Antibody and Biologic Design
Antibody and biologic design sits where structure, sequence, immunology, and manufacturing meet. A designed binder is useful only if affinity, specificity, expression, stability, safety, and developability survive the same program. RFantibody (Bennett et al., Nature 2026) extends the RFdiffusion lineage to antibody scaffolds with experimental validation, and is the current peer-reviewed reference. Antibody-specific structure-prediction (IgFold) and language-model (AbLang) tools handle adjacent tasks. None of this AI work removes the developability cascade required for any clinical biologic.
Use this chapter to:
- Design and evaluate antibodies, binders, and biologics where affinity, specificity, developability, immunogenicity, and manufacturability all matter.
- Antibody structure and sequence models are helpful, but CDR behavior, epitope biology, glycosylation, liabilities, and assay context determine translation.
Prerequisites: Protein Structure Prediction for confidence-metric discipline; Protein Design and Engineering for the generative-design stack that RFantibody extends.
Summary: Design and evaluate antibodies, binders, and biologics where affinity, specificity, developability, immunogenicity, and manufacturability all matter. The field has strong structure and representation tools, with de novo antibody design advancing but still constrained by experimental validation and developability filters.
Key point: Antibody structure and sequence models are helpful, but CDR behavior, epitope biology, glycosylation, liabilities, and assay context determine translation. Open question: whether design gains survive epitope biology, developability, manufacturability, and immune-context testing.
Bottom line: Antibody AI links molecular design to immunology, oncology, infectious disease, protein engineering, cell therapy, and biomanufacturing.
What is this field trying to solve? Design and evaluate antibodies, binders, and biologics where affinity, specificity, developability, immunogenicity, and manufacturability all matter.
What is the core idea? Antibody structure and sequence models are helpful, but CDR behavior, epitope biology, glycosylation, liabilities, and assay context determine translation.
What is the current state of the field? The field has strong structure and representation tools, with de novo antibody design advancing but still constrained by experimental validation and developability filters.
What do we know, and what remains open? Known reference points include RFantibody, IgFold, AbLang, antibody language models, SAbDab, OAS, Therapeutic Antibody Profiler, developability assays, and binding benchmarks. What remains open is whether design gains survive epitope biology, developability, manufacturability, and immune-context testing.
Why does this matter? Antibody AI links molecular design to immunology, oncology, infectious disease, protein engineering, cell therapy, and biomanufacturing.
Introduction
Protein design systems now support binder and scaffold design, but antibodies add constraints from immune repertoires, CDR loop geometry, epitope context, glycosylation, Fc behavior, and developability. RFdiffusion and AlphaFold 3 are relevant to this space because they address structure and interaction modeling (Watson et al., 2023; Abramson et al., 2024). RFantibody (Bennett et al., 2026) extends RFdiffusion to antibody scaffolds with peer-reviewed experimental validation.
Antibody biology before antibody AI
An antibody program starts with biology, not with a sequence generator. The antigen may be soluble, cell-surface, conformational, post-translationally modified, strain-variable, or masked in the relevant tissue. The epitope may need to block a ligand, agonize a receptor, recruit immune effector function, deliver a payload, or avoid cross-reactivity with related proteins. Those requirements shape the paratope, heavy-light chain pairing, isotype, Fc engineering, and format before AI enters the workflow.
The CDR loop problem is why antibody-specific models matter. Antibody binding depends heavily on loop geometry, especially CDR-H3, and generic protein structure models can be less reliable in that region than their global confidence scores imply. IgFold, AbLang, RFantibody, and antibody-tuned evolution workflows should therefore be treated as antibody-specific layers, not as interchangeable substitutes for generic protein tools.
Developability is a product filter, not a final checklist
Therapeutic antibodies occupy a constrained biophysical envelope. Jain and colleagues showed that clinical-stage antibodies have recognizable developability profiles, which makes outlier detection useful before a candidate consumes expensive work (Jain et al., 2017). Raybould and colleagues operationalized this idea in computational form by comparing candidates against therapeutic-antibody property ranges (Raybould et al., 2019). Hötzel and colleagues linked nonspecific binding to rapid clearance risk, a reminder that affinity for the intended antigen is only one part of in-vivo behavior (Hötzel et al., 2012).
The cascade should run early: expression, aggregation, thermal stability, polyspecificity, viscosity, post-translational modification, immunogenicity risk, pharmacokinetics, and formulation. Treating developability as a cleanup step after affinity optimization creates false progress because many high-affinity molecules are poor products.
Format changes reset parts of the evidence
Monospecific IgG design is not the same as bispecific design, antibody-drug conjugate design, VHH design, or Fc-fusion design. Bispecifics add geometry, valency, chain-pairing, cytokine-release, and tissue-distribution concerns. ADCs add linker stability, payload potency, bystander effect, internalization, and manufacturing controls. VHHs change size, tissue penetration, half-life, and immunogenicity profile. A model result in one format should not be generalized to the others without format-specific assays.
What is demonstrated?
Demonstrated capability includes structure-guided binder design, antibody structure modeling, affinity maturation, and sequence optimization workflows for selected targets. RFdiffusion demonstrated protein binder design tasks with experimental follow-up (Watson et al., 2023). AlphaFold 3 demonstrated biomolecular interaction prediction that includes protein complexes relevant to binder assessment (Abramson et al., 2024). RFantibody demonstrated atomically accurate de novo antibody designs with experimental validation (Bennett et al., 2026). IgFold demonstrated fast antibody-specific structure prediction (Ruffolo et al., 2023).
Protein language models also support antibody optimization when the claim is framed narrowly. Hie and colleagues showed that general protein language models can guide efficient antibody evolution and affinity maturation (Hie et al., 2024). Shanker and colleagues showed structure-informed language-model evolution for protein and antibody complexes (Shanker et al., 2024). These are optimization and evolution results, not evidence that a model can move from antigen sequence to clinical biologic without immunology, developability, and manufacturing work.
Developability has its own evidence base. Clinical-stage antibodies occupy a constrained biophysical-property range (Jain et al., 2017); computational profiling can flag candidates outside common therapeutic-antibody envelopes (Raybould et al., 2019); and nonspecific binding or fast-clearance risk remains a pharmacokinetic liability that needs program-specific mitigation (Hötzel et al., 2012). These tools are triage, not clearance.
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| RFantibody | Atomically accurate de novo antibody design with wet-lab validation | Developability and clinical translation are separate evidence |
| RFdiffusion (general) | Designed binders and constrained protein generation | Antibody-specific behaviour requires antibody-specific tools |
| AlphaFold 3 | Complex prediction for biomolecular interactions | Antibody-antigen loop geometry has documented limits |
| IgFold | Fast antibody structure prediction | Structure prediction is upstream of design |
| AbLang | Antibody sequence language modelling | Sequence completion, not design end-to-end |
| ProteinMPNN | Sequence design around backbones | Developability filters remain external |
| Hie / Shanker PLM optimization | Affinity maturation and complex optimization | Optimization evidence, not end-to-end therapeutic design |
What is theoretical?
Theoretical capability includes antibody libraries designed around target epitopes with predictable affinity maturation and low developability risk. That goal is plausible for constrained domains, but target biology and manufacturability still determine program success. Reliable design of bispecific and multispecific formats is in an earlier stage than monospecific antibody design. Production-grade end-to-end pipelines that span design, developability triage, and manufacturability are partial.
What is beyond current capability?
Beyond current capabilities includes end-to-end biologic development from antigen sequence to clinical candidate without immunological, pharmacological, and manufacturing testing. No model removes the need for those gates. Fully autonomous antibody discovery without expert chemistry-and-immunology review is not demonstrated.
What would make this more promising?
Antibody claims would strengthen if independent groups reproduced de novo design across multiple antigen classes, epitope types, and antibody formats while reporting the full developability profile beside affinity. Stronger evidence would include prospective campaigns where AI-designed candidates survive expression, aggregation, viscosity, polyspecificity, immunogenicity-risk, pharmacokinetic, and manufacturing screens without manual rescue. Claims about bispecifics, ADCs, VHHs, and Fc-engineered formats need format-specific denominators rather than transfer from monospecific IgG evidence.
What should researchers, biotech teams, funders, and program leaders do with this?
- Keep epitope, paratope, isotype, format, and intended mechanism separate.
- Screen for aggregation, viscosity, expression, liabilities, immunogenicity, and polyspecificity in cascade rather than serially after design.
- Use orthogonal assays when structural models disagree with binding data.
- Treat high-affinity designs without specificity data as incomplete.
- Cite the peer-reviewed RFantibody Nature paper rather than the bioRxiv preprint when available.
- Audit licence terms for generative antibody tools before integrating into a commercial pipeline.