Compare · 11 models
Antibody design and analysis models you can run online
Computational antibody tools number, model, humanise, and design immunoglobulin sequences and structural coordinates, supporting researchers across therapeutic antibody engineering and repertoire analysis.
What antibody design and analysis models do
Antibody engineering requires addressing several distinct computational tasks: standardising variable domain nomenclature, predicting three-dimensional structures, evaluating developability traits such as solubility, and designing novel or humanised sequences. A primary challenge across these workflows is that antibodies share a conserved immunoglobulin fold, while antigen recognition relies on hypervariable complementarity-determining region (CDR) loops.
Methods divide between sequence-based and structure-based approaches. Sequence-focused models rely on standard numbering alignments or language models trained on repertoire collections such as the Observed Antibody Space to capture chain pairings and sequence diversity. Structure-focused tools exploit conserved framework geometry to generate all-atom coordinates at high speed, identify binding interfaces, or guide CDR sequence design through inverse folding and all-atom co-folding.
Compared
- ProGen2Salesforce ResearchGenerates protein and antibody sequences conditionally using autoregressive language modelling.Commercial OKReady
- AbLang2Oxford Protein Informatics GroupAn antibody language model for generating paired-chain sequence embeddings and guiding CDR design.Commercial OKOn demand
- ANARCIOxford Protein Informatics GroupNumbers antibody sequences and annotates framework and CDR regions across multiple numbering schemes.Commercial OKOn demand
- AntiFoldHummer et al., DTUAn antibody inverse folding model that designs CDR sequences tailored to antibody backbone frameworks.Commercial OKOn demand
- ImmuneBuilderOxford Protein Informatics GroupFast structure prediction for antibodies, nanobodies, and T-cell receptors across full repertoires.Commercial OKOn demand
- MAMMAL solubilityIBM ResearchBinary protein solubility classification directly from amino-acid sequences.Commercial OKOn demand
- p-IgGenopigAn autoregressive protein language model for designing diverse paired antibody heavy and light chain sequences.Check termsOn demand
- SapiensAstraZeneca / Prihoda et al.Germline-guided antibody humanisation tool proposing mutations that increase humanness while preserving affinity.Commercial OKOn demand
- IgFoldRuffolo et al., Johns HopkinsPredicts antibody 3D structures from heavy and light chains in sub-second runtimes.Non-commercialNot yet
- ParagraphDeane Lab / OxfordPredicts antibody binding interface residues from structure to support humanisation and immunogenicity studies.Commercial OKNot yet
- Promerabjing-mitGenerative model for biomolecular co-folding, binder filtering, and miniprotein and nanobody design.See its pageNot yet
How to choose
For sequence preprocessing and annotation, ANARCI executes instantaneously on CPU hardware, assigning standard Chothia, IMGT, or Kabat numbering schemes without requiring structural coordinates. When generating sequences de novo, model choice depends on the required conditioning and chain format. For paired-chain work, p-IgGen uses a GPT-NeoX autoregressive architecture requiring 16 GB of GPU memory to design paired heavy and light chains, while AbLang2 uses an 8 GB GPU footprint to generate paired-chain embeddings and guide CDR design from paired inputs. For therapeutic optimisation, Sapiens proposes germline-guided mutations from paired heavy and light chain sequences to enhance humanness while maintaining affinity, validated directly on clinical-stage antibodies.
When three-dimensional structures are needed from sequence alone, ImmuneBuilder and IgFold both produce all-atom coordinate files in sub-second runtimes with per-residue pLDDT confidence metrics. While IgFold is restricted to paired heavy and light chains, ImmuneBuilder also handles nanobodies and T-cell receptors across whole repertoires. Promera combines all-atom co-folding with masked-sequence prediction, allowing structural prediction that surpasses OpenFold3-p2 and Boltz-2 on therapeutic categories alongside epitope-conditioned design of nanobodies and miniproteins. If coordinate files already exist, Paragraph maps paratope residues from antibody structures alone, while AntiFold applies inverse folding to antibody backbones and CDR loops, outputting sequence designs scored with per-position log-odds metrics.
For evaluating developability attributes without 3D coordinates, MAMMAL solubility classifies protein solubility directly from primary amino-acid sequences using a 458-million-parameter foundation model. However, MAMMAL solubility provides only a binary classification without accounting for environmental factors like pH or temperature, and requires 24 GB of GPU memory. In contrast, tools such as Sapiens, AbLang2, AntiFold, ImmuneBuilder, and Paragraph operate within an 8 GB GPU requirement, while ANARCI runs without GPU acceleration.
Questions
Can I run these models without a GPU or installing anything?
8 of the 11 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which tools predict 3D antibody structures from sequence data?
ImmuneBuilder, IgFold, and Promera predict three-dimensional coordinates. ImmuneBuilder and IgFold run in sub-second execution times and provide per-residue pLDDT confidence scores. ImmuneBuilder covers antibodies, nanobodies, and T-cell receptors, whereas IgFold requires paired heavy and light chain sequences. Promera performs all-atom co-folding coupled with sequence design.
How does AntiFold differ from sequence generative models like ProGen2 and p-IgGen?
AntiFold is an inverse folding model that requires pre-existing three-dimensional coordinates in PDB format to design sequences tailored to specific antibody backbones and CDR loops. In contrast, ProGen2 and p-IgGen are autoregressive language models that generate sequence completions directly from primary sequence prompts without requiring structural coordinate inputs.
Which models require paired heavy and light chain inputs?
AbLang2, p-IgGen, Sapiens, and IgFold specifically require paired heavy and light chain sequence inputs. By contrast, ANARCI and MAMMAL solubility process individual sequence inputs, ProGen2 generates from single-sequence prefix prompts, and ImmuneBuilder and Promera support single-chain modalities such as nanobodies.
What tools are available for antibody humanisation?
Sapiens is built specifically for antibody humanisation, proposing germline-guided mutations from paired heavy and light chains to improve humanness while preserving target affinity, and is validated on clinical-stage molecules. Paragraph can also assist humanisation and immunogenicity workflows by predicting paratope residues directly from antibody coordinate structures.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.