Bispecific Antibodies from Hybridoma Engineering to AI-Assisted In-silico Design

Abstract

Bispecific antibodies (BsAbs) bind two different epitopes or antigens in one molecular product. By physically co-localizing targets, they enable mechanisms unavailable to conventional monospecific IgG: immune-cell redirection, dual-pathway blockade, receptor agonism, conditional activation, coagulation-factor bridging, and targeted delivery. This review follows the field from early chemical cross-linking and quadroma antibodies through recombinant fragment formats, Fc engineering, and current AI/ML-assisted design. It emphasizes what computation can and cannot establish: models can prioritize architectures, sequences, interfaces, and developability, but potency, specificity, pharmacology, immunogenicity, manufacturability, and clinical benefit remain experimental questions.

Historical development

The first BsAb concepts joined two antibody specificities by chemical conjugation or hybrid-hybridoma (quadroma) technology. Early products were limited by heterogeneous pairing, immunogenicity, low yields, and difficult purification. Recombinant DNA enabled defined Fab, scFv, diabody, tandem scFv, and single-domain formats. Later Fc-containing architectures restored IgG-like half-life and manufacturability while using knobs-into-holes, electrostatic steering, common-light-chain, CrossMab, DuoBody, and controlled Fab-arm exchange to enforce correct heavy/light-chain pairing. Fragment-only BiTE-like molecules improved immune synapse formation but often traded half-life for activity and required continuous infusion or half-life extension.

Therapeutic mechanisms

Immune-cell engagers

CD3×tumor-antigen molecules bridge T cells to malignant cells; examples include blinatumomab (CD19×CD3), teclistamab (BCMA×CD3), mosunetuzumab and glofitamab (CD20×CD3), epcoritamab (CD20×CD3), talquetamab (GPRC5D×CD3), and tarlatamab (DLL3×CD3). NK-cell engagers use CD16 or NK-cell receptors. Clinical design must balance target density, synapse geometry, affinity, valency, cytokine release, and on-target/off-tumor risk.

Dual-pathway blockade and receptor modulation

BsAbs can inhibit two ligands/receptors simultaneously, block a resistance route, or agonize a receptor only when both antigens are present. Examples include amivantamab (EGFR×MET), zenocutuzumab (HER2×HER3), and HER2 biparatopic zanidatamab. Similar logic is being explored in inflammation, fibrosis, ophthalmology, infectious disease, and metabolic disorders.

Coagulation and non-oncology uses

Emicizumab (FIXa×FX) functionally substitutes for factor VIII in hemophilia A. Other non-oncology strategies include blood-brain-barrier transcytosis shuttles, antiviral neutralization, cytokine neutralization, and tissue-selective immune modulation.

Targeted delivery and conditional activity

One specificity can recognize a tumor or tissue antigen while the second recruits a payload, receptor, enzyme, or masked activation mechanism. BsAb-drug conjugates and immune-cell-recruiting constructs are especially sensitive to linker geometry, internalization, valency, and systemic exposure.

Pre-AI/ML development workflow

Before modern ML, BsAb discovery was an iterative design-build-test process:

1. Select target biology and desired geometry (cell-cell bridge, dual blockade, agonism, or delivery).
2. Choose format, valency, Fc activity, half-life, and linker architecture.
3. Obtain parental antibodies by immunization, hybridoma, phage/yeast display, or synthetic libraries.
4. Humanize or affinity-mature variable regions by CDR grafting, targeted mutagenesis, and display selection.
5. Solve chain-pairing problems through common light chains, knobs-into-holes, CrossMab, DuoBody, or orthogonal interfaces.
6. Screen expression, assembly, aggregation, thermal stability, non-specific binding, and target binding.
7. Measure cell killing, cytokines, receptor modulation, pharmacokinetics, and animal efficacy/toxicity.

Computational tools were mainly homology modelling, docking, molecular mechanics, sequence liability filters, and statistical design-of-experiments. These methods reduced search space but did not remove the need for large experimental screens.

AI/ML-era design

Modern workflows combine protein language models, antibody-specific structure predictors, generative sequence models, inverse folding, diffusion, Bayesian optimization, and active learning. A defensible BsAb pipeline is multi-objective:

• Architecture: enumerate formats and linker/valency choices consistent with target geometry and pharmacology.
• Sequence generation: use antibody language models, ProteinMPNN/AbMPNN, IgGM, or constrained CDR design while preserving framework and pairing rules.
• Structure: predict each arm and the assembled BsAb; use AlphaFold-Multimer/ColabFold, IgFold, NanoNet, Boltz, or equivalent models with model agreement and PAE/pLDDT gates.
• Interface: evaluate contacts, clashes, buried surface, electrostatics, hydrogen bonds, and relative Rosetta/FoldX energies; do not treat any single score as affinity.
• Developability: assess aggregation patches, hydrophobicity, charge, chemical liabilities, immunogenicity, polyspecificity, expression, and proteolysis risk.
• Optimization: use Bayesian or active-learning loops where measured binding, expression, cytokine release, and stability—not model scores alone—update the next design round.
• Evidence and provenance: retain sequence hashes, model versions, weights, random seeds, input structures, rejected candidates, and assay results.

BiSpec Pairwise AI illustrates ML for selecting target combinations. Deep-learning antibody libraries and sequence-based antibody models demonstrate prospective developability prioritization. Generative and structure-guided models can propose novel binders, but BsAb-specific training data are sparse, format-dependent, and biased toward successful programs. Leakage-resistant validation and prospective experiments are essential.

Product-development risks

BsAbs create coupled risks: incorrect chain pairing, half-antibody and aggregate species, target-mediated clearance, cytokine-release syndrome, immunogenicity, asymmetric pharmacokinetics, and an unfavorable therapeutic window. AI can prioritize these risks; it cannot certify them away. A candidate should advance only after orthogonal structure/interface evidence and experimental confirmation of expression, monomer content, affinity/kinetics, cell-based function, specificity, PK, and safety.

Recommended NextELN workflow

NextELN should capture target pair, desired mechanism, format, valency, Fc policy, linker constraints, parental sequences, antigen structures, and assay acceptance criteria. The public service stores a signed request and input artifact. An isolated GPU worker performs model inference and returns versioned PDB/FASTA/JSON artifacts and a limitations report. The admin portal exposes download, processing state, provenance, and result upload; it must never present a computational score as measured efficacy.

Selected references

• Biology drives the discovery of bispecific antibodies as innovative therapeutics. PMID 33928225.
• Kontermann & Brinkmann, Bispecific antibodies, DOI 10.1016/j.drudis.2015.02.008.
• Brinkmann & Kontermann, The making of bispecific antibodies, DOI 10.1080/19420862.2016.1268307.
• Bispecific antibodies: design, therapy, perspectives. PMCID PMC5784585.
• Symmetry breaking: bispecific antibodies, beginnings, 50 years on. PMCID PMC3380354.
• Staerz et al., Hybrid antibodies can target a specific T cell receptor, DOI 10.1038/314628a0.
• Bargou et al., Tumor regression in cancer patients by very low doses of a T-cell-engaging antibody, DOI 10.1126/science.1158545.
• Oldenburg et al., Emicizumab prophylaxis in hemophilia A with inhibitors, PMID 28691557.
• Heier et al., Faricimab in neovascular age-related macular degeneration, DOI 10.1016/S0140-6736(22)00010-1.
• Park et al., Amivantamab in EGFR exon 20 insertion-mutated NSCLC, DOI 10.1200/JCO.21.00662.
• Expanding the Boundaries of Biotherapeutics with Bispecific Antibodies. PMID 30132211.
• The present and future of bispecific antibodies for cancer therapy. PMID 38448606.
• A pivotal decade for bispecific antibodies? PMCID PMC10936642.
• FDA staff: Bispecific Antibodies—An Area of Research and Clinical Applications. FDA workshop document.
• Applications and challenges in designing VHH-based bispecific antibodies: leveraging machine learning solutions. PMID 38666503.
• Applications and challenges in designing VHH-based bispecific antibodies (full text). PMCID PMC11057648.
• Jumper et al., Highly accurate protein structure prediction with AlphaFold, DOI 10.1038/s41586-021-03819-2.
• Abanades et al., ImmuneBuilder, DOI 10.1038/s42003-023-04927-7.
• Dauparas et al., Robust deep learning-based protein sequence design using ProteinMPNN, DOI 10.1126/science.add2187.
• AntiFold antibody inverse folding, arXiv:2405.03370.
• tFold de novo epitope-specific antibody design, DOI 10.1038/s41467-025-67361-9.
• BiSpec Pairwise AI: guiding selection of bispecific target combinations. PMID 38713378.
• In silico proof of principle of machine learning-based antibody design. PMCID PMC8986205.
• Deep learning-based design and experimental validation of a medicine-like human antibody library. PMID 39851074.
• Artificial intelligence-driven computational methods for antibody design and optimization. PMID 40677216.
• Best practices for machine learning in antibody discovery and development. arXiv:2312.08470.
• Applying computational protein design to therapeutic antibody discovery. arXiv:2503.00913.
• In Silico Approaches to Deliver Better Antibodies by Design. arXiv:2305.07488.

Conclusion

BsAbs have progressed from difficult experimental assemblies to a clinically validated modality with multiple mechanisms. AI/ML can make architecture selection, sequence generation, structure prediction, and developability triage more efficient, but the winning workflow remains closed-loop: model, build, measure, learn, and document uncertainty.