Protein language models and the wider protein AI field
· 45 min read
Digests are written primarily by LLMs with only readability edits by me. If this one reads like slop, please shoot me an email at max.zhuang@berkeley.edu.
Protein AI has several major families, each learning a different relationship between a protein’s sequence, shape, and behavior. ESM made large-scale learning from amino-acid sequences a reusable foundation. AlphaFold made accurate structural prediction widely useful. Chai, Boltz, RoseTTAFold, and newer generative systems extend the field toward molecular interactions and the creation of new proteins. Their contributions become clearer when we compare the questions they answer.
This is a broad guide to the major research families and companies, including their important predecessors and newer releases. The comparison is current to 17 September 2026. Sources include peer-reviewed papers, recent preprints, and company announcements. No survey can enumerate every model or reveal private architectures; where a company has not disclosed its internals, that limit is part of the comparison.
The central distinction is what the model represents and predicts.
The tasks and architectures
A protein’s sequence is its ordered chain of amino acids. An amino acid within that chain is called a residue. Its structure is a three-dimensional arrangement of atoms; its function is what it does in a particular biological setting. Proteins also move, interact with other molecules, and behave differently under different conditions. The backbone is the repeating main chain; side chains are the chemical groups that distinguish amino acids. All-atom models include this atomic detail beyond a simplified backbone representation.
| Model category | What it learns or predicts | Representative families |
|---|---|---|
| Sequence representation | Numerical features describing residues or whole proteins | ESM-2, ESMC, ProtT5, AMPLIFY, Ankh |
| Sequence generation | A distribution over amino-acid sequences | ProGen, ProtGPT2, RITA, EvoDiff, DPLM |
| Family-context modeling | Sequence relationships informed by related proteins | MSA Transformer, PoET, E1, ProtMamba |
| Multimodal representation or generation | Relationships among sequence, structural descriptions, and function | ESM3, DPLM-2, SaProt, ProstT5, ProTrek |
| Folding and cofolding | Coordinates for a specified protein or molecular complex | AlphaFold, Chai-1, Boltz-2, ESMFold2, RF3 |
| Inverse folding | Sequences compatible with a supplied structure | ProteinMPNN, LigandMPNN, ESM-IF1, PiFold |
| Structural generation and co-generation | New shapes, or new sequences and shapes together | RFdiffusion, Chroma, BoltzGen, La-Proteina, Latent-X |
| Conformational ensembles | Multiple structural states of the same protein | AlphaFlow, ESMFlow, BioEmu |
Several architectural terms recur. An encoder turns an input into learned numerical representations, often called embeddings. A decoder produces an output from those representations or from earlier output tokens. Attention lets parts of an input exchange information. A transformer organizes that exchange into repeated layers. A token may represent an amino acid, a short sequence fragment, or a compressed description of structure.
The learning objective is a separate choice. Masked language modeling learns to recover hidden tokens using the visible context. Autoregressive modeling predicts successive tokens from preceding context. Diffusion learns to reverse corruption, which can affect discrete tokens or continuous coordinates. Flow matching learns a velocity field: the direction and rate of change along a continuous transformation from a simple distribution to the data distribution. The last two are mathematical generation processes; their intermediate states need not describe how a real protein folds.
Conditioning means supplying information on which a prediction or generation depends. Retrieval supplies relevant examples found outside the model’s stored parameters. Both can change the evidence available to a model without changing the size of its neural network.
Geometry introduces another distinction. An equivariant network changes its coordinate outputs consistently when the input is rotated or translated. Other networks learn geometric behavior through their data and training. A mixture-of-experts model routes computation through selected neural subnetworks, so its total parameter count exceeds the parameters active for each token.
Sequence foundation models
Model sizes below use M for millions and B for billions of parameters.
ESM: Meta FAIR, EvolutionaryScale, and Biohub
ESM-1 and ESM-1b established an influential approach: hide amino acids in natural protein sequences and train a bidirectional transformer to recover them. The resulting representations carried information about contacts (residues close in space), secondary structure (local arrangements such as helices and sheets), and evolutionary relationships even though the model had learned from sequences. ESM-1b became a widely used 650-million-parameter reference model. ESM-1v emphasized predicting the effects of sequence variants without a separately trained predictor for each experiment. ESM foundations, ESM-1v.
ESM-2 scaled this masked-transformer approach across a family extending to 15 billion parameters. Its contribution was both stronger sequence representations and the foundation for ESMFold, a separate network that converts those representations into coordinates. ESM-2 by itself does not output an atomic structure. The older Meta collection also includes MSA Transformer and ESM-IF1, which consume different kinds of input. Meta’s model inventory.
ESM Cambrian, also written ESM C or ESMC, continues the sequence representation branch. EvolutionaryScale introduced it in 2024; that team subsequently joined Biohub. The current 300M, 600M, and 6B models emphasize stronger representations per parameter. Biohub’s 2026 release connects ESMC to ESMFold2 and an expanded ESM Atlas. The Atlas is a collection of indexed sequences and predictions, rather than another neural architecture. ESMC, Biohub’s 2026 release.
The family now has three distinct central roles: ESMC represents proteins, ESMFold2 predicts molecular geometry, and ESM3 models multiple protein modalities together. ESM3 is discussed below. Keeping those roles separate makes clear which ESM model a given comparison refers to.
ProtTrans and ProtT5: the Rostlab lineage
ProtTrans compared protein adaptations of several language-model architectures: ProtBERT, ProtALBERT, ProtELECTRA, ProtT5, ProtTXL, and ProtXLNet. ProtALBERT shares parameters; ProtELECTRA learns to detect replaced tokens; ProtTXL retains recurrent context; and ProtXLNet changes the order in which tokens are predicted. The study showed how broadly trained protein representations could support many biological prediction tasks. ProtTrans paper.
ProtBERT is a masked encoder. ProtT5 uses an encoder–decoder denoising architecture, although many applications use only its encoder’s embeddings. ProtT5-XL-U50 names a particular widely used variant. Using only the encoder explains why full-model and deployed-model parameter counts differ. ProtT5 provides a reusable representation that many later models build on, including the structure-token model ProstT5. ProtTrans models and documentation.
Ankh: protein-specific encoder–decoder development
Ankh, associated with Proteinea and Elnaggar’s collaborators, develops T5-derived models through experiments in protein data, masking objectives, and the allocation of capacity between encoder and decoder. Ankh2 extends training and changes the nonlinear function used inside its neural layers; Ankh3, reported in 2025, combines variable-corruption denoising with sequence completion. Its contribution is an alternative route to useful representations through training design and encoder–decoder structure. Ankh, Ankh2 model card, Ankh3.
ProGen: Salesforce origins and Profluent’s successors
ProGen is a conditional autoregressive transformer. Its original formulation combined amino-acid sequences with metadata describing biological context. Subsequent experiments demonstrated functional generated proteins in the tested families. ProGen2 expanded the family to 6.4B parameters and examined how training-data composition affects generation and variant prediction. Its contribution was evidence about both scale and the distribution of proteins used for training. ProGen, experimental study, ProGen2.
ProGen3, from Profluent, adds sparse mixture-of-experts layers and learns both sequence continuation and completion of missing spans. Its largest reported model has 46B total parameters, with about 27% active in a forward pass, so its computation per token differs from that of a dense 46B-parameter network. Its research contribution combines larger curated sequence data, sparse scaling, and experiments connecting model scale to the diversity of viable outputs. ProGen3 paper.
The public release is smaller than the largest reported model. The checked repository lists checkpoints through 3B, with Apache-licensed code and separately licensed noncommercial weights. Profluent’s E1 is another distinct family, focused on representations and retrieved evolutionary context rather than ProGen-style generation. ProGen3 releases.
Other major sequence architectures
| Family and developers | Architecture | Distinct contribution |
|---|---|---|
| ProtGPT2, Ferruz, Schmidt, and Höcker | Autoregressive GPT-2-style transformer; tokens can represent amino-acid subsequences | An influential early open sequence generator, showing protein-like distributions without first specifying a structure. Paper |
| RITA, Hesslow and colleagues | Autoregressive transformer family up to 1.2B parameters | Systematic study of sequence-model scaling, variant prediction, and functional representations. Paper |
| xTrimoPGLM, BioMap and Tsinghua | Large transformer combining masked representation learning and autoregressive reconstruction | A shared backbone for representation and generation, plus an associated xT-Fold structural model. Public artifacts include 100B weights stored at reduced numerical precision, called weight quantization; availability has noncommercial restrictions. Paper, repository |
| AMPLIFY, Amgen and Mila/Chandar lab | Efficient masked transformers at 120M and 350M parameters | Careful data curation and modern transformer components make strong representations available at modest scale. Training code, data, and checkpoints make the approach inspectable. Paper, repository |
| AIDO.Protein / GB.Protein, GenBio | Sparse mixture-of-experts encoder; the 16B model activates about 4.5B parameters per token | Large representation capacity with selective computation. Its retrieval-augmented branch adds family and structural context; current naming includes GB.Protein-RAG. Paper, current model card |
| CARP, Microsoft Research | Dilated convolutions combine spaced sequence positions through shared local filters; training recovers masked tokens | Shows that the language-model objective can work with convolution rather than attention, with favorable sequence-length scaling. Its architecture also underlies EvoDiff’s sequence branch. Paper, repository |
These families learn statistical regularities of protein sequences.
Evolutionary context and retrieval
A general sequence model stores patterns learned across many proteins in its weights. A second source of information is the set of proteins related to the particular sequence under study. Homologs share evolutionary ancestry. A multiple sequence alignment, or MSA, places corresponding positions in related sequences into a common set of columns.
Family-context models make that information available during prediction. They contribute access to relevant examples, so a small context-aware model and a large single-sequence model may be working from different information.
the learned weights.
MSA Transformer
MSA Transformer, from Meta FAIR and collaborators, alternates attention along the rows and columns of an alignment. Row attention examines positions within a sequence; column attention compares related sequences at aligned positions. Masked pretraining across families teaches a reusable model to interpret these evolutionary comparisons. The resulting architecture combines broad pretraining with explicit family evidence. MSA Transformer.
PoET and PoET-2: OpenProtein.AI
PoET models protein families as collections of sequences, using hierarchical attention within and between them. It learns ordered relationships within each sequence while reducing dependence on the order of the family examples. It can learn from homologous context without requiring every position to be supplied in an explicit alignment. PoET.
PoET-2 adds optional structural information, a hierarchical encoder, and two related decoding modes: causal sequence generation and bidirectional representation learning. This makes family context useful for both generating sequences and representing their potential properties. Its distinctive claim is that relevant context can provide information that otherwise has to be learned through larger models or more task-specific data. PoET-2.
E1, Tranception, and ProtMamba
Profluent E1 is a masked encoder with separate attention patterns within and across retrieved homologous sequences. The homologs need not be aligned residue by residue. Its contribution is an explicitly retrieval-aware representation model; it also supports single-sequence inputs. E1 and ProGen3 therefore represent different strategies within the same company. E1 paper, official release.
Tranception, from the Marks lab and Oxford research groups, combines a general autoregressive model with alignment-derived residue statistics. Those statistics contribute a separate prediction; PoET and E1 instead attend directly to related sequences. Tranception also incorporates convolutional components for local patterns. TranceptEVE combines general sequence modeling and family-specific evolutionary modeling. Their major role is variant-effect assessment: combining broad learned regularities with local family evidence. Tranception, ProteinGym and TranceptEVE.
ProtMamba explores a different network family. Its state-space architecture maintains and updates a learned state as it processes sequences, avoiding the full pairwise attention cost of an ordinary transformer. It learns from concatenated homologous proteins without explicit alignment. ProtMamba makes long evolutionary context cheaper to compute. ProtMamba.
Structural tokens and multimodal language models
Structures can be represented as coordinates, but they can also be compressed into discrete symbols. A structural alphabet assigns symbols to recurring local geometric patterns. Assigning continuous geometry to discrete codes is structural quantization, a different operation from storing model weights at lower numerical precision. This lets language-model machinery process aspects of shape, with an unavoidable tradeoff: a compact token representation does not preserve every atomic detail.
Letters become tokens. The model learns contextual features from patterns across sequences.
Local structural information becomes compact codes. Each model chooses its own alphabet and how to combine it with sequence.
The geometric generator works directly with atom positions. Learned features guide its coordinate updates.
One backbone atom per residue stays explicit. Sequence and remaining atomic detail are compressed into learned variables.
ESM3: sequence, structure, and function together
ESM3 learns across parallel tracks for sequence, structure, and function. Its masked transformer can recover missing information within and across these tracks. Unlike ESM-2 and ESMC, it models geometry and function explicitly, instead of leaving them to be recovered from sequence embeddings. ESM3 paper.
ESM3’s contribution is flexible multimodal completion within one model. Its evidence includes an experimentally characterized fluorescent protein. The paper’s “500 million years” description refers to an estimated evolutionary distance, not a simulated trajectory. The largest described model is 98B, while the publicly downloadable open variant is 1.4B. Current Biohub documentation lists the open model under MIT terms, so older access summaries can be misleading. original release, current ESM3 documentation.
DPLM and DPLM-2: discrete diffusion
DPLM, from ByteDance and academic collaborators, learns protein sequences through discrete denoising. Its generation process can reconstruct a whole sequence from corrupted information, while the same learned representations support prediction tasks. DPLM-2 adds quantized structural tokens and learns joint relationships between sequence and shape. Unlike coordinate diffusion, these models reconstruct discrete sequence and structure symbols. DPLM, DPLM-2.
Later branches address specific limitations. The DPLM-2.1 follow-up, including the released DPLM-2 Bit component, examines the fidelity lost in structural tokenization, including modeling the bits that compose structural codes. DPLM-Evo, an ICML 2026 paper, incorporates substitutions, insertions, and deletions through a latent alignment representation. That broadens denoising beyond a fixed set of sequence positions. These changes address representation and variable length. DPLM-2 follow-up, DPLM-Evo.
SaProt, ProstT5, and ProSST
| Family | How it connects sequence and structure | What it adds |
|---|---|---|
| SaProt, Westlake collaborators | A joint vocabulary pairs amino-acid identity with 3Di symbols from Foldseek’s discrete structural alphabet | Structure-aware representations with familiar masked-transformer machinery. Earlier and newer checkpoints differ in how well they handle sequence-only input. Paper and current models |
| ProstT5, Rostlab and Steinegger collaborators | A ProtT5-derived encoder–decoder translates between amino-acid and 3Di sequences | A common model for sequence/structure-token translation, representations, and structure-sensitive search. Predicting 3Di symbols does not itself produce atomic coordinates. Paper |
| ProSST, Shanghai Jiao Tong and collaborators | Learned quantized structural tokens with attention that separately handles sequence and structural information | An alternative to a fixed combined alphabet, emphasizing how the two information streams interact. Paper |
Connecting proteins to human language
Human-language descriptions provide a different kind of supervision. They name functions and relationships that may be difficult to identify from sequence statistics alone. Contrastive learning trains representations of matched protein–description pairs to be more similar than mismatched pairs.
ProtST, ProteinCLIP, and ProtCLIP connect sequence representations to descriptions through related alignment objectives. ProteinCLIP refines existing embeddings toward functional similarity; ProtCLIP adds function-informed objectives at finer sequence scales. Despite the similar names, these are separate projects. ProtST, ProteinCLIP, ProtCLIP.
ProTrek aligns three modalities (sequence, structural representation, and text) in a shared representation space. Its distinctive role is searching across those modalities, including finding proteins associated with a functional description. ProTrek.
Prot2Text / Prot2Text-V2 and ProtT3 connect protein encoders to text decoders to produce functional descriptions. ProtLLM represents proteins within interleaved protein–text inputs. ProLLaMA adapts a text-model backbone for protein-related tasks. ProteinGPT connects sequence and structure encoders to a language model. Their contribution is an interface between molecular information and human-language reasoning. Prot2Text, Prot2Text-V2, ProtT3, ProtLLM, ProLLaMA, ProteinGPT.
ProteinDT makes the text-to-sequence connection explicit. Its ProteinCLAP component aligns protein and scientific-text embeddings; another learned component connects their representations; an autoregressive or diffusion decoder models sequences. The architecture links text descriptions to protein representations and generation. ProteinDT.
Folding and molecular complexes
Structure predictors estimate geometry for molecules whose identities are specified. Cofolding extends this to multiple components, such as several protein chains or a protein with a small molecule. A ligand is a binding partner; in many cofolding benchmarks, the term refers to a small molecule.
AlphaFold: from distance geometry to atom diffusion
AlphaFold1, from DeepMind, used a residual convolutional network to predict distances between residues, then converted those predictions into a structural model. Its historical contribution was richer learned geometry than a simple prediction of which residues touch. AlphaFold1.
AlphaFold2 introduced a much more integrated architecture. Its Evoformer exchanges information between an MSA and a pair representation: learned features describing each pair of residues. Triangle operations connect relationships among triples, helping the model reason consistently about geometry. A separate structure module uses invariant point attention, which compares features in residue-centered coordinate frames, to produce a three-dimensional prediction. Recycling feeds earlier predictions back through the network for refinement. AlphaFold2.
The contribution was accurate end-to-end structure inference using explicit evolutionary and structural evidence. AF2 contains a masked-MSA learning objective, but its defining product is a structure predictor. Its output is also a structural hypothesis, rather than the time-resolved physical path by which a protein folds. AlphaFold-Multimer adapts the family to assemblies of supplied protein chains. AlphaFold-Multimer.
AlphaFold3, developed jointly by Google DeepMind and Isomorphic Labs, extends the modeled chemistry to proteins, nucleic acids, small molecules, ions, and modified residues. Reduced MSA processing feeds a Pairformer that emphasizes token and pair representations. An atom-coordinate diffusion module then generates geometry conditioned on those representations. This replaces AF2’s protein-specific structure module with a more general way to place heterogeneous atoms. AlphaFold3.
AF3’s major value is unified molecular structure prediction. Its paper also documents failures involving stereochemistry, clashes, and disordered regions. Different samples from the diffusion model are not calibrated samples of a protein’s physical equilibrium distribution.
Chai-1, Chai-2, and Chai-3
Chai-1 combines an AF3-style geometric architecture with pretrained protein-language-model embeddings and optional evolutionary or structural context. Pair-biased attention lets pairwise relationships influence the exchange of token information, and diffusion produces atom coordinates. The language-model features help when an explicit MSA is absent. MSAs can still improve predictions. Chai-1 technical report.
Its distinctive contribution is an accessible general molecular predictor that combines these sources of evidence. The released code and weights use Apache-2.0 terms. The scope includes mixed molecular complexes, rather than only the shape of one isolated protein. Chai-1 repository.
Chai-2, announced in 2025, shifts the emphasis to generative molecular design. The company describes multimodal models combining all-atom prediction and generation, with experimental characterization of generated molecules. The public reports establish that system-level scope, but do not disclose enough internals to assign it a reproducible detailed network architecture. Chai-2 announcement.
Chai-3 appears in the company’s 2026 partnership announcements. It belongs in any current survey, but those announcements provide company-reported capabilities rather than a complete technical specification or independent cross-model benchmark. Chai-1’s open release does not extend to Chai-2 or Chai-3. Chai-3 company disclosure.
Boltz-1, Boltz-1x, and Boltz-2
Boltz-1 brought AF3-class cofolding into an openly released model family. Boltz-1x added physical constraints during coordinate generation to improve chemical plausibility. Boltz-2 combines structural representations, a coordinate denoiser, confidence prediction, and an additional affinity module. Its code and weights are released under MIT terms. Boltz repository.
Boltz-2’s main addition is the affinity module. This separate Pairformer-based component predicts both binding likelihood and a continuous binding-strength estimate for protein–small-molecule settings. It adds an explicit learned property prediction to a model of molecular geometry. Boltz-2 paper.
The continuous estimate is learned from heterogeneous experimental measures, including equilibrium dissociation, inhibition, and assay-concentration measurements. Because the training data mixes these assay types, the estimate is not an assay-independent physical constant. BoltzGen and BoltzProt-1 are the family’s separate generative branch.
RoseTTAFold and RF3: the University of Washington lineage
RoseTTAFold introduced a three-track network in which sequence, residue-pair relationships, and coordinates repeatedly inform one another. That architecture offered a distinct organization of structural reasoning. RoseTTAFoldNA, also associated with the RoseTTAFold2NA repository name, extends the framework to protein–DNA and protein–RNA complexes. RoseTTAFold, RoseTTAFoldNA.
RoseTTAFold All-Atom combines polymer-level sequence representations with atom-and-bond descriptions for other chemistry. Its contribution is a shared representation for proteins, nucleic acids, small molecules, and modifications. It also supplied the foundation for RFdiffusion All-Atom, a related generator. RoseTTAFold All-Atom.
RoseTTAFold3 / RosettaFold-3, usually RF3, is a newer all-atom transformer-and-diffusion predictor. Released with AtomWorks, it places substantial emphasis on consistent molecular data processing, training artifacts, and representations of molecular handedness, or chirality. It contributes reusable open infrastructure as well as predictions. RF3 and RFdiffusion3 are different models: one predicts structures; the other generates designs. RF3 and AtomWorks, RF3 documentation.
ESMFold2 and other single-sequence folding models
ESMFold demonstrated large-scale structural prediction from ESM-2 representations without a per-query MSA search. Its contribution was the connection between sequence pretraining and rapid folding across enormous sequence collections. ESMFold paper.
ESMFold2, released by Biohub in May 2026, uses ESMC 6B representations with an all-atom diffusion-based structure architecture. It supports molecular complexes and optional MSAs, and has a faster model variant. The main change is breadth: current ESM folding models are no longer only fast predictors for isolated protein chains. Biohub releases the current ESMC and ESMFold2 weights under MIT terms. 2026 technical report, current model documentation.
OmegaFold, from HeliXon, combines OmegaPLM sequence representations with a geometry-focused Geoformer and a coordinate-producing module. RGN2 combines AminoBERT with a recurrent geometric network and a compact description of backbone geometry. Both were important independent demonstrations of single-sequence folding, though their original evidence and scope differ from today’s all-atom complex predictors. OmegaFold, RGN2.
Open reproductions and newer cofolding families
Sharing an architectural lineage does not make these efforts redundant. Training access, data curation, speed, molecular coverage, and reproducibility can be substantial contributions.
| Family and organization | Architectural relationship | Distinct contribution and current scope |
|---|---|---|
| OpenFold, AlQuraishi lab and consortium | Trainable PyTorch reproduction of AF2 | Exposes training machinery and supports controlled investigation of how data affect folding accuracy. Paper |
| OpenFold3 / OpenBind-0 | Open AF3 reproduction; OpenBind-0 uses a nearly final OpenFold3 architecture | Public training artifacts and a 2026 protein–ligand-focused release. OpenBind-0 became the default weights for recent OpenFold3 versions; preview and final-release claims should remain distinct. Repository, OpenBind-0 release |
| Protenix / Protenix-v2, ByteDance | Trainable AF3-derived family, with capacity and training improvements | Open reproduction, smaller-model branches, and current prediction/design development. Version 2 arrived in April 2026; checkpoints using different data cutoffs support different comparisons. Repository and release history |
| HelixFold / HelixFold-Single / HelixFold3.2, Baidu PaddleHelix | AF2 and AF3 reproductions; a language-model-based single-sequence branch | A substantial independent implementation, with its own hosted services. Model-specific availability differs from the surrounding software’s license. HelixFold-Single, HelixFold3, current release |
| Uni-Fold / UF-Symmetry, DP Technology | AF2/Multimer reproduction with symmetry-aware extensions | Open training and efficient treatment of large symmetric assemblies. Repository |
| IntelliFold2, IntelliGen-AI | Enlarged Pairformer, revised atom attention, and reinforcement-learning adaptation of diffusion | Explores model capacity and post-training for cofolding; 2026 developer-reported evidence. Technical release |
| OpenDDE, Aureka | Pairformer followed by intermediate structural tokens and atom diffusion | Adds a representation between whole residues and individual atoms for relational geometric reasoning. A July 2026 preview, with emerging evidence. Paper |
| Promera, Jing and collaborators | Structural predictor, diffusion coordinate generator, and assessment heads | Connects cofolding, design assessment, and generation in an open 2026 project. This integration has emerging preprint evidence. Preprint, repository |
Four further families add useful contrasts to the common AF3-derived design:
- SimpleFold, from Apple, uses standard transformer blocks and flow matching, dispensing with specialized triangle attention and explicit pair representations. It tests how far scale and structural data can replace folding-specific architectural components. Its folding and ensemble results should be evaluated within their stated protein scope. SimpleFold.
- SeedFold / SeedFold-Linear, from ByteDance Seed, explores wider pair representations and larger distilled structural datasets, meaning structures predicted by another model and then used as training examples. The Linear variant reduces the cost of triangle attention. This is a distinct named effort from Protenix, with an explicit architecture/compute tradeoff. SeedFold.
- NeuralPLexer / NeuralPLexer2 / NeuralPLexer3, from Iambic Therapeutics, develops molecular complex prediction. NP3 combines a Pairformer-based encoder with an atom-level flow decoder and a physically motivated polymer-shaped starting distribution. Its emphasis includes geometric validity and conformational changes associated with ligand binding. The released NPBench benchmark does not imply release of every NP3 artifact. NeuralPLexer3.
- Pearl, from Genesis Molecular AI, emphasizes rotationally equivariant diffusion, synthetic physical training data, and protein–ligand geometry. DeCAF-Pearl trains a new generator to imitate the original model with fewer generation steps, another use of model distillation. These company reports present physical training information and computational cost as central contributions. Pearl report, DeCAF-Pearl.
TDFold adds an emerging 2026 alternative: diffusion first models pairwise distances and orientations, producing geometric templates that a later stage uses to predict coordinates. It generates the two-dimensional relationships between residues first and the full structure afterwards. TDFold.
IsoDDE, from Isomorphic Labs, is a significant proprietary counterpart. Its February 2026 report describes a broader drug-discovery system spanning complex prediction, binding properties, and molecular binding pockets. The published evaluations explain its claimed advances, but do not expose an independently reproducible architecture. Its system-level claims therefore have more public detail than its underlying model components. IsoDDE technical report.
Inverse folding and protein generation
Generative models differ in which parts of a molecule are variable. An inverse folding model holds geometry fixed and models sequence. A backbone generator models the protein’s main-chain shape. A joint model represents sequence and shape together, sometimes including side chains and other molecules. These are different modeling problems, even when the resulting systems share the word “design.”
ProteinMPNN, LigandMPNN, ESM-IF1, and PiFold
ProteinMPNN, from the University of Washington Institute for Protein Design, represents backbone geometry as a graph. A message-passing neural network exchanges information among neighboring residues; an autoregressive decoder models compatible amino-acid identities. Its contribution is a comparatively compact geometric model with substantial experimental validation. It helped make inverse folding a reliable, widely reused component of protein research. ProteinMPNN.
LigandMPNN adds information about surrounding atoms, including small molecules, nucleic acids, and metals. Additional message-passing components let those atoms influence residue representations. This addresses a specific limitation of backbone-only modeling: the chemically relevant environment may include much more than the protein main chain. LigandMPNN.
ESM-IF1 uses geometric vector perceptrons, which process scalar and vector features consistently with three-dimensional rotations, followed by a transformer. Its contribution combines geometric architecture with learning from millions of predicted structures. The original GVP-GNN work supplied a broader architectural foundation for this class of geometric reasoning. ESM-IF1, GVP.
PiFold emphasizes parallel prediction of amino-acid identities through learned residue features and graph layers. It explores the speed/accuracy tradeoff of avoiding a sequential decoder. Together, these families show why inverse folding is a distinct field: the supplied geometry changes both the information available to the model and the architecture that can exploit it. PiFold.
RFdiffusion, RFdiffusion2, and RFdiffusion3
The original RFdiffusion adapts RoseTTAFold into a denoising generator of protein backbones. Its geometric state uses residue frames: positions plus orientations. The contribution was a broadly useful structural generator supported by experimental results across multiple classes of protein architecture. RFdiffusion.
RFdiffusion All-Atom extends that lineage to molecular context using RoseTTAFold All-Atom’s mixed residue/atom representation. RFdiffusion2 develops a hybrid representation that can describe local geometry at atomic resolution. Despite its name, RFdiffusion2 is trained with a flow-matching objective. Its advance concerns more expressive geometric representation and conditioning, rather than a larger copy of the first model. RFdiffusion All-Atom, RFdiffusion2.
RFdiffusion3 moves to an atom-coordinate diffusion model with a transformer U-Net organization: information is compressed from atoms to residue-level tokens, processed, and expanded back to atomic updates. Its conditioning network is smaller than the large structural trunk of AF3 and omits expensive triangle operations. The contribution is a purpose-built generative architecture with broader atomic detail and improved efficiency. It is released through RosettaCommons Foundry. RFdiffusion3, Foundry.
The related ProteinGenerator takes another route: sequence-space diffusion coupled to RoseTTAFold’s structural reasoning. RFantibody and RFpeptides specialize structural generation for antibody geometry and cyclic-peptide topology. They are distinct branches with narrower molecular scope, rather than interchangeable names for the general RFdiffusion model. ProteinGenerator, RFantibody, RFpeptides.
Chroma: Generate:Biomedicines
Chroma uses a polymer-informed diffusion process, geometric neural networks, and a random graph architecture that limits the cost of modeling long-range relationships. It connects structural generation with sequence and side-chain modeling and supports multiple forms of conditioning. Its distinctive contribution is a scalable probabilistic treatment of protein structure with experimental validation. Chroma paper.
Generate:Biomedicines also develops proprietary models, datasets, and therapeutic programs. The published Chroma model is one part of the company’s current system. Its source code and model weights have different access terms. Chroma repository and licenses.
FrameDiff, FrameFlow, FoldFlow, and Genie
These research families model the geometry of generation explicitly. SE(3) names the space of three-dimensional rotations and translations; working with it lets a model reason about residue positions and orientations.
| Family | Architectural idea | Distinct contribution |
|---|---|---|
| FrameDiff | Diffusion over residue frames with geometric symmetry built into the model | A principled backbone generator that does not require initialization from a pretrained folding network. Paper |
| FrameFlow | Flow matching over structural frames | Studies efficient geometric generation through a learned continuous transformation. Paper |
| FoldFlow / FoldFlow-2 | Flow matching on rotations and translations; the successor adds sequence conditioning | Explores the geometry and statistical formulation of structural flows. This is a separate family from FrameFlow. FoldFlow, FoldFlow-2 |
| Genie / Genie 2, AlQuraishi lab | Coordinate noise with equivariant denoising using local frames | A scalable geometric diffusion lineage; Genie 2 expands structural training and conditional capabilities. Genie, Genie 2 |
| Genie 3 | Partial atomization: known atomic regions are represented as branched polymers, while unknown regions retain backbone representations | Combines finer chemical context with geometric equivariance. The May 2026 model is not direct full-sequence/all-atom co-generation. Paper, repository |
Joint sequence and atomic generation
The representational difficulty is that amino-acid identity determines which side-chain atoms exist, while those identities may themselves be unknown. Different joint models solve this in different ways.
Protpardelle represents possible side-chain states while coupling structural generation to sequence prediction. Its successor Protpardelle-1c strengthens conditional and multichain generation. MultiFlow instead couples a discrete flow over amino-acid identities with a continuous flow over backbone frames. The latter jointly models sequence and backbone, without implying that every side-chain atom is directly generated. Protpardelle, Protpardelle-1c, MultiFlow.
BoltzGen, released in 2025, extends the Boltz/AF3-style Pairformer and coordinate-denoising architecture to generation. It represents residue identity geometrically using virtual atoms, letting a continuous-coordinate model encode both sequence and shape. Its distinctive contribution is a common geometric representation for molecular prediction and generation, with publicly released code and weights. BoltzGen report, release.
BoltzProt-1, announced in June 2026, is a successor system combining an improved generator with BoltzPPI. BoltzPPI adds interaction and confidence prediction heads to Boltz-2-derived representations. That shifts assessment toward whether molecules actually interact, a different quantity from confidence in a predicted arrangement. BoltzProt-1 is proprietary; it does not inherit BoltzGen’s MIT access terms. BoltzProt-1 paper, company availability disclosure.
PXDesign, from ByteDance, extends the Protenix architecture into structural generation. Its PXDesign-d diffusion model exchanges information between atomic and residue representations, retaining backbone placeholders for unknown residues and atomic detail for molecular context. The wider suite also evaluates structure-based assessment. This work differs from raising a cofolder’s accuracy. PXDesign report.
SeedProteo, from ByteDance Seed, adapts a folding architecture to all-atom diffusion generation and feeds previous structural predictions into later generation steps. Its revised 2026 report includes experimental validation. It is a separate named generative effort associated with the SeedFold line; downloadable weights are not currently available. SeedProteo.
NVIDIA: Proteína, La-Proteina, and Proteina-Complexa
Proteína uses a scalable transformer with flow matching over backbone coordinates. Its original contribution emphasizes scale, computational efficiency, and learning the distribution of protein shapes. Proteína.
La-Proteina keeps alpha-carbon coordinates explicit, one central backbone atom per residue, while compressing sequence and the remaining atomic information into per-residue latent variables: learned numerical representations. A variational autoencoder learns that compression and reconstruction, while a flow model generates those coordinates and latent variables together. A decoder recovers the sequence and full atomic geometry. This hybrid representation reduces the need for the main generator to reason over every individual atom. La-Proteina.
Proteina-Complexa extends the approach to molecular complexes. Its contributions include joint sequence/structure generation, interaction training data, and research into additional computation during inference. The 2026 reports include a substantial experimental study. method report, experimental report.
NVIDIA’s BioNeMo is the wider software and model collection around work like this. It supports training and deployment of many biological models. BioNeMo is infrastructure, not a competing protein architecture. BioNeMo overview.
AlphaProteo and Latent Labs
AlphaProteo, from Google DeepMind, is a protein generation-and-assessment system with substantial experimental characterization. Its report describes a sequence/structure generator and a model for assessing likely experimental success. The distinctive evidence concerns the performance of that combined system. The public report does not justify labeling its undisclosed generator as an AF3 fine-tune or assigning it a detailed diffusion architecture. AlphaProteo technical report.
Latent-X, from Latent Labs, jointly models sequence and all-atom structure. Latent-X2 extends the work with particular attention to properties relevant to developing a molecule into a usable product. These developability properties include whether a molecule can be produced and remain sufficiently stable, alongside its intended biological interaction. The reports also examine selected immune-response measurements. Latent-X, Latent-X2.
Those experimentally measured properties are a substantive part of the company’s contribution. The low-level architecture is less publicly disclosed than that of several open families. Tested candidates also undergo computational selection, so measured success rates reflect the selected subset, not raw model output. Latent-Y is an agent built around the discovery platform, rather than another name for the base protein model. The X2 report also distinguishes the experimentally studied model from its later X2.1 serving version. Latent-X2 report, Latent-Y.
Other important sequence-generation approaches
EvoDiff, from Microsoft Research, performs discrete diffusion over amino-acid sequences. Its sequence branch uses a CARP-derived convolutional architecture; its alignment branch uses MSA Transformer. The contribution is generation over sequence variables without making coordinate prediction the central representation. It is distinct from the genomic Evo models. EvoDiff.
ProtBFN and AbBFN, from InstaDeep, use Bayesian flow networks. Their evolving state is a probability distribution over amino-acid choices. The network updates that uncertainty, providing another formulation of sequence generation alongside autoregressive and diffusion models. AbBFN specializes the approach to an antibody sequence domain. Bayesian flow networks for proteins.
Raygun, published in 2026, studies changes to protein sequence length. Its encoder–decoder uses a fixed-dimensional probabilistic representation derived from language-model embeddings. The distinctive idea is to represent proteins of different lengths in a common space. This addresses a limitation of models whose natural operation is to replace residues at a fixed set of positions. Raygun.
Protein motion and conformational ensembles
A protein can occupy several structural states. Predicting those states and their relative populations is a different problem from predicting a single well-resolved structure. An ensemble is a distribution of conformations, and its usefulness depends on more than producing visually diverse shapes. Molecular dynamics simulates atomic motion; kinetics concerns the rates of transitions between states.
one possible shape
AlphaFlow and ESMFlow adapt AlphaFold and ESMFold through flow matching. They learn sequence-conditioned distributions of structures, including training on ensembles from molecular dynamics. Their contribution is to repurpose accurate single-structure networks into models of conformational variation. AlphaFlow and ESMFlow.
BioEmu, from Microsoft Research and collaborators, uses a geometry-aware denoising diffusion model conditioned on protein information. Its training combines static structures, molecular-dynamics simulations, and experimental stability measurements. Its distinctive goal is an equilibrium ensemble: both the range of possible states and their relative probabilities. The work evaluates thermodynamic properties as well as geometric diversity. BioEmu paper.
BioEmu’s generated structures are not a time-ordered molecular trajectory, because the model samples equilibrium states rather than simulating dynamics. BioEmu model card.
ESMDynamic, a 2026 model, addresses a more compressed output: dynamic contact maps from a single sequence. It estimates how contacts vary rather than generating a complete atomic trajectory. A model can predict a summary of motion when the question does not need full conformational generation. ESMDynamic.
Specialist models and historical foundations
Antibody and immune-receptor models
Antibodies have characteristic sequence organization and paired-chain relationships. Conventional antibodies pair a heavy chain with a light chain; features of their interaction can matter beyond the properties of either alone. A specialized model can learn that distribution more directly than a general model.
| Family | Architecture or specialization | Distinct contribution |
|---|---|---|
| AntiBERTa / AntiBERTa2 | Antibody-focused masked transformers; AntiBERTa2 uses rotary position information, and AntiBERTa2-CSSP adds sequence–structure alignment | Representations adapted to antibody sequence organization, with structural supervision extending the sequence model. AntiBERTa, AntiBERTa2-CSSP |
| AntiBERTy | Antibody-specific masked language model | A domain-specific sequence representation, also used by IgFold. Author repository |
| AbLang / AbLang2 | Antibody sequence encoders; AbLang2 addresses inherited-sequence bias | Reduces the tendency to overemphasize frequently observed inherited, or germline, residues. AbLang2 |
| IgBert / IgT5 | BERT- and T5-based learning from paired and unpaired antibody chains | Explicit treatment of chain pairing within large antibody-language-model training. Paper |
| BALM-paired / BALM-unpaired, Burbach and Briney | Matched antibody-language-model experiments with paired or unpaired sequences | Isolates the contribution of native heavy/light-chain pairing. These models are distinct from other projects sharing the BALM acronym. Paper |
| IgFold | AntiBERTy embeddings plus geometric graph/attention modules | Rapid prediction of antibody and nanobody structures. Paper |
| ImmuneBuilder | Specialized geometric predictors with ensembles | ABodyBuilder2, NanoBodyBuilder2, and TCRBuilder2 address antibodies, nanobodies, and T-cell receptors. Paper |
| ABodyBuilder3 | ImmuneBuilder-derived prediction with language-model features | Improves variable-region geometry and predicted structural accuracy. Paper |
| AntiFold | Antibody adaptation of ESM-IF1’s geometric encoder and transformer | Domain-specific inverse folding. Paper |
| tFold, Tencent | Interaction-aware ESM-PPI representations coupled to geometric folding modules | Multichain language information for specialized tFold-Ab, tFold-Ag, and tFold-TCR models. Paper, repository |
| DiffAb | Joint modeling of sequence, residue coordinates, and orientations through diffusion | An influential early antibody sequence/structure generator. Paper |
| IgGM / IgGM2, Tencent and collaborators | IgGM combines language features, interaction encoding, and consistency-based sequence/structure generation; IgGM2 extends to all-atom immune receptors | Specialized prediction and co-generation. IgGM2 is a July 2026 preprint. IgGM, IgGM2 |
MINT addresses interactions more generally. It starts from ESM-2 and adds cross-chain attention, allowing each sequence’s representation to depend on other interacting sequences. Its contribution is an interaction-aware language representation. MINT.
Variant effects and older representation models
DeepSequence and EVE learn family-specific evolutionary distributions with variational autoencoders. These models compress related sequences into a latent representation and reconstruct their distributions. They model dependencies among residues and use evolutionary evidence to assess sequence variation. They helped establish that useful protein modeling could come from a strong family-specific statistical model. DeepSequence, EVE.
AlphaMissense uses an AlphaFold-derived architecture for missense-variant effect prediction, drawing on evolutionary and structural information. Its output concerns variation in human proteins. AlphaMissense.
Two earlier representation models are also important. UniRep uses a multiplicative recurrent neural network, showing the value of transferable protein representations before transformers dominated the field. ProteinBERT combines local and global representations, convolution, and global attention while learning from sequences and functional annotations. ProteinBERT is a different model from ProtTrans’s ProtBERT. UniRep, ProteinBERT.
TAPE helped establish common transfer-learning evaluations rather than one definitive model architecture. Earlier structural-statistics approaches such as dTERMen also show that protein design has a longer history than large language models: recurring structural fragments can supply learned sequence preferences without a transformer. TAPE, dTERMen.
Companies, experimental data, and platforms
Some companies publish a named base model with downloadable weights. Others publish a multi-model system or experimental results while retaining most internals. Their contribution may come from proprietary measurements, assessment models, production software, or learning from repeated experiments. The comparisons below describe what each company has disclosed.
The major named systems from Biohub, Google DeepMind, Isomorphic Labs, Chai, Boltz, Profluent, BioMap, Generate:Biomedicines, Latent Labs, NVIDIA, Iambic, Genesis, GenBio, ByteDance, and Baidu are covered above. Several additional companies are important to the experimental and product side of the field.
Cradle and CRADLE-1
Cradle combines generative protein models with predictors learned from laboratory measurements. The 2026 CRADLE-1 report describes pretrained models, adaptation to relevant evolutionary data, and further learning from experimental feedback. Its contribution is a system that connects general protein priors to the properties actually measured in a project. CRADLE-1, Cradle’s technical discussion.
This makes data handling and multi-property learning part of the product’s value, though its published description does not reduce to one fully disclosed checkpoint.
Nabla Bio and JAM
Nabla’s Joint Atomic Modeling, or JAM, combines molecular generation with a model for assessing experimental viability. Public sequence and structural information is connected to proprietary biologics measurements. Its distinctive emphasis is a connection between atomic modeling and properties relevant to real molecules, including developability. JAM technical report, platform description.
The generation/assessment split is publicly described, but a detailed layer-level architecture and downloadable weights are not available.
Absci: IgDesign and Origin-1
IgDesign is Absci’s published antibody inverse-folding work. Its later Origin-1 platform contains multiple components: AbsciDiff for all-atom structural diffusion, IgDesign2 combining geometric features with a paired-antibody language model, and AbsciBind for assessment. Proprietary experimental data are central to the company’s approach. IgDesign, Origin-1 architecture disclosure.
The distinctive contribution is the integration of antibody sequence, geometry, and experimental assessment. Origin-1 is a system of models; IgDesign2 develops the earlier inverse-folding work within that larger system.
Prescient Design and Genentech
Prescient Design is part of Genentech. Its published LaMBO combines a denoising autoencoder with an uncertainty-aware property predictor based on Gaussian processes. LaMBO-2 studies guided discrete diffusion. These methods connect generative sequence modeling with learning from limited experimental data and representing uncertainty about properties. Genentech’s Prescient Design, LaMBO, LaMBO-2.
Their contribution is the relationship between generation and experimental learning, beyond the choice of a sequence-model backbone.
Infrastructure and adjacent fields
ColabFold combines efficient evolutionary search and accessible folding software. Its contribution is usability and computation, rather than a separate foundation-model architecture. ModelGenerator, from GenBio, similarly supports adapting and evaluating biological models. ColabFold, ModelGenerator.
The same scope distinction applies to genomic models. Evo and Evo 2, associated with Arc Institute and collaborators, model nucleotide sequences. Their context can include protein-coding information, but their primary modeled object is genetic sequence. They are genomic models, not amino-acid language models (and are unrelated to Microsoft’s EvoDiff). Evo 2.
Comparing the contributions fairly
The survey suggests several recurring sources of progress. Large sequence datasets provide broad statistical priors. Related proteins supply specific evolutionary evidence. Structures provide geometry. Experimental measurements teach properties that cannot be established by sequence plausibility or structural confidence alone. Architectures differ in how directly they represent and combine these sources.
Input information and evaluation
Much of this field shares a data foundation. UniProt collects protein sequences and annotations; UniRef groups related sequences into clusters at different similarity levels. The Protein Data Bank archives experimentally determined structures. Predicted-structure collections add far more geometric examples, with the biases of the models that produced them. A count of raw sequences, cluster representatives, or training tokens therefore describes different quantities. UniRef, Protein Data Bank.
A model supplied with an MSA or a structure has evidence unavailable to a sequence-only model. A larger training dataset or a later data cutoff can also change a result without any architectural improvement. Comparisons should therefore name the available inputs, training-data cutoff, and evaluation task.
ProteinGym supplies common benchmarks for variant effects, including substitutions and insertions/deletions. Those tasks are useful for assessing sequence representations and evolutionary modeling; they do not rank every cofolder or generative system. Likewise, a structural benchmark measures geometry, and a laboratory study measures specific biological outcomes. ProteinGym.
For experimental claims, the denominator matters. The success rate among computationally selected candidates is different from the success rate among all generated candidates. The number and diversity of tested systems, the definition of success, and the measurement method also matter. For these reasons, the studies above cannot fairly be ranked in a single table.
Parameters, representations, and computation
Parameter count mixes several quantities: total model capacity, active parameters in a sparse model, and parameters in components that may be discarded after training. It also says little about how much work a model does during generation. A structural denoiser may run repeatedly; a retrieval model may process a large context; an encoder may produce its output in one pass.
The corresponding architectural comparison is more informative. ProGen3 and AIDO.Protein explore sparse capacity. PoET and E1 emphasize retrieved context. SaProt and DPLM-2 make structural representation explicit. AF3 and its relatives invest in pairwise geometry and atom generation. La-Proteina compresses atomic detail into learned variables. CARP and ProtMamba explore different ways to process long sequences. Each choice changes what the model can represent efficiently.
Access and reproducibility
“Open” can refer to inference code, model weights, training code, training data, or commercial-use rights. These should be checked separately and for the exact release.
| Release checked | What the current disclosure establishes |
|---|---|
| ESMC and ESMFold2 | MIT access to current released weights; larger ESM3 offerings have their own delivery arrangements. Biohub documentation |
| AlphaFold3 | Code changed to Apache-2.0 in June 2026; separate parameter-use restrictions remain. Releases, parameter policy |
| Chai-1 | Apache-2.0 code and weights; this does not establish access to Chai-2/3. Repository |
| Boltz-2 and BoltzGen | MIT model/code releases; BoltzProt-1 is a separate proprietary product. Boltz, BoltzGen, successor terms |
| ProGen3 | Apache-2.0 code, separately noncommercial weights, and released checkpoints smaller than the paper’s flagship. Repository |
| Ankh3 | Downloadable models with noncommercial share-alike weight terms. Model card |
| OpenFold3 / OpenBind-0 | Public training-oriented releases; OpenBind-0 explicitly releases code, data, weights, and training materials under Apache-2.0. OpenBind-0 |
A new architecture, a better experimental dataset, an improved assessment model, and a fully reproducible implementation are all meaningful contributions. The papers cited here establish their value on different questions.