Protein structure prediction & protein design

Knowing a protein's 3D shape used to take months of crystallography or cryo-EM. Deep-learning models now predict structures and complexes from sequence in minutes, and generative models design new proteins (binders, enzymes, antibodies) that are then synthesised and tested in the lab. The benefit is fewer wet-lab cycles per candidate, not the elimination of experiments.

Problem & value

• Today: structure determination (X-ray, cryo-EM, NMR) is slow and costly; classical protein engineering relies on directed evolution and large screening campaigns with low hit rates.
• AI value: structure hypotheses for targets without experimental structures, in-silico screening of thousands of designs before ordering DNA, and faster design-build-test-learn cycles.
• Evidence: AlphaFold 2 reached near-experimental accuracy in CASP14; the 2024 Nobel Prize in Chemistry honoured computational protein design (Baker) and structure prediction (Hassabis, Jumper) [1][2]. The AlphaFold Database (hosted by EMBL-EBI, Hinxton/UK, with EU research funding ties) provides predicted structures for over 200 million proteins as a free resource [3]. Hit-rate claims for de-novo binders vary by target and vendor; treat vendor numbers as unverified until reproduced on your target.

Variants

• Structure prediction only: understanding targets, mutations, complexes (protein-protein, protein-ligand, protein-nucleic acid).
• De-novo binder / mini-protein design: against a target epitope (therapeutics, diagnostics, biosensors).
• Antibody and nanobody design/optimisation: affinity maturation, developability filtering.
• Enzyme engineering: stability, activity, specificity for industrial biotech, food, detergents, bioremediation.
• Academic / public research: free web servers and open models, no commercial pipeline.
• Private use: none meaningful beyond education (browsing AlphaFold DB).

How it works

1. target definition (sequence, structure, epitope, function)
2. structure prediction / complex modelling (AF3-class co-folding)
3. generative design (backbone diffusion
4. sequence design) or language-model proposals
5. in-silico filtering (re-prediction, confidence ipTM/pLDDT, energy, developability, immunogenicity)
6. biosecurity & sequence screening
7. gene synthesis
8. wet-lab assays (binding, activity, stability)
9. results feed back into the next design round (active learning)

• Structure predictors (AlphaFold 3, Boltz, Chai, OpenFold3) are diffusion-based "co-folding" models predicting proteins together with ligands, nucleic acids and ions [4][5].
• Generative design: RFdiffusion generates backbones, ProteinMPNN designs sequences for them; protein language models (ESM family) propose and score sequences [6][7].
• Agents/LLMs orchestrate tool chains, summarise literature and write pipeline code; they do not replace the structural models.
• Human-in-the-loop: scientists define targets, review filtered candidates, decide what is synthesised, and interpret assay data. Every design must be experimentally validated.

Data preparation

Typical inputs are protein and target sequences, PDB/mmCIF structures, multiple sequence alignments and wet-lab assay results. In an SMB-sized biotech or institute they live in electronic lab notebooks (ELN), LIMS, shared drives, Excel sheets and sometimes only in PDFs of publications or instrument exports.

Typical gaps: assay data in free-form spreadsheets without units or consistent target IDs, negative results never recorded, sequence variants named inconsistently, structures from different sources without provenance, and no clear owner for each data set. Access rights are often too broad for unpublished sequences.

Preparation steps: 1. Inventory sources and owners, then map what each design loop needs (see PREP-001, PREP-002, PREP-004). 2. Convert spreadsheets into tidy tables with fixed columns for construct ID, sequence, assay, unit and date (PREP-011). 3. Agree on a naming scheme and a small controlled vocabulary for targets and assays (PREP-012, PREP-019). 4. Extract machine-readable data from papers and instrument PDFs (PREP-006, DOC-014) and keep provenance. 5. Build a held-out test set with the lab scientists (PREP-018). 6. Restrict access to unpublished sequences (PREP-024).

Effort share: high - roughly half of the project effort. Predictors work out of the box, but design loops only improve when assay results come back in clean, linked form. See also GUIDE-003.

Implementation paths

Which path for whom

• Enterprise: Path C - sequences and designs are core IP, and open co-folding and design models (Boltz, OpenFold3, RFdiffusion, ProteinMPNN) run on an own GPU cluster (HW3) without sending targets to a third party.
• Developer: Path C - the open models are well documented and easy to wrap into pipelines; cloud APIs (Path B) are a good bridge for bursts.
• Public: Path C - research institutes can run the open stack on university or sovereign HPC/cloud, which also keeps results reproducible and publishable.

Path A — Buy (off-the-shelf)

• Free web tools: AlphaFold Server (non-commercial use, limited jobs/day) and the AlphaFold DB [3][4]; ColabFold for quick runs.
• Commercial platforms: Cradle (NL, EU; ML-guided protein engineering), Generate:Biomedicines, Profluent, Absci, Isomorphic Labs (partnership-based), NVIDIA BioNeMo-based offerings, and Schrodinger / Dassault BIOVIA / Chai Discovery / Latent Labs (UK) tooling.
• Choose A when: small team without ML engineers, a defined engineering goal (e.g. enzyme stability) and budget for a platform plus wet-lab partner or CRO.

Path B — Build on cloud APIs

• Hosted models: NVIDIA BioNeMo / NIM microservices (OpenFold, Boltz, RFdiffusion, ProteinMPNN), AWS, Azure and Google Cloud GPU instances in EU regions, EvolutionaryScale ESM models via API/open weights [7].
• Orchestration: Nextflow or Snakemake plus a workflow engine; LLM agent layer for planning and reporting.
• Choose B when: you need custom pipelines and elastic GPU capacity, with sequences possibly being IP-sensitive (check data-processing terms; prefer EU regions and no-training clauses).

Path C — Self-host (open source)

• Structure prediction: Boltz-1/Boltz-2 (MIT licence, includes binding-affinity prediction), Chai-1 (Apache 2.0), OpenFold3 (OpenFold consortium, Apache 2.0), ESMFold, ColabFold [5][8]. AlphaFold 3 code and weights are available only under non-commercial academic terms [4].
• Design: RFdiffusion, ProteinMPNN, BindCraft, ESM3-open (small, non-commercial licence terms apply).
• Hardware: HW3 (one 48-141 GB GPU) for complexes of typical size and design campaigns; HW2 for small monomers; HW4 for large-scale campaigns. Check each model's licence before commercial use.
• Ops effort: MSA databases need several TB of storage; plan an ML/bioinformatics engineer.

Path D — Custom training

Fine-tune or train when you own proprietary assay data (thousands of labelled variants) for a specific target class such as antibodies or enzymes. Typically a property-prediction head or a fine-tuned language model on top of public weights; full pretraining of a structure model needs HW5 and is rarely justified.

State of the art (2026)

• Co-folding models (AlphaFold 3, Boltz-2, Chai-1, OpenFold3) predict biomolecular complexes and now estimate binding affinity (Boltz-2) [4][5]. Open reproductions closed much of the gap to AlphaFold 3, with permissive licences enabling commercial use.
• De-novo design: diffusion-based design has produced functional binders, enzymes and antibody fragments with experimentally confirmed activity; antibody-specific generative models are advancing quickly in industry and academia [6].
• Protein language models (ESM3 and successors) generate and score sequences, including conditioning on function [7].
• Limits and failure modes: confidence scores are not the same as functional success; models struggle with conformational dynamics, intrinsically disordered regions, novel folds far from training data, allosteric effects, and antibody-antigen docking; memorisation of training-set structures can inflate benchmarks. Affinity predictions remain much less reliable than structure predictions. Always validate in the lab.

Shared resources & combinations

• Same stack (scientific-fm, HW3 GPU): SCI-002 virtual screening and molecule generation uses the same co-folding and docking models; SCI-017 target identification supplies targets; SCI-021 DNA/RNA sequence design covers the gene/construct side; SCI-009 genomics provides variants whose structural effect is assessed.
• Lab loop: SCI-007 self-driving labs close the design-build-test cycle; SCI-011 AI-scientist agents orchestrate hypotheses and pipelines.
• Shared infrastructure: GPU cluster, MSA/sequence databases, workflow engine, ELN/LIMS.

Effort, cost & KPIs

• Path A: weeks to onboard; platform subscription or project fees plus wet-lab costs, which usually dominate.
• Path B: 2-4 months to a reliable pipeline; GPU hours are minor next to synthesis and assays.
• Path C: 3-6 months; GPU server (HW3), storage, ML engineer.
• Path D: 6-12+ months; needs substantial proprietary assay data.
• KPIs: experimental hit rate per designed candidate, number of design-build-test cycles to reach target potency, correlation of predicted vs. measured stability/affinity, cost per validated candidate, time from target to first validated binder.

People & change

Involve early: management (budget, IP policy) and the department lead of R&D define the target and success criteria. Data owners (lab, LIMS/ELN) and IT release data and GPU resources. IT security and legal/compliance clarify IP protection, export control and biosecurity screening of designed sequences before synthesis. End users are the computational and wet-lab scientists; involve them from the first pilot, not at rollout.

Works council: depends. The models process no personal data and do not evaluate staff. Co-determination under §87(1) no. 6 BetrVG becomes relevant only if the platform logs individual usage (job queues, user-level logs) in a way suited to monitor performance - then agree on logging limits. In any case inform the works council in good time under §90(1) no. 3 BetrVG; public institutes involve the Personalrat. Details in GUIDE-004.

Data protection: usually no personal data, so no DPIA; check user accounts and logs with the DPO.

AI literacy: support scientists with short training on confidence scores, failure modes and why every design needs wet-lab validation (GUIDE-006, Art. 4 AI Act).

Communication and pilot: start with a volunteer group of 3-5 scientists on one target and share hit rates openly (GUIDE-008, GUIDE-009).

Typical concerns: fear of replacing wet-lab work (it reduces cycles, not experiments) and of black-box results. Scientists stay in control: they choose targets, filters and what is synthesised.

Risks, regulation & compliance

• EU AI Act: minimal risk for research and development; scientific R&D systems are largely out of scope (Art. 2(6)), and AI literacy (Art. 4) still applies to staff. Outputs used in a medical device or clinical decision would trigger separate regimes (MDR/IVDR); drug development follows EMA/GMP rules for the resulting product, not for the design tool.
• GDPR: generally no personal data (sequences of proteins). Becomes relevant if patient genomic or clinical data is used (see SCI-009, SCI-017).
• Biosecurity (dual use): AI protein design could in principle be misused to design harmful agents. Mitigations at policy level: use providers that screen synthesis orders (IGSC Harmonized Screening Protocol), apply the community commitments on responsible AI-enabled protein design [9], restrict access to tools without safeguards, keep audit logs, and involve your biosafety officer. In Germany/EU, genetic engineering law (GenTG) and the EU dual-use export rules (Regulation 2021/821) may apply to certain agents and know-how.
• IP and licensing: model licences differ (non-commercial vs. permissive); AI-designed proteins raise open questions on inventorship in patents (human contribution required at the EPO).
• Confidentiality: target and sequence data is valuable IP; prefer EU regions, contractual no-training clauses.

Getting started

1. Define one concrete target and success criterion (e.g. binder with nM affinity, enzyme with +X degrees stability).
2. Run public tools (AlphaFold Server/DB, Boltz or Chai on a cloud GPU) on 5-10 known complexes to calibrate against known structures.
3. Pilot a small design campaign (tens to hundreds of designs) with RFdiffusion/ProteinMPNN or a Path A vendor, with CRO validation.
4. Set up filtering, sequence-order biosecurity screening and a tracking sheet linking predictions to assay results.
5. Review licences and IP position before moving to commercial use.

Sources

1. Jumper et al., Highly accurate protein structure prediction with AlphaFold, Nature 2021 - https://www.nature.com/articles/s41586-021-03819-2
2. Nobel Prize in Chemistry 2024 - https://www.nobelprize.org/prizes/chemistry/2024/press-release/
3. AlphaFold Protein Structure Database (EMBL-EBI) - https://alphafold.ebi.ac.uk/
4. Abramson et al., Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature 2024 - https://www.nature.com/articles/s41586-024-07487-w
5. Boltz (open-source co-folding and affinity) - https://github.com/jwohlwend/boltz
6. Watson et al., De novo design of protein structure and function with RFdiffusion, Nature 2023 - https://www.nature.com/articles/s41586-023-06415-8
7. EvolutionaryScale ESM - https://github.com/evolutionaryscale/esm
8. OpenFold3 - https://github.com/aqlaboratory/openfold-3
9. Community values, guiding principles and commitments for the responsible development of AI for protein design - https://responsiblebiodesign.ai/
10. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj