AI models in antibody discovery help scientists analyze antibody sequences, predict structures, design new variants, assess developability, and model antibody-antigen interactions. Recent 2025 reviews show that AI is becoming more closely integrated with high-throughput experimentation, especially when large antibody candidate pools need to be prioritized before wet-lab testing123.
AI models for antibody design and optimization are advancing so rapidly that keeping track has become a challenge in itself. To help, I break the field down into categories and give biologists a practical overview below. I am Julia Pizzolato, PhD, Commercial Scientific Liaison at evitria in Zurich, with a background in protein engineering, antibody design, data science and biomanufacturing analytics. In my work at evitria, I advise customers and project teams through technical consultation, feasibility assessments and training.
1. Protein Language Models: Reading the Grammar of Antibodies
Protein language models work like language models for text, but they learn from amino acid sequences instead of words. For non-ML biologists, the simplest explanation is that these models learn the “grammar” of proteins and antibodies by studying very large sequence datasets.
After training, protein language models convert antibody sequences into numerical representations called embeddings. These embeddings can support sequence clustering, missing residue completion, mutation ranking, paratope prediction, and early candidate prioritization. Recent reviews describe protein language models as a core component of AI-driven antibody discovery because they can rapidly process sequence libraries before experimental screening13.
Common examples include ESM-2, ESM3, ProtT5, AntiBERTa, AntiBERTy, AbLang, IgBert, IgT5, BALM, and IgLM. General protein models learn from broad protein space, while antibody-specific models are trained on antibody repertoire data and are often more useful for variable heavy and light chain analysis.
The problem is that sequence plausibility does not guarantee expression success. An antibody can look strong in silico but still express poorly, aggregate, or fail in downstream assays. evitria helps solve this by converting AI-prioritized sequence lists into CHO-expressed, purified antibodies that can be experimentally compared.
2. Structure Prediction Models: From Sequence to 3D Hypothesis
Antibody function depends strongly on three-dimensional structure, especially the complementarity-determining regions that form the antigen-binding surface. The CDR-H3 loop is particularly important because it is highly diverse and often central to antigen recognition.
Structure prediction models take an antibody sequence and predict its likely 3D structure. Recent antibody design reviews highlight structure prediction as an essential step for connecting sequence-level AI outputs to binding-site geometry, CDR conformation, and antigen recognition hypotheses23.
Common examples include AlphaFold2, AlphaFold3, RoseTTAFold, IgFold, DeepAb, ImmuneBuilder, ABodyBuilder2, NanoBodyBuilder2, and BALMFold. These tools are powerful, but predicted structure is not the same as functional performance. Binding depends on CDR flexibility, antigen conformation, glycosylation, expression quality, and assay context.
evitria helps researchers move from predicted structures to physical molecules by producing candidates in a CHO-based mammalian expression system suitable for downstream binding and characterization assays.
3. Generative Models: Designing New Antibody Candidates
Generative models move AI from analysis to design. Instead of only reading existing sequences, they create new antibody sequences, CDR loops, or structures that may satisfy a desired goal.
A generative model is like a design engine. It learns the rules of functional proteins and then proposes new molecules that may still follow those rules. Recent 2025 reviews describe rapid growth in generative AI approaches for antigen-conditioned antibody design, CDR engineering, affinity maturation, and multi-property optimization23.
Different generative approaches are used. Autoregressive models generate amino acids one at a time. Variational autoencoders sample new variants from compressed sequence spaces. Diffusion models gradually transform noise into plausible structures. Inverse-folding models start from a backbone and propose amino acid sequences likely to fold into it.
Common examples include ProteinMPNN, RFdiffusion, IgLM, AbGPT, PALM-H3, DiffAb, AbEgDiffuser, AntiBARTy Diffusion, LaMBO-2, DyAb, Property Enhancer, and newer geometry-aware co-design tools. Each has their own strength, application and use case. For example, ProteinMPNN generates sequences for a given structure, while AbEgDiffuser can be used for antibody sequence-structure co-design. GeoGAD, published in 2026, uses geometric information from the antigen-antibody interface to guide the generation of CDR sequences and structures for affinity optimization456.
The problem is that AI-generated antibodies can fail during expression, purification, binding, or stability assessment. evitria’s high-throughput antibody production helps AI teams test whether generated sequences can become real, purified, assay-ready antibodies.
4. Developability and Property Predictors: Finding Risks Earlier
A strong binder is not always a good therapeutic candidate. Antibodies can fail because of aggregation, poor solubility, high viscosity, weak expression, chemical liabilities, polyreactivity, immunogenicity risk, or poor manufacturability.
Developability and property predictors aim to identify these risks earlier. They may use sequence features, structural features, surface charge, hydrophobicity, language model embeddings, or experimental high-throughput assay data. Recent work emphasizes that large, standardized experimental datasets are essential for training AI and ML models that predict antibody developability more reliably71.
Common examples include TAP, CamSol, SoluProt, Aggrescan-type tools, SAP-based methods, humanness predictors, immunogenicity predictors, and newer ML models trained on high-throughput developability datasets.
The problem is that developability prediction depends on clean, comparable experimental data. evitria supports this need through standardized analytical quality control, including titer measurement, HPLC-SEC for purity and aggregation, CE-SDS for structural integrity, and endotoxin measurement. This helps connect computational developability predictions with real antibody behavior.
5. Docking and Interaction Models: Predicting Antibody-Antigen Binding
Docking and interaction models predict how an antibody may bind to an antigen. They help estimate the antibody-antigen interface, paratope-epitope contacts, and binding orientation. Docking can be understood as molecular fitting. The model places the antibody and antigen together in many possible orientations and ranks which complexes appear most plausible.
Common examples include ClusPro, HADDOCK, RosettaDock, SnugDock, AlphaFold-Multimer, AlphaFold3, and AI-augmented docking workflows. Recent AI-augmented docking work shows that machine learning can support antibody-antigen complex prediction when combined with physics-based docking strategies8.
The problem is that docking predictions can miss biological context. Flexible CDR loops, antigen conformation, and assay conditions can all affect binding. evitria helps by providing purified antibody material suitable for downstream binding and functional assays, allowing teams to test whether predicted interactions translate into real biological value.
How AI Models Fit into an Antibody Discovery Pipeline
In practice, AI models are rarely used alone. A typical AI-assisted antibody discovery workflow may look like this:
Protein language model embeddings → virtual screening → structure prediction → docking → developability filtering → CHO expression → purification and QC → functional testing
This pipeline matters because each AI model has limitations. Language models capture sequence patterns but not full biological function. Structure models predict shape but not always binding. Docking models suggest interfaces but may miss flexibility. Developability predictors reduce risk but cannot replace analytical testing. Recent 2025 reviews emphasize that AI-driven antibody discovery is strongest when computational prioritization is combined with high-throughput experimental validation and iterative model refinement12.
For AI-designed or AI-prioritized antibodies, evitria’s CHO-based high-throughput antibody production provides the physical validation step. AI can identify which sequences look promising. evitria helps determine which candidates can actually be expressed, purified, characterized, and moved forward.
The Future: Generalist Biological AI and Agentic AI
The field is moving toward generalist biological AI. These are foundation models trained across multiple biological data types, including DNA, RNA, proteins, structures, small molecules, and cellular data. The goal is to support many biological tasks without extensive task-specific retraining.
For antibody discovery, this could allow more integrated prediction. Future systems may evaluate whether an antibody is likely to fold correctly, bind a target, express in mammalian cells, avoid developability risks, and function in a relevant biological context. ESM3 reflects this direction within protein biology by connecting sequence, structure, and function, while a 2026 Nature Biotechnology review describes generalist biological AI as a broader framework for modeling the “language of life” across DNA, RNA, proteins, and cellular systems910.
Agentic AI in Biotech
A second direction is agentic AI: teams of large language model (LLM) agents that plan a project, choose tools, and run them in sequence instead of answering a single question. The use of LLMs has so far been largely limited to answering specific scientific questions rather than performing open-ended research11.
Antibody design already has a first experimental example. In a 2025 Nature study, a “Virtual Lab” of LLM agents, working with a human researcher, built a nanobody design pipeline that combined the protein language model ESM, AlphaFold-Multimer, and the Rosetta software, and designed 92 new nanobodies against SARS-CoV-2. Experimental validation found a range of functional nanobodies with promising binding profiles across variants11.
This result needs to be read carefully. It is one study on one target family, and a human researcher stayed in the loop. Designed sequences are not validated therapeutics, and the study does not show that agents generalize to other antigens, antibody formats, or manufacturing requirements. What it shows is that agents can connect the model classes described above into one workflow and hand a set of candidates to the lab.
That makes the experimental step more important, not less. If agents can propose candidates faster, the bottleneck moves to expressing, purifying, and testing them.
Even with generalist and agentic AI, experimental data will remain essential. The most valuable workflows will be closed-loop systems where AI design is followed by high-throughput expression, analytical testing, and model refinement. evitria is positioned to support this next phase by providing consistent CHO-expressed antibody material for AI-driven discovery programs.
Key Takeaway
The main AI models used in antibody discovery and engineering are protein language models, structure prediction models, generative models, developability predictors, and docking or interaction models. Each model supports a different decision in the antibody discovery process.
But AI alone does not create a validated antibody. The real value comes when computational predictions are connected to reliable experimental expression, purification, and characterization.
For teams working with AI-designed or AI-prioritized antibody candidates, evitria helps bridge the gap between digital design and biological validation through high-throughput CHO-based antibody production, purification, and analytical QC.
Frequently Asked Questions
The main AI models are protein language models, structure prediction models, generative models, developability predictors, and docking or interaction models.
No. AI can reduce the number of candidates entering the lab, but expression, binding, stability, specificity, and function still require experimental testing.
AI can generate or prioritize many sequences. evitria helps convert selected sequences into CHO-expressed, purified, assay-ready antibodies for real experimental validation.
They may fail because of poor expression, aggregation, instability, low binding, poor specificity, or unfavorable developability properties. Therefore, it is important to test in silico designed antibodies early in the laboratory.
evitria supports parallel CHO expression, purification, and analytical QC, helping teams compare antibody candidates under standardized conditions.
Generalist biological AI refers to foundation models trained across biological data types such as DNA, RNA, proteins, structures, small molecules, and cellular data.
Agentic AI uses LLM-based agents that plan and run several computational tools in sequence. In a 2025 Nature study, a team of agents working with a human researcher designed 92 SARS-CoV-2 nanobodies, and experiments found functional binders. The candidates still required laboratory validation.
References
- Matsunaga R, Tsumoto K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. Journal of Biomedical Science. 2025;32:46. https://doi.org/10.1186/s12929-025-01141-x ↩︎
- Vecchietti L, Wijaya BN, Armanuly A, Hangeldiyev B, Jung H, Lee S, Cha M, Kim HM. Artificial intelligence-driven computational methods for antibody design and optimization. mAbs. 2025;17:2528902. https://doi.org/10.1080/19420862.2025.2528902 ↩︎
- Park S, Jeong W, Kim Y, Lee CH, Seok C. Artificial intelligence in therapeutic antibody design: Advances and future prospects. Current Opinion in Structural Biology. 2025;94:103084. https://doi.org/10.1016/j.sbi.2025.103084 ↩︎
- Dauparas J, Anishchenko I, Bennett N, et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science. 2022;378(6615):49-56. https://doi.org/10.1126/science.add2187 ↩︎
- Zhu Y, Shi X, Zhang J, Sun W, Wang L. AbEgDiffuser: Antibody sequence-structure codesign with equivariant graph neural networks and diffusion models. Journal of Chemical Theory and Computation. 2025;21:11307-11317. https://doi.org/10.1021/acs.jctc.5c00990 ↩︎
- Wei S, Zhang J, et al. GeoGAD: geometry-aware antibody design framework for sequence-structure co-modeling, CDR design, and affinity optimization. Bioinformatics. 2026;42:btag042. https://doi.org/10.1093/bioinformatics/btag042 ↩︎
- Arsiwala A, Bhatt R, Van Niekerk L, et al. A high-throughput platform for biophysical antibody developability assessment to enable AI/ML model training. mAbs. 2025;17:2593055. https://doi.org/10.1080/19420862.2025.2593055 ↩︎
- Gaudreault F, Sulea T, Corbeil CR. AI-augmented physics-based docking for antibody-antigen complex prediction. Bioinformatics. 2025;41:btaf129. https://doi.org/10.1093/bioinformatics/btaf129 ↩︎
- Hayes T, Rao R, Akin H, et al. Simulating 500 million years of evolution with a language model. Science. 2025;387:850-858. https://doi.org/10.1126/science.ads0018 ↩︎
- Rao VM, Zhang S, Plosky BS, et al. Generalist biological artificial intelligence in modeling the language of life. Nature Biotechnology. 2026. https://doi.org/10.1038/s41587-026-03064-w ↩︎
- Swanson K, Wu W, Bulaong NL, et al. The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies. Nature. 2025;646:716-723. https://doi.org/10.1038/s41586-025-09442-9 ↩︎