Protein Design and Engineering

Published

August 22, 2026

Protein design inverts the structure-prediction problem. Instead of asking what shape a sequence adopts, design asks which sequence (or scaffold, or assembly) satisfies a target constraint. The current stack combines diffusion-based backbone generation, sequence design with inverse folding, multimodal generative models that handle structure and sequence jointly, and active-learning loops that use experiments to steer the next round. The 2024 Nobel Prize in Chemistry recognized this work alongside structure prediction. None of it removes the validation campaign at the bench: what changes is the rate at which design hypotheses can be produced and triaged.

Learning Objectives

Use this chapter to:

  • Design protein backbones, sequences, binders, enzymes, and functional proteins that can be built and tested experimentally.
  • A design is not a result until expression, folding, binding, activity, stability, and specificity are measured in the intended context.

Prerequisites: Protein Structure Prediction recommended; Foundation Models for Biology for the ESM-3, ProGen2, and EvoDiff sections.

Summary: Design protein backbones, sequences, binders, enzymes, and functional proteins that can be built and tested experimentally. Backbone generation and sequence design are now credible for selected tasks; general function design remains far harder than producing plausible structures.

Key point: A design is not a result until expression, folding, binding, activity, stability, and specificity are measured in the intended context. Open question: whether designs keep working across target classes after expression, function, specificity, and developability are measured.

Bottom line: Protein design connects structure prediction to therapeutics, enzymes, synthetic biology, vaccines, biomaterials, and autonomous design-build-test loops.

Field Guide

What is this field trying to solve? Design protein backbones, sequences, binders, enzymes, and functional proteins that can be built and tested experimentally.

What is the core idea? A design is not a result until expression, folding, binding, activity, stability, and specificity are measured in the intended context.

What is the current state of the field? Backbone generation and sequence design are now credible for selected tasks; general function design remains far harder than producing plausible structures.

What do we know, and what remains open? Known reference points include RFdiffusion, ProteinMPNN, LigandMPNN, ESM3, Chroma, ProGen2, EvoDiff, EVOLVEpro, COMPSS, Rosetta, PDB, ProteinGym, and wet-lab design campaigns. What remains open is whether designs keep working across target classes after expression, function, specificity, and developability are measured.

Why does this matter? Protein design connects structure prediction to therapeutics, enzymes, synthetic biology, vaccines, biomaterials, and autonomous design-build-test loops.

Introduction

The 2024 Nobel Prize in Chemistry recognised computational protein design alongside structure prediction (Nobel Prize, 2024). The Baker laboratory’s methodological lineage includes Rosetta (Leaver-Fay et al., 2011), RFdiffusion (Watson et al., 2023), ProteinMPNN (Dauparas et al., 2022), LigandMPNN (Dauparas et al., 2025), and RFantibody (Bennett et al., 2026). These tools expand the set of proteins that can be proposed and experimentally tested; they do not remove the experimental denominator.

The change is real but easy to overstate. AI protein design has produced experimentally validated binders, miniproteins, enzymes, and antibodies at rates that would have been implausible a decade ago. AI protein design has not produced general functional design from natural-language prompts, reliable design of complex multi-domain biologics without iteration, or removal of the developability and immunogenicity steps that follow any biologic candidate from bench to clinic.

The chapter is organised around four questions a working researcher asks:

  1. Which generative system should I use for this design task?
  2. What does the published evidence actually support?
  3. What experimental validation does a designed candidate require?
  4. Where do current systems still fail predictably?

Demonstrated capability

Backbone generation: RFdiffusion

RFdiffusion (Watson et al., 2023) is a denoising diffusion model fine-tuned from RoseTTAFold weights. It generates three-dimensional protein backbones conditioned on user-specified constraints: motif scaffolding (place a functional motif within a generated scaffold), symmetric oligomer design, binder design against a target surface, and unconditional generation.

The Nature paper reported experimental validation across monomer design, oligomer design, motif scaffolding, and binder design against selected targets. The reported success depended on the task, filtering, and experimental definition. Public code and weights made the method inspectable and reusable, but later use does not transfer the original validation to every target class.

What RFdiffusion does not do: predict sequences. It produces a backbone without final amino-acid identities; sequence design is a separate downstream step.

RFdiffusion3 extends the lineage to an all-atom diffusion model that generates protein structures in contexts containing ligands, nucleic acids, and other non-protein atoms. In a developer preprint, the authors reported improved in-silico benchmark performance and approximately tenfold faster generation than RFdiffusion2 for typical protein lengths; experimental validation covered two selected design tasks, so the evidence does not establish general performance across target classes (Butcher et al., 2025, preprint). The developers released inference code, training code, and model weights.

Sequence design: ProteinMPNN and LigandMPNN

ProteinMPNN (Dauparas et al., 2022) is a graph neural network that designs amino-acid sequences for a fixed backbone. Trained on PDB structures, it predicts the residue at each position conditioned on the surrounding backbone geometry and, optionally, neighbouring residue identities. The Science paper reported sequence-recovery results and experimental tests across selected design tasks. It is a widely adopted inverse-folding tool, but task-specific baselines and assays remain necessary.

LigandMPNN (Dauparas et al., 2025) extends ProteinMPNN to handle atomic ligand context: small molecules, metals, nucleic acids, and modified residues. It is a relevant sequence-design option when ligand or non-protein atomic context matters, subject to the target class and evaluation.

A common composition pattern is RFdiffusion → ProteinMPNN for protein-only contexts, or RFdiffusion → LigandMPNN for ligand-aware design. Their model scores can support triage, but should not be described as calibrated probabilities without a task-specific calibration analysis.

Joint and language-model approaches

Several systems generate sequence and structure jointly or skip explicit structure altogether:

  • Chroma (Ingraham et al., 2023) is a programmable generative model for protein space that supports conditioning on symmetry, shape, sub-structure, and natural-language-like constraints. The Nature paper demonstrated experimental expression of generated designs.
  • ESM3 (Hayes et al., 2025) from EvolutionaryScale is a multimodal protein language model that models sequence, structure, and function tokens in one architecture. The Science paper included experimental validation of esmGFP, a designed green fluorescent protein with substantial sequence divergence from natural GFPs.
  • ProGen2 (Nijkamp et al., 2023) is an autoregressive protein language model with the same scaling story as natural-language LMs, useful for sequence-only generation conditioned on functional family.
  • EvoDiff applies diffusion directly to amino-acid sequences without an explicit structural step (Alamdari et al., 2023, preprint). The cited evidence remained a preprint at the August 12, 2026 review cutoff.
  • ProteinGenerator (Lisanza et al., 2025) uses RoseTTAFold sequence-space diffusion for multistate and functional design. Its practical contribution is conditioning sequence generation on structural and functional constraints, including experimentally guided design settings.
  • AlphaProteo (Zambaldi et al., 2024, preprint) is DeepMind’s reported binder generator. Hit rates in the preprint are notably high; the work is unpublished, the code is unreleased, and external reproducibility is not yet established. Treat as a preprint with the corresponding evidence weight.
  • Claude binder campaign (Anthropic, August 2026) is a company-reported experimental result. Anthropic stated that Claude (Opus 4.8 and Mythos Preview), given a human-expert prompt and access to public specialist design and folding models, produced de novo binders against 14 of 15 targets. Adaptyv Bio and Twist Bioscience independently produced and tested the designs. Reported hit rates were 22 to 35%, against a field-typical 10 to 15%; some designs bound several times more tightly than the best published de novo binder for that target. Reported targets included PD-L1, TREM2, VEGF-A, IL-7Rα, latent GDF-8, RBX1, and Nipah G. Designs and assay data were released. The record is a company announcement, not a peer-reviewed paper, and Anthropic stated that the binders are not drugs.

The general capability claim is established: large generative models can produce sequences that fold and function. The boundary conditions, which tasks, which targets, which success rates under independent reproduction, are still being mapped.

Experimental calibration and active learning

The highest-value design loop uses wet-lab data to steer the next round. COMPSS showed that computational filters for generated enzymes become more useful when calibrated against experimental activity rather than treated as universal in-silico scores (Johnson et al., 2025). EVOLVEpro used protein language embeddings and few-shot active learning to guide protein activity optimization across RNA production, genome-editing, and antibody-binding tasks (Jiang et al., 2025).

These papers sharpen the operational rule: AI protein design is strongest when the model proposes, the assay measures, and the next model round learns from measured failures. A design pipeline that lacks experimental feedback is a generator, not an engineering system.

Antibody-specific design: RFantibody

Antibody design is its own subfield because the immunoglobulin fold, the hypervariable loop structure, and the affinity-maturation history make antibodies systematically different from general proteins. Generic protein-design systems are weak on antibody loops.

RFantibody (Bennett et al., 2026) extends RFdiffusion to antibody scaffolds and reports atomically accurate de novo antibody designs validated by experimental binding and structural characterisation. The Nature paper supersedes the earlier bioRxiv preprint and is the current peer-reviewed reference for AI-enabled antibody design.

Complementary systems for antibody-specific tasks (not de novo design):

For developability, immunogenicity prediction, and manufacturability, see the Antibody and Biologic Design chapter; design quality and developability are different evaluations and require different evidence.

Germinal adds evidence for epitope-targeted antibody generation. Across four protein targets, the study tested 43 to 101 designs per antigen and reported functional binders, including selected nanomolar-affinity designs and structural or mutational validation (Mille-Fragoso et al., 2026). This is evidence of experimentally connected design on four targets. It does not establish therapeutic efficacy, broad target-class generalization, or manufacturability at scale, and several authors disclosed a related provisional patent application.

Iterative small-molecule binder design

Neural iterative selection-expansion combines sequence design and protein-ligand co-structure prediction in a repeated computational selection loop. Fry and colleagues reported binding for all four tested exatecan designs and five of six tested apixaban designs, with selected high-affinity results (Fry et al., 2026). Those denominators make the experiment informative. The scope remains two small molecules and selected scaffolds. Binding does not establish in vivo function, developability, safety, pharmacokinetics, or therapeutic utility, and the authors disclosed a related provisional patent application.

Evidence anchor summary

Evidence Anchor What It Supports Practical Constraint
RFdiffusion De novo backbone generation under constraints Backbone only; sequence is a separate step
RFdiffusion3 All-atom biomolecular interaction design Developer preprint with selected experimental tasks; general performance is not established
ProteinMPNN Inverse-folding sequence design for fixed backbones Quality depends on backbone realism and target distribution
LigandMPNN Ligand- and modification-aware sequence design Extends ProteinMPNN; supersede for ligand-aware tasks
Chroma Programmable joint structure-and-sequence generation Conditioning interface differs from RFdiffusion stack
ESM3 Multi-modal sequence + structure + function generation Compute scale; reported esmGFP validation but generalisation under external reproduction is still being mapped
ProGen2 Autoregressive sequence generation, family-conditioned Sequence-only; structural realism is downstream
EvoDiff Sequence-space diffusion at evolutionary scale Preprint; no peer-reviewed venue at this writing
ProteinGenerator Sequence-space diffusion with structural and functional conditioning Task-specific validation determines value
EVOLVEpro Few-shot active learning for activity optimization Requires a measurable assay and round-by-round experimental data
COMPSS Calibration of computational filters for generated enzymes In-silico scores alone are insufficient
RFantibody De novo antibody design, experimentally validated Antibody-specific; not a general design tool
Germinal Epitope-targeted antibody generation on four tested targets Binding evidence does not establish therapeutic development
Neural iterative selection-expansion Experimentally tested small-molecule binder design on two targets Target diversity and downstream validation remain limited
AlphaProteo Reported strong binder generation Restricted release; code unavailable; cite as preprint
Claude binder campaign Company-reported experimental binders on 14 of 15 targets Company announcement, not peer-reviewed; binders are not drugs

Theoretical capability

Several capabilities are plausible given current methods but not yet routine.

Natural-language design specification. A pipeline that takes “design a thermostable, low-immunogenicity binder to PD-1 that does not aggregate at 100 mg/mL” and produces a candidate set would compose existing systems with structured specification languages and developability scorers. Components exist; integration at production quality does not.

Reliable enzyme design. Enzymes require precise active-site geometry, catalytic mechanism, and substrate access. Backbone design plus inverse folding produces candidates; the fraction that show measurable catalytic activity remains low and highly target-dependent. The community has produced individual successes (de novo retro-aldol, Diels-Alder, Kemp eliminase work going back to Baker-lab studies in the 2000s and 2010s) but no general design recipe.

The early enzyme-design record is the right calibration point. Computationally designed retro-aldolases and Kemp eliminases demonstrated that active-site geometry can be designed into new scaffolds, but the reported catalysts required experimental screening and optimization and did not establish a general recipe for arbitrary enzymatic function (Jiang et al., 2008; Röthlisberger et al., 2008).

Allosteric and switchable proteins. Conformational-cycle proteins (kinases, GPCRs, transporters) require sequence designs that adopt multiple states. Current generative systems predict one geometry; the design of an explicit conformational landscape is an open problem.

Cell-context design. Proteins that depend on cellular context, including localisation, post-translational modification, and regulation, are often designed in cell-free biophysical assays. The cellular-fidelity gap remains.

Closed-loop autonomous design. Coupling generative design, automated DNA synthesis, robotic expression and characterisation, and a model that updates from each round (covered in detail in Self-Driving Laboratories and Agentic Science Workflows) is feasible at limited scale. Production deployment that meaningfully accelerates real programs is still emerging.

Beyond current capability

A few framing claims are not supported by current evidence.

General-purpose functional design. A prompt-to-protein pipeline that reliably produces active, manufacturable, low-immunogenicity biologics for arbitrary therapeutic targets without iterative assay cycles is not demonstrated. The recipe-of-recipes for biological function does not exist.

Removal of developability and immunogenicity assessment. AI design can accelerate candidate generation; it does not establish acceptable aggregation, viscosity, polyspecificity, immunogenicity, or manufacturability. Therapeutic programs still need risk-proportionate computational and experimental assessment of these properties. AI tools help triage; they do not replace experimental characterization.

Reliable design across all protein families. Designed proteins have systematically different biophysical properties from natural ones (different stability distributions, different aggregation behaviour, lower immunogenicity in some cases, higher in others). Some families (small globular folds, helical bundles) work well; others (membrane proteins, intrinsically disordered regions, large complex assemblies) remain weak.

The general rule: a design system that promises any of these without reporting independent experimental hit rates is making a claim that the current state of the field does not support.

Evidence that would change the assessment

Protein-design claims become more promising with denominator-complete reports from independent teams across target classes: candidates generated, ordered, expressed, purified, folded, active, specific, stable, and developable. Stronger evidence would show closed-loop design outperforming expert and directed-evolution baselines under matched experimental budgets, not only producing attractive in-silico scores. Therapeutic claims would need manufacturability, immunogenicity, formulation, and pharmacology evidence tied to the same candidates that succeeded in the primary assay.

Implications for research and program decisions

For working researchers running a design programme:

  • Compose the stack. RFdiffusion for backbones → ProteinMPNN or LigandMPNN for sequences → AlphaFold 2/3 or Boltz / Chai for in-silico structural triage. For antibody work, use RFantibody for de novo and IgFold or AlphaFold 3 for structure prediction of natural antibodies.
  • Filter in silico before bench work. Per-residue pLDDT distribution, sequence recovery against natural homologs, predicted developability scores, basic biophysical filters (charge, hydrophobicity, predicted solubility). The cost of expressing and characterising candidates dominates; aggressive in-silico filtering is well-paid effort.
  • Define success as an assay result. A designed protein that scores well on every in-silico metric and does not bind in the assay is not a successful design; it is a failed in-silico-to-bench transfer. Report assay results, not scores.
  • Build the negative set. Designs that fail teach as much as designs that succeed. Track and analyse failure modes (no expression, no fold, no binding, off-target binding, aggregation, low stability) so each round biases the next.
  • Use developability tools as triage, not validation. Predicted immunogenicity (NetMHC-class predictors), aggregation propensity (SAP, AggreScan), polyspecificity scores: useful for ranking, not for clearance. Therapeutic candidates need experimental developability work regardless.
  • Watch for restricted releases. AlphaProteo’s reported hit rates are notable; the absence of open code and independent reproduction limits how much weight to put on the numbers. Treat as a preprint until the field reproduces the claims.
  • Cross-check with structure prediction. Predicted designs benefit from the confidence-interpretation discipline: pLDDT, PAE, ipTM: applied the same way as for natural proteins.
  • Document everything before bench work. Sequence, scaffold specification, generation parameters, in-silico scores, software version, training data range. Designed-protein reproducibility lives in those records.