably not a single model but a whole toolchain of structure prediction, language models, format constraints and experimental feedback.

Perhaps in a few years BindCraft's significance will not be how many binders it produced, but that it made the field realize something for the first time: design is not about letting the model do as much as possible, but about knowing where it has to do a little less.

Would it output a disordered plate of spaghetti, or would it miraculously fold into something sensible and bind the target? The surprising answer is the second. The MIT Sergey lab and the Baker lab at the University of Washington found that new-generation structure predictors such as AlphaFold3 have a powerful structure hallucination ability: even given a meaningless, all-X sequence, they denoise and hallucinate a structurally reasonable three-dimensional shape out of the protein-structure priors learned during training, sometimes one that binds a target specifically.

Original paper: Protein Hunter: exploiting structure hallucination within diffusion for protein design.

Building on that observation, they developed Protein Hunter: a framework that needs no fine-tuning, is efficient and general, and turns the AI's hallucination "flaw" into a weapon for zero-shot binder design.

01 — From structure prediction to structure hallucination

Existing design methods fall roughly into two families.

• Generative methods such as RFdiffusion can generate new folds with a diffusion model, but because they build a backbone first and design the sequence afterwards, they need heavy sampling and filtering, and it is hard to satisfy structural and sequence constraints at the same time.
• Optimisation methods such as BindCraft and BoltzDesign alternate between structure and sequence with gradient descent. They are precise but converge slowly, get stuck in local minima, and need enormous compute for larger systems and multimers.

The Sergey lab noticed that diffusion-based predictors such as AlphaFold3, Boltz-2 and Chai-1 show a "hallucination" behaviour on extreme inputs: given a fully random or unknown sequence, they still output a plausible, high-confidence protein structure. It resembles the way image diffusion models hallucinate scenes out of noise.

The core idea of Protein Hunter is to convert that hallucination from an error into a tool: use the predictor's internal prior to generate foldable protein structures.

"What is distinctive here is that, unlike earlier hallucination models such as BindCraft, it does not co-optimise sequence and structure. It trusts AlphaFold3's diffusion module for structure and MPNN's inverse folding for sequence, using an all-X input to design structure first and sequence later. The initial conformation is not constrained by any sequence, which lets AF3's structural prior do more work."

02 — From an all-X sequence to a high-quality design

Protein Hunter requires no fine-tuning and no changes to pretrained parameters; it relies entirely on an existing AlphaFold3-class model to do the design.

The workflow is one loop:

1. Structure hallucination: feed an all-X sequence (unknown residues) and let the AF3 diffusion module hallucinate an initial structure and binding pose.
2. Inverse folding: use ProteinMPNN (or LigandMPNN) to design the most likely amino acid sequence for that backbone.
3. Structure re-prediction: feed the new sequence back into AF3 to obtain the updated structure.

Then iterate until structure and sequence converge on a high-confidence protein, in the same spirit as AF2Cycler's use of recycling to improve design accuracy. The loop inherits the generation speed of a diffusion model and the stability of an optimisation method.

On compute, Protein Hunter needs about 10 seconds for a complete design cycle on a 100-residue protein, and 130 seconds even for a 900-residue one. The whole process relies on no external fine-tuning or diverse sampling; quality comes from the cooperative convergence of hallucination, inverse folding and re-prediction.

03 — One framework, many tasks: from unconditional folding to all-atom binder design

Thanks to AlphaFold3's all-atom architecture and prediction accuracy, Protein Hunter is strikingly general. The authors validated it across several design tasks.

Unconditional de novo design. For proteins of 100 to 700 residues, Protein Hunter reaches pLDDT around 0.8 within five cycles while keeping high structural designability (stable TM-score). Compared with a single diffusion pass, structures are more compact and folds more reasonable.

Protein–protein binder design. Against RFdiffusion, Protein Hunter paired with Boltz-2 scored higher ipTM on 9 of 11 targets, and with Chai-1 on 7. It can even design binders against a trimer such as TNF-α without disrupting the trimer.

Small molecules and cyclic peptides. Through the iterative loop, Protein Hunter beat RFdiffusion-AA and BoltzDesign on success rate across four small-molecule targets. For cyclic peptides, adding cyclic positional encoding let generated peptides bind targets such as MDM2 stably, with a clear improvement in structural confidence.

Multi-motif scaffolding and partial redesign. For a heme-binding protein, key motifs can be fixed while the model re-hallucinates the surrounding scaffold. In antibodies, the framework can be kept while only the variable loops (CDRs) are optimised. Both show that Protein Hunter supports precise partial redesign.

Cross-molecule systems. The framework also supports generating RNA- and DNA-binding proteins, raising interface confidence (ipTM) over several rounds to design nucleic-acid recognition sites in a controllable way.

04 — Less helix, more folding

Diffusion models are powerful generators but tend to produce α-helical topologies, at the cost of structural diversity (an old complaint).

Protein Hunter addresses this with a simple mechanism: adding a negative helix bias to the pair features of the AF3 Pairformer module. The bias suppresses helix formation during diffusion and encourages the model to explore β-sheet conformations. The share of β-rich structures rises markedly and the folded space becomes more diverse.

This also shows that biasing features during diffusion can directly steer the secondary-structure distribution of generated proteins. The experiments confirm Protein Hunter outperforms plain single-pass RFdiffusion on all design tasks. The authors plan to add sparse autoencoders to identify structural features and guide the model into even richer fold space.

"This is close to the earlier BindCraft from the same lab, which also had a helix reward/penalty parameter. It also shows, indirectly, that without extra constraints both AF2-based (BindCraft) and AF3-based (Protein Hunter) models converge readily to helical conformations and lose sampling diversity."

05 — Closing

Protein Hunter puts protein design back into the era of "no need to build a new model". It exploits the generative structure hallucination inside AlphaFold3 and, through a structure-first, sequence-later loop, systematically removes the bias of single-pass diffusion to deliver fast, general, high-success design.

Because it builds on an all-atom structure prediction model and works zero-shot, it does not depend on extra training data or fine-tuning, whether the predictor is today's AlphaFold3 or a stronger AF3-class model yet to come. Protein Hunter represents a new "structural hallucination driven" line of thinking, pushing generative design toward something more efficient and more flexible.

"In only two years, RFdiffusion went from the residue-level RFD1 to the atom-level resolution of RFD3, like an artist moving from a pencil sketch of outlines to a scientist rendering atomic detail with light and shadow. Every stroke of protein design is starting to touch reality."

On September 18, 2025, David Baker's group posted RFdiffusion3 (RFD3) on bioRxiv: the first diffusion model capable of all-atom de novo design of molecular interactions.

Original paper: De novo Design of All-atom Biomolecular Interactions with RFdiffusion3.

The last five years of protein design have moved at the pace of an industrial revolution, and each new generation of tools reshapes how we read the protein landscape. The RFdiffusion series in particular has carried generative AI into structural biology and is changing how we imagine designing new molecules of life. This post follows the three generations and the paradigm shift they represent.

01 — RFdiffusion1: from hallucination to controllable generation

On July 11, 2023, the Baker lab published De novo design of protein structure and function with RFdiffusion in Nature, introducing the RoseTTAFold Diffusion (RFdiffusion) system. It turned RoseTTAFold2 into a diffusion-based backbone generator, producing protein structure in three dimensions by adding noise and then denoising. After several denoising rounds, amino acids deform into complex shapes carrying features of real proteins, α-helices and β-sheets among them.

"Anyone who has worked on computer vision will find this familiar: it is the DALL-E recipe for image generation, with polypeptide chains in place of pixels."

Before RFdiffusion, design relied on hallucination and inpainting. The first creates random amino acids and optimises them with an AI model; the second takes a motif and builds the rest around it. Both lack precision.

RFD1's breakthrough was to go beyond the blurriness of those early hallucination methods and generate precise backbones under user-specified topology or interface constraints. Combined with ProteinMPNN for fast inverse folding and AlphaFold2 for scoring and ranking, the pipeline's success rate rose by two orders of magnitude: from tens of thousands of screens to a few hundred designs for a high-affinity binder.

That made it possible to generate proteins that resemble real ones but have never existed in nature, and it marked a watershed. Protein design moved from inefficient, experience-driven work to a scalable pipeline. Many early breakthroughs, including the SARS-CoV-2 and PD-1/PD-L1 inhibitor designs, were done in the RFD1 framework.

02 — RFdiffusion2: atomic-level side-chain constraints

RFD1 could generate new folds and binding interfaces, but it was weaker at more complex functional design, such as precise chemical interactions, and its resolution stayed at the residue level.

In April 2025 the Baker lab released RFdiffusion2 (RFD2), which keeps the residue-precision diffusion model and adds conditional constraints on key side-chain atoms. You can specify the critical residues of a catalytic site, or particular atoms in a ligand-binding pocket. The model keeps global conformational plausibility while giving precise local control, which the authors call "point control". RFD2 still cannot freely construct every side chain the way an all-atom model could, but it broke ground on small-molecule binding and catalytic residue placement.

In short, RFD2 took RFD1's ability to generate plausible structures to the point of interactions that actually work, opening the era of atomically precise protein design.

03 — RFdiffusion3: all-atom diffusion and heterogeneous systems

On September 18, 2025, only five months later, the Baker team released RFdiffusion3 (RFD3), completing the jump in modelling resolution: it abandons residue-level coarse graining and diffuses directly at the atomic level. Backbone, side chains, every atom position can be modelled and controlled during denoising.

RFD3 uses a unified atomic representation: each residue is represented as 14 atoms (4 backbone and 10 side-chain atoms). For residues with small side chains, such as glycine, virtual atoms fill the gap so that every residue has the same representation. This uniformity lets the model learn structure generation without depending on amino acid identity.

Architecturally RFD3 is a Transformer U-Net with three parts:

• A downsampling module that encodes noisy atomic coordinates together with residue-level features.
• A sparse transformer that processes tokenised information through sparse attention, letting only spatially neighbouring atoms and residues attend to each other, which improves efficiency and fights overfitting.
• An upsampling module that fuses fine atomic features with coarse residue features to predict the denoised coordinate update.

The simplification also buys computational efficiency. By removing AlphaFold3's compute-heavy 48-layer Pairformer and triangle attention, RFD3 drops to 168 million parameters, roughly half of AlphaFold3, and infers about ten times faster than RFdiffusion2, which makes large-scale screening and rapid iteration practical.

"RFdiffusion3 has not released its source yet, so the technical details are still guesswork. Knowing the Baker lab, the code will not be far behind. Interestingly, RFD3 swapped RoseTTAFold-All-Atom for AlphaFold3, so some of the performance gain may come from the change of predictor."

04 — The functional leap: designing all biomolecules

The functional advances are several.

First, atomic precision. Researchers can specify hydrogen-bond donors and acceptors, solvent-accessible surface area, ligand burial and similar atomic-level constraints, so generated structures satisfy functional requirements at atomic resolution.

Second, RFD3 handles non-protein molecules naturally: small molecules, DNA, RNA, mixed systems. Designing a DNA-binding protein, for instance, it co-generates the protein scaffold and the DNA shape without pre-fixing the DNA conformation, and the designed binders reach 5.89 ± 2.15 μM affinity. For small molecules it diffuses ligand and protein together, sampling a range of binding poses.

Third, enzyme design. The team designed cysteine hydrolases, screened 190 designs, and found 35 with multiple-turnover activity; the most active enzyme far exceeded all previous designs, confirming that RFD3 can embed an active site inside a complex scaffold.

"I will not go into the rest for length; read the paper if you are interested. Worth noting: protein PTMs, glycoproteins and other long-standing hard cases are still unsolved, which shows how far the current models are from true all-atom design."

A summary of the three generations:

05 — Closing

Early in 2024, as a third-year undergraduate, I was working on antibody design with Protein Generator built on RFdiffusion1. Back then I could not have imagined that all-atom RFdiffusion3 would exist by 2025.

The path shows not only a leap in model capability but a shift in the field's question, from can we generate proteins to can we generate proteins that are actually useful, which is the real milestone for AIDD. From RFD1 to RFD3, two years turned protein design from a residue sketch into an atomic hologram. That is not just a resolution increase but a change in the underlying logic. It points at a new possibility: future drugs and molecules will not be found by time and chance but rendered, atom by atom, the way text-to-image models render a picture.

"'Code to Complex' works two ways. Code is both computer code and the code of life spelled out in a protein sequence, and the title captures the essence of the change: we are moving from reading and writing the genetic code to writing functional modules of life."

If you were a programmer, how would you "code" a protein?

This is not a privilege reserved for the creators of Greek myth. It is happening now. Over the past few years AI has learned not only to predict how proteins fold but to design proteins that nature has never made: proteins that bind a virus precisely, neutralize a toxin, or tune the immune system. All of it comes from a quiet computational revolution.

On September 1, 2025, Cell's Structure published the review Code to complex: AI-driven de novo binder design. Having read it, I would call it a high-quality entry point to AI binder design: a precise and generous map of how AI drives de novo protein design from theory toward therapeutics. We want to design not only structures but functions; not only imitate nature but begin to surpass it. So let us follow the review's own thread through forty years of protein design.

01 — Introduction

The original paper: Code to complex: AI-driven de novo binder design, Structure, Cell Press, open access.

02 — Milestones: from rational design to generative AI

The review's main figure condenses the milestones of de novo binder design. Looking back, the history of protein design reads like an evolution from experience toward automation.

• As early as 1988, DeGrado and co-workers built the first successful artificial four-helix bundle by rational design. What they relied on was a deep understanding of how hydrophobic cores and polar residues should be arranged, what you might call the first principles of proteins. The achievement was modest in size but proved a fundamental proposition: humans really can design proteins from scratch.
• In the 21st century, computational methods made design systematic. In 2003, the Baker lab's TOP7 became a milestone: the first de novo protein generated entirely by computation, with no natural template, that folded stably. The method was physical energy function optimisation with Monte Carlo sampling. From then on protein design was no longer only an art of structure; it became a science you could compute and optimise.

"Baker's legendary road as the hand of God begins here, and protein design formally enters the computable stage. (PS: your author was born right about then.)"

• Over the following decade researchers aimed at harder functional design: binders against the conserved region of influenza hemagglutinin, screened at high throughput with yeast display. But those methods still leaned heavily on expert intuition, with limited success rates and poor generality.
• The real turning point came in 2020: AlphaFold2 arrived and reached near-experimental accuracy in the 14th Critical Assessment of protein Structure Prediction (CASP14). Predictions were astonishing, but more importantly the deep learning architecture underneath became the foundation for generative design models. Suddenly AI could not only predict protein structures; it could create them.

"DeepMind pushed AI prediction so close to experiment that CASP14 conceded the single-chain structure prediction problem was largely solved, which set off an industrial revolution in protein prediction and design."

• In the last two years a series of tools — RFdiffusion, ProteinMPNN, BindCraft — raised design success rates by one to two orders of magnitude. Today a researcher can generate a few dozen candidates in silico and expect a decent chance of a high-affinity, high-specificity binder. Protein design has entered its automation era.

03 — The AI toolbox: how the main strategies work

Despite the proliferation of tools, the review groups current design strategies into four classes, each a different design philosophy.

Hallucination grew out of the trRosetta framework. The idea is neat: invert the structure prediction network. A model built to go from sequence to structure is repurposed to generate new foldable backbones from random noise. Because it depends on no existing template, it greatly widens the range of designable structures.

MaSIF (molecular surface interaction fingerprints) takes a more visual approach. Instead of starting from global structure, it analyses the chemical and geometric features of protein surfaces to find patches likely to interact. Once such a patch is found on a target, MaSIF can search for or design a complementary surface for precise docking. This suits targets with no obvious active site, which classical methods handle poorly. It has produced binders against the SARS-CoV-2 spike protein and PD-L1.

RFdiffusion + ProteinMPNN is the popular current workflow. RFdiffusion is a diffusion-based backbone generator: an architect sketching the overall three-dimensional shape. ProteinMPNN is the interior designer, assigning an amino acid sequence that folds stably and functions. AlphaFold2 then scores the result (pLDDT, pTM) and predicts how it binds the target. The success rate of this pipeline is high enough that many groups need only screen a few dozen designs to obtain a usable binder.

The newer BindCraft represents the next leap and was formally published in Nature days ago. It co-folds target and binder from the start, simulating their interaction in real time. That dynamic strategy handles the flexible conformational changes of a binding interface better, reaching 10–100% success on some difficult targets: precise design in one step.

04 — Bench to bedside: where designed proteins are already used

If AI protein design is a new scalpel, its applications are spreading quickly across biomedicine.

Against toxins: a new route to antivenom. Snake venom has long been a hard problem for antiserum therapy, which traditionally requires immunizing animals and extracting antibodies, an expensive and cumbersome process. The Baker lab used RFdiffusion against α-neurotoxins and cytotoxins to design a series of small binding proteins. They are extremely thermostable (some Tm above 95 °C) and reach nanomolar to picomolar affinity. Most striking: in animal models, dosing even 15 minutes after envenomation gave 100% survival. This may be the shape of future antitoxin therapy: no animal immunization, designed entirely in silico, able to respond quickly to new toxin variants.

"Note that the Baker lab did not design antivenom antibodies directly but de novo mini-binders. Antibodies are harder to design, and mini-binders are more stable and easier to produce at scale."

Precise control of the immune system. The complexity of immune signalling makes drug development a game of whack-a-mole. Conventional TNF-α inhibitors block both TNFR1 and TNFR2, even though TNFR2 is anti-inflammatory. With AI design, researchers obtained antagonists that target only TNFR1, avoiding that side effect. More neatly, by switching to a partial diffusion optimisation strategy, the same workflow yielded agonists for OX40 and 4-1BB that activate T cells and strengthen antitumour immunity. These designs beat natural ligands on activity and add small size, high stability and potential oral availability, giving immunotherapy a new toolbox.

Emerging frontiers: gene editing and antimicrobial resistance. Applications keep expanding. Anti-CRISPR proteins (Alcrs) give gene editing a brake, enabling finer spatiotemporal control. Antibacterial proteins designed against the heme-uptake systems of resistant bacteria offer a new angle on global antimicrobial resistance. AI has even produced high-affinity binders against intrinsically disordered protein regions (IDPs), long considered undruggable. Together these describe a future in which protein drugs are not limited to a handful of targets but can be tailored to a wide range of biological processes.

05 — Challenges: technical limits

Like every review, this one sets out the open problems.

Technical barriers and resource distribution. Running BindCraft, RFdiffusion and similar models efficiently needs serious GPU capacity, which not every non-computational lab or small company can afford. Successful design also requires a deep understanding of the target's structural biology, so teams need both computational and experimental expertise.

The gap between prediction and experiment. Metrics such as pLDDT and pAE_interaction correlate reasonably with experimental success, but far from perfectly. Some designs that look excellent in silico fail to fold, fail to stay soluble, or fail to bind. Closing that gap needs more high-quality experimental data feeding back into training; open databases such as the Protein Design Archive are helping.

Limited target scope. Most successes are still on relatively rigid, surface-exposed targets. Highly dynamic membrane proteins, glycosylated proteins and protein–nucleic acid complexes remain hard. Widening the range requires algorithms that handle flexibility, interfacial water and chemical modifications better.

06 — "Code to Complex": a new era for binder design

"Code to Complex" carries two meanings at once: code as computer code, and code as the code of life in a protein sequence. The title captures the essence of the change: we are moving from reading and writing the genetic code to writing functional modules of life.

What this review shows is not just a series of technical advances but the birth of a new paradigm. Protein design is turning from an art into an engineering discipline; from intuition and experience toward data and algorithms; from imitating nature to surpassing it.

Future protein design will be more precise, more general and more accessible. Perhaps soon, designing a binder against a newly emerged pathogen will be as simple as building a web page today: enter the target, choose the design parameters, press generate.

At that point we may truly realize Richard Feynman's line: "What I cannot create, I do not understand." AI is giving us that creative ability: not only to understand life, but to design it.

"We want AI protein models to learn the underlying logic that goes beyond the folding rules of solved natural proteins, and to guide us in sampling biologically plausible regions of protein space that we have not yet explored."

Hallucination design is a generative approach to protein design that runs a structure prediction model backwards to design new sequences. Concretely: given a target protein structure (or a complex), you start from random sequences, repeatedly feed those sequences through a structure prediction module (AlphaFold2, RoseTTAFold), read out confidence metrics such as pLDDT and ipTM, and back-propagate on those metrics to optimise the sequence. Sequence and structure are co-designed this way, converging step by step on a binder with high predicted confidence that satisfies the design goal.

The method was first proposed by the Baker lab for generating monomers and symmetric assemblies (the 2022 Science paper on hallucinating symmetric protein assemblies), and in recent years it was extended to binder design. Unlike classical design built on physical energy functions, hallucination design needs no predefined fold scaffold. It explores sequence space using the "energy landscape" a trained deep model has already learned.

01 — What we mean by hallucination design

The background is the breakthrough in deep learning for structure prediction. Models such as AlphaFold2 and RoseTTAFold predict structures with very high accuracy, and the confidence distributions they output (pLDDT, pAE) can be read as a proxy objective for how well a sequence matches a structure. Researchers found that optimising gradients on those confidence metrics can "hallucinate" entirely new protein structures. Binder design builds on this: the metrics from the structure prediction module are combined into a loss function, and back-propagation generates sequences that bind a target. BindCraft, for instance, uses AlphaFold2 weights, feeds the target together with a random sequence, and keeps adjusting the sequence to minimise the loss. The practical consequence is that a designer can explore structure space directly inside the neural network, reaching novel binder sequences that template-based or docking-based pipelines struggle to cover.

02 — The main models

In practice, the leading hallucination-style binder design tools all use a pretrained deep structure predictor as both evaluator and optimiser.

ColabDesign is a collection of open-source pipelines from the Sergey Ovchinnikov lab supporting several design methods; its core module, AfDesign, is also the core of BindCraft. The platform contains P(structure | sequence) design modules (TrRosetta-Design, AfDesign and others) and P(sequence | structure) modules (ProteinMPNN). AfDesign implements AlphaFold2-based binder hallucination: the user supplies a target structure and a binder length, locally or on Google Colab, and the system iteratively optimises the sequence to satisfy the structural constraints. No retraining is required; gradients are computed through Alpha-Multimer weights.

"Wang Jue, the author of ColabDesign, did a number of interesting things to make AlphaFold2 support back-propagation; I will not go into them here for length. Because AlphaFold2 is built on JAX, ColabDesign is JAX too, which is a bit of a challenge for readers used to PyTorch. If you are curious, there are three detailed posts by Peldom on Zhihu, AI + protein design | gradient back-propagation through AlphaFold2."

BindCraft (Pacesa et al., 2024) is one of the representative systems that combined these pieces into high-quality designs, reporting success rates of 10–100%. The pipeline: take a target structure and a randomly initialised binder sequence, call AlphaFold-Multimer to predict the complex, read ipTM, pLDDT and pAE, take a weighted sum as the loss, and back-propagate into the binder sequence (one-hot encoded) to raise ipTM and pLDDT. No experimental data and no known binding site are required; everything comes from AlphaFold-Multimer's parameters. ProteinMPNN is used alongside to improve sequence manufacturability and diversity. The whole pipeline is automated: one design run produces hundreds of candidate sequences, which are then tested experimentally.

"Strictly speaking BindCraft is a hybrid: ColabDesign plus ProteinMPNN. You can think of it as using AlphaFold-Multimer back-propagation to co-design structure and sequence at the binder interface, then using ProteinMPNN to redesign the scaffold region away from the interface. If the wet-lab success rates (10–100%) hold up, separating interface design from scaffold design is a promising recipe for binders more generally."

Several other systems extend the same core. BindEnergyCraft (2025 preprint) reinterprets AF2 confidence as an energy model (pTMEnergy); changing nothing but the loss function measurably improves hallucination design and beats both BindCraft and RFdiffusion in vitro. FoldCraft adds an extra constraint (contact-map similarity) for scaffold-conditioned design, but has only been evaluated in silico with AF3, so its value is still open. Broadly, these methods share a modular shape: a structure predictor at the centre, several loss terms on top (prediction accuracy, interface affinity), and gradient or sampling search through sequence space.

Beyond AlphaFold2-based back-propagation, AlphaFold3 opened a second route. BoltzDesign1 (2025) is built on AlphaFold3 (via the open Boltz-1 implementation) and widens the design target to small molecules, nucleic acids, metals and modified proteins. Without any experimental optimisation, hallucination driven by Boltz-1 structure prediction produced binders with high predicted affinity for a range of ligands, including drug molecules, DNA, RNA, metal ions and covalently modified proteins. Experimental validation is still pending, but reported in-silico success rates exceed 90%.

03 — Where it gets used

Antibodies and nanobodies are one important setting. In work by Takafumi and others, AfDesign was used to redesign antibody CDR loops, optimising antigen binding affinity without a structural prior: given a known antigen–antibody complex, CDR sequences are generated at random and iteratively optimised, improving binding after substitution.

Peptides and cyclic peptides are another active area. Kosugi et al. (2022) used AfDesign binder hallucination to design peptides against an antibody–protein interface and added a solubility loss to control hydrophilicity; adjusting the loss weights kept affinity high while markedly improving solubility. Other work hallucinates cyclic peptides against protein–protein interaction interfaces, generating soluble binding peptides in a targeted way.

Allergens, cell receptors and similar targets are where BindCraft's own wet-lab validation lives. BindCraft was tested experimentally on a range of targets, including pollen allergens (designing binders that reduce IgE binding), Cas9 (modulating gene-editing activity) and bacterial enterotoxins (reducing cytotoxicity). BindCraft also developed AAV vector retargeting: designing binders that bind a specific cell-surface receptor redirects the viral vector to a chosen cell type. Binders, then, are not only inhibitors; they can also be delivery modules.

04 — How it compares

Against docking or template-driven methods, hallucination design differs in kind. Classical pipelines first screen or prepare a library of protein scaffolds (template libraries, random protein libraries), then use docking algorithms such as Rosetta Docking to evaluate binding modes and energies. That requires a great deal of prior information (a known binding site, or structures of similar binders), and because physical force fields and scoring functions are imperfect, you often have to design thousands or tens of thousands of candidates to find a few binders. The Rosetta design work from the Baker lab, for example, needed tens of millions of candidates across 13 target sites, and only a handful reached sub-micromolar affinity. Such methods adapt poorly: with no suitable template or library you cannot design a new structure at all, and the match to the binding-site shape has to be close. Hallucination needs no scaffold template and can create entirely new folds in sequence space, which makes design more flexible.

Against sequence-based methods (statistical co-evolution models, language model generation, and ProteinMPNN-style P(sequence | structure) predictors), the difference is that hallucination methods evaluate folding dynamically. Sequence-based approaches learn sequence statistics or predict a sequence for a given backbone, but they do not re-evaluate whether the result folds. Hallucination considers sequence and structure together and, at every iteration, predicts the complex and optimises against structural confidence. That makes it stronger at generation: it can explore folds outside known sequence and structure space, whereas sequence models tend to stay close to their training distribution.

Affinity prediction differs too. Docking methods use empirical force fields to compute binding energy; hallucination relies on confidence metrics from the structure predictor. Those metrics are approximations of structural quality, but they do reflect the network's confidence that a complex is feasible. On adaptability, hallucination design absorbs new loss terms easily (a solubility objective, say) and can use different strategies for different target types (protein, nucleic acid, small molecule). Classical methods usually need a workflow hand-tuned per target, whereas the deep learning approach is modular and iterates faster.

05 — Where hallucination design still struggles

One problem is design accuracy versus stability. Current methods optimise a prediction model's confidence, but experimental validation still finds mismatches and unstable structures. In Wicky et al.'s symmetric assembly hallucination work, many designs failed to express at all, which points at an overfitting risk: the network may generate sequences that are not stable in a biological system. Improving expressibility and foldability (through more constraints or extra losses) is a priority.

Another is model limits and generalisation. AlphaFold2 is powerful but depends on multiple sequence alignments, which may bias predictions for purely designed sequences; some work uses single-sequence AF2 or fine-tuned models to reduce that dependence. A third challenge is scope: hallucination today mostly targets protein–protein interfaces, and highly flexible targets remain difficult.

"Back-propagation requires dropping AlphaFold2's three recycles to zero, which costs some structure prediction accuracy and can feed through into design quality."

Looking forward: any in-silico design and optimisation ultimately has to come back to wet-lab validation. Pairing hallucination design with automated wet-lab loops (microfluidic high-throughput screening, machine-learned immunogenicity prediction) would accelerate the closed loop. As stronger structure predictors and better optimisers arrive, hallucination-style binder design should work through its current limitations and move toward more reliable rational drug design.

06 — A closing note

I have been an AIDD intern at XtalPi for four months now, and I am grateful to the colleagues who have looked after me. In this field, diffusion models, LLMs and GNNs are all doing their own spectacular thing, the architectures are broad and the iteration is fast, and keeping up feels like trying to shake a tree as an ant. I started this account to share what is happening in AI4Protein and to organise what I have learned so far. I am far from expert, so corrections are very welcome, and I am happy to argue about where the boundaries of AIDD actually are.