Zero-shot de novo antibody design to a chosen epitope with sub-nanomolar affinity and drug-like developability
Hit rates rose by orders of magnitude in 2024-25 and company reports claim about half of targets yield binders in one round, so a path is visible and difficulty (50) sits below de novo enzymes (66) and scaffold-agnostic affinity (63); the open part is sub-nanomolar affinity, epitope fidelity and developability together, demonstrated outside company reports. Application value 74 is second in this area, just below general affinity prediction (76): therapeutic antibodies are a core pharmaceutical modality, but the capability is confined to one modality, where affinity prediction serves every structure-based programme; 74 places it above cryo-CMOS control (60) and near the surface-code milestone (70). Leverage 43 because it reshapes antibody discovery without other fields depending on the answer, comparable to secure speculation (40).
The problem
Using AI to write an antibody from scratch that sticks to a chosen spot on a target.
- 1Pick target and spot
- 2AI writes antibody loops
- 3Computer filters designs
- 4Lab tests binding and stability
- 5Goal: drug-quality in one round
Antibody structure. The antigen-binding part consists of variable domains with six loops, the complementarity-determining regions (CDRs). CDR-H3 is long, varied and flexible, which makes it the hardest to model and the main driver of specificity.
Design pipeline. A generative model (such as RFdiffusion fine-tuned for antibodies) places CDR loops against a specified epitope, an inverse-folding model proposes sequences, and filters based on predicted structure remove poor designs. Candidates are expressed (for example on yeast) and sorted by binding.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostaticsComputational Chemistry · D52
- One classical force field accurate for folded proteins, disordered proteins and weak protein–protein associationComputational Chemistry · D51.5
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceComputational Medicine · D64
- Step 2
- Prospective protein–ligand binding free energies at 1 kcal/mol RMSE across diverse, difficult targetsComputational Chemistry · D50
- Step 3
- Physically valid protein-ligand co-folding for unseen chemotypes and pocketsComputational Medicine · D47
- Step 4
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Reverse-mode automatic differentiationAlgorithm
- Fast matrix multiplicationAlgorithm
- Step 2
- The transformer architectureAlgorithm
- Step 3
- Step 4
- Protein language models and generative protein designFrontier · host
- Then
- Zero-shot de novo antibody design to a chosen epitope with sub-nanomolar affinity and drug-like developabilityThe problem
State of the art
- RFantibody (Nature 2025-26). Bennett et al. fine-tuned RFdiffusion to design VHHs, scFvs and full antibodies to user-specified epitopes, screened by yeast display. Initial designs bound with modest affinity (tens to hundreds of nM); OrthoRep-based affinity maturation reached single-digit nM; cryo-EM and crystal structures confirmed atomic-level agreement for selected designs. Success rates per design are low and the work needed library-scale screening. An unreviewed commentary on Zenodo disputes the accuracy claims; it should be weighed as a non-refereed critique.
- Chai-2 (bioRxiv preprint, July 2025). A company technical report claims a 16% per-design hit rate in fully de novo antibody and nanobody design across 52 targets with no antibody or nanobody in the PDB, at most 20 designs per target in a 24-well-plate format, and at least one hit for about 50% of targets in one round; the abstract states hits "often" have strong affinity and favourable drug-like profiles but does not give K_D values or developability data. Not peer reviewed in sources reviewed; the model is not openly released.
- Structure prediction baseline. AlphaFold 3 antibody-nanobody docking had high-accuracy success rates of roughly 9-13% with single-seed sampling and about 60% outright failure for docking in a 2024 study (Proteins/bioRxiv), limited by CDR H3 modelling; Isomorphic Labs reports better high-accuracy antibody-antigen prediction than AlphaFold 3 for its closed IsoDDE engine; this is a company claim with no released model and no independent replication, and the reported margin could not be confirmed from a primary source for this entry.
- Not shown. Sub-nanomolar zero-shot binders with documented developability at the target success rates have not been reported in a peer-reviewed study.
Realistic applications
Antibodies are the largest class of biologic drugs; zero-shot epitope-targeted design would shorten discovery from months to weeks, enable targeting of conserved or poorly immunogenic epitopes and membrane proteins, and cut animal use. It would also provide a general interface-design method for other biologics (minibinders, bispecifics).
What counts as solved
A peer-reviewed study (or an independent replication of a company result) satisfying criteria (i)-(iv) on a pre-registered panel of at least 50 targets, with sequences, affinity data and developability assays released, and with at least one independent group reproducing the pipeline on new targets.
How it gets verified
ObservationSettled by new measurements or observations.
Needs new binding and developability assays: a peer-reviewed study (or an independent replication of a company result) must meet criteria (i) to (iv) on a pre-registered panel of at least 50 targets, with sequences, affinity data and assay data released and at least one independent group reproducing the pipeline on new targets.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Most of the remaining work is experimental: affinity, epitope confirmation and developability data on many targets, with model improvements guided by them.
Bottlenecks
ranked by votes- #1EmpiricalDevelopability and epitope confirmation data
Hit rates are reported by binding; polyreactivity, stability, expression and structural epitope confirmation exist for few designs.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Nov 5, 2025RFantibody in Naturesource
Epitope-targeted VHH, scFv and antibody designs; structures confirm atomic accuracy; OrthoRep maturation to single-digit nM.
- Jul 5, 2025Chai-2 technical report (preprint)source
Claims 16% per-design hit rate over 52 targets and about 50% of targets with a binder in one round; no K_D or developability detail in the abstract.
- Sep 24, 2024AlphaFold 3 antibody docking limitssource
High-accuracy antibody/nanobody docking succeeds in roughly 9-13% of cases (single seed); CDR H3 modelling is the bottleneck.