ABodyBuilder3

What is ABodyBuilder3?

ABodyBuilder3 predicts the 3D structure of an antibody variable domain (Fv) from its paired heavy and light chain sequences. It is the third generation of the ABodyBuilder line from the Oxford Protein Informatics Group, built on the ImmuneBuilder framework, and it specializes in the part of antibody modeling that general structure predictors handle worst: the hypervariable CDR loops, especially CDR-H3.

Most of an antibody Fv is conformationally conserved, so the framework regions are easy to model. The binding loops are the hard part. CDR-H3 has no canonical templates, varies in length and sequence across antibodies, and largely determines where and how the antibody binds. ABodyBuilder3 focuses its accuracy there.

How to use ABodyBuilder3 online

Paste an antibody heavy chain (VH) and light chain (VL) sequence into ProteinIQ, pick a model checkpoint, and run the prediction on cloud GPUs. The result is a single Fv structure in PDB format with heavy and light written as separate chains, plus a summary of chain lengths and native confidence for the LM and pLDDT checkpoints. No installation, no antibody numbering setup, and no MSA step.

Inputs

Input	Format	Notes
Heavy chain (VH)	Amino-acid sequence or FASTA	Standard 20-letter alphabet; X is accepted for an unknown residue. Provide one variable domain of at most 200 residues, not the full heavy chain.
Light chain (VL)	Amino-acid sequence or FASTA	Required. Provide one variable domain of at most 200 residues. ABodyBuilder3 models paired Fv, so both chains are needed.
Precomputed ProtT5 embeddings	.pt file, up to 10 MiB	Optional with the LM checkpoint. The source-format dictionary must contain a floating-point plm_embedding tensor of shape (VH length + VL length, 1024), with heavy-chain rows first. Leave empty to generate embeddings.

Both inputs expect the variable domain sequence. Trimming constant-region residues before submitting avoids modeling regions the tool is not trained on. To assign and trim variable domains by antibody numbering, use ANARCI.

Settings

Setting	Default	Description
Model checkpoint	ABodyBuilder3-LM loss checkpoint	Selects a published trained model. The LM checkpoint uses ProtT5 embeddings. Both LM and pLDDT checkpoints predict per-residue confidence; the base checkpoint uses one-hot sequence features and has no confidence head.

The three checkpoints in the published model archive are available:

• ABodyBuilder3-LM loss checkpoint: the default. Uses ProtT5 embeddings and returns native per-residue pLDDT.
• ABodyBuilder3 pLDDT checkpoint: uses one-hot sequence features with a confidence head and returns native per-residue pLDDT.
• ABodyBuilder3 base loss checkpoint: uses one-hot sequence features without a confidence head.

An embedding file must correspond to the exact submitted sequences in heavy-then-light order. ProteinIQ checks tensor shape and finite values, but cannot establish the sequence provenance of an arbitrary embedding file. Tensor values are passed to the model without conversion or reordering. Files requiring executable pickle objects are rejected.

Results

Output	Format	Notes
Predicted Fv structure	PDB	Native heavy (H) and light (L) chains, residue numbering and atom records are preserved. Confidence is supplied separately; native PDB B-factors are not replaced with pLDDT.
pLDDT data	JSON	Returned for the LM and pLDDT checkpoints. Includes the original pLDDT vector plus chain-aware residue identities, 1-based positions, and scores. This file can be reused in workflows.
Summary table	Spreadsheet	Heavy and light chain lengths, total residues, model checkpoint, and mean pLDDT when available.
Run log	LOG	Native inference output, warnings and errors.

How ABodyBuilder3 works

ABodyBuilder3 inherits the ImmuneBuilder architecture: a graph-based network predicts inter-residue geometry, which is folded into atomic coordinates and refined. The two changes that define this version are the sequence representation and the confidence estimate.

Language model embeddings

ABodyBuilder2 encoded each residue as a one-hot vector. ABodyBuilder3 replaces that with embeddings from the ProtT5 protein language model, which has seen billions of natural protein sequences and carries evolutionary and structural context that a one-hot vector cannot. This richer representation stabilizes training and improves modeling of the variable loops, with the largest effect on CDR-H3 and CDR-L3. The gain on CDR-H3 backbone RMSD is real but modest; the more reliable benefit is more robust training and a model that no longer needs an ensemble to estimate confidence.

pLDDT confidence instead of an ensemble

Earlier ImmuneBuilder models estimated per-residue uncertainty by training an ensemble of networks and measuring their disagreement, which multiplies both training and inference cost. ABodyBuilder3 adds a pLDDT head that predicts local accuracy directly, in the same 0 to 100 scale used by AlphaFold. A single model with this head matches the ensemble's uncertainty estimate at a fraction of the compute.

Interpreting pLDDT

pLDDT	Meaning
> 90	High confidence, backbone very likely correct.
70 to 90	Reliable, framework and well-defined loops.
50 to 70	Treat with caution, common in long or unusual CDR-H3 loops.
< 50	Low confidence, the local geometry may be wrong.

Confidence tends to be high across the framework and lower through the CDRs. A low score concentrated in CDR-H3 is expected for long or atypical loops and is the signal to validate that region before relying on it.

When to use ABodyBuilder3 vs alternatives

ABodyBuilder3 is the right choice when the target is a paired antibody Fv and the goal is a fast, accurate model of the binding loops. It is purpose-built for paired VH/VL variable domains, runs in seconds, and needs no MSA. For a single-chain nanobody, use ImmuneBuilder with NanoBodyBuilder2.

For a full antigen-antibody complex rather than the unbound Fv, a co-folding model such as Chai-1 or Protenix predicts antibody and antigen together. For general (non-antibody) single-chain prediction, use ESMFold or AlphaFold 2. To redesign CDR sequences or graft new loops onto a predicted scaffold, pair the structure with AntiFold, Proteo-R1, or the antibody-specialized RFantibody and DiffAb.

A typical workflow numbers and trims the chains with ANARCI, predicts the Fv here, then docks against an antigen or feeds the structure into a design step. To compare a prediction against a crystal structure or another model, measure backbone agreement with the RMSD calculator.