ABodyBuilder3¶

Antibody structure prediction model that predicts 3D Fv region coordinates from paired heavy/light chain sequences using a GNN architecture with optional ProtT5 language model embeddings.

License: Apache-2.0 · Molecules: antibody · Tasks: structure_prediction

Actions: fold · Variants: 2

At a glance¶

Use it when

You need rapid Fv structure prediction for paired antibody sequences to support structure-based CDR design, paratope identification, or epitope mapping
You want to provide predicted structures as input for AntiFold inverse folding to design structurally compatible CDR variant libraries
You need per-residue confidence scores (pLDDT) to assess which regions of an antibody structure are reliably predicted before using them for downstream analysis
You are characterizing an antibody library at structural level and need fast throughput without MSA computation -- the pLDDT variant runs on CPU for high-volume screening
You want the highest-accuracy antibody-specific structure prediction available (language variant) for a small number of high-priority candidates
You need predicted Fv structures as input for structure-based developability feature extraction (biophysical descriptors that require 3D atomic coordinates)

Strengths

Two complementary variants: a fast CPU-only pLDDT variant for high-throughput screening and a language-enhanced L40S GPU variant that leverages ProtT5 embeddings for improved CDR accuracy
Antibody-specialized GNN architecture trained on SAbDab crystallographic structures, achieving state-of-the-art CDR loop prediction accuracy among antibody-specific methods at time of publication
Per-residue pLDDT confidence scores (0-100 scale) enable users to identify well-predicted framework regions versus uncertain CDR loops, guiding downstream interpretation
Predicts full-atom Fv structures (backbone and side chains) in PDB format, directly usable as input for structure-based design tools (AntiFold and molecular docking)
Language variant uses ProtT5 protein language model embeddings for enhanced accuracy, particularly in CDR-H3 where sequence diversity makes purely structural approaches less effective
Fast inference compared to AlphaFold2: single-sequence prediction without MSA computation, enabling rapid structural characterization of antibody libraries
Apache-2.0 license with no commercial restrictions, from the same Exscientia/Oxford group that published IgBERT and IgT5
Published successor to the widely used ABodyBuilder lineage (ABodyBuilder, ABodyBuilder2), with documented accuracy improvements

Limitations

Requires paired heavy and light chain sequences; cannot predict nanobody (VHH) structures -- use ImmuneBuilder (NanoBodyBuilder2) or ImmuneFold for nanobodies
Predicts Fv region only; constant domains (CH1, CL, Fc) are not modeled, limiting utility for full-length antibody structural analysis
Does not predict antibody-antigen complex structures; only models the unbound Fv -- use ImmuneFold for complex prediction with antigen context
Cannot predict TCR structures; antibody-only model -- use ImmuneBuilder or ImmuneFold for TCR structure prediction
Language variant requires L40S GPU (48 GB VRAM), making it one of the more expensive models to deploy in this catalog
CDR-H3 prediction remains challenging despite improvements; pLDDT scores in this region are typically lower and predictions should be interpreted with caution for long CDR-H3 loops
Single action (fold) with no embedding or sequence scoring capabilities -- purely a structure prediction tool

Reach for something else when

You need to predict nanobody (VHH) structures -- use ImmuneBuilder's NanoBodyBuilder2 variant, which is specifically trained on single-domain antibody structures
You need to predict antibody-antigen complex structures with bound antigen context -- use ImmuneFold, which accepts antigen PDB input for complex modeling
You need TCR structure prediction -- use ImmuneBuilder (TCRBuilder2/TCRBuilder2Plus) or ImmuneFold (TCR variant)
You need the broadest immune protein coverage (antibodies + nanobodies + TCRs) in a single model -- use ImmuneBuilder, which has dedicated sub-models for all three
You need ESM-2-enhanced accuracy and are working with antibodies or TCRs -- use ImmuneFold, which uses ESM-2 3B representations for potentially higher CDR accuracy
You need sequence-level analysis (embeddings, scoring, generation) rather than structure prediction -- use AbLang2, IgBERT, or IgT5
You need general protein structure prediction for non-antibody targets -- use ESMFold or Chai-1

Alternatives

Model	Better when	Worse when
immunebuilder	Covers antibodies, nanobodies (VHH), and TCRs in one family with CPU-only inference for all variants; use when you need nanobody/TCR folding or a single endpoint for mixed immune proteins.	Its ABodyBuilder2 antibody sub-model predates ABodyBuilder3 and has lower CDR accuracy; ABodyBuilder3's ProtT5 language variant is the accuracy upgrade for paired antibodies.
immunefold	Uses ESM-2 3B representations for strong CDR-H3 accuracy and supports antibody-antigen complex prediction from an antigen PDB.	Requires a T4 GPU and is newer with less community validation; ABodyBuilder3's pLDDT variant offers CPU-only speed for high-throughput screening.
esmfold	General protein structure predictor that accepts the same antibody chains (as a ':'-separated complex) and folds them with no antibody-specific setup; also handles arbitrary non-antibody chains in the same job.	Not antibody-specialized -- lower CDR (especially CDR-H3) accuracy and no antibody-region-calibrated confidence; ABodyBuilder3 is faster and more accurate for standalone Fv prediction.
chai1	AlphaFold3-class biomolecular predictor that can fold antibody chains together with antigens, ligands, or nucleic acids from the same sequence input for full-complex context.	Heavier and slower, GPU-bound, and not antibody-specialized; ABodyBuilder3 is the faster, more accurate choice for unbound paired-Fv structure prediction.

API & schema¶

Variants

Variant	Endpoint slug	GPU	CPU	Memory
language	`abodybuilder3-language`	l40s	4.0	12 GB
plddt	`abodybuilder3-plddt`	CPU	2.0	8 GB

Call an action with POST /api/v1/{slug}/{action} — the request envelope is {"items": [...], "params": {...}} and a success returns {"results": [...]}. See the HTTP API page for the base URL, error shape, and full contract.

`fold`¶

Call it

curl -X POST http://127.0.0.1:8000/api/v1/abodybuilder3-language/fold \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "heavy_chain": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIE",
      "light_chain": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIE"
    }
  ]
}'

Request — AbodyBuilder3FoldRequest

Field	Type	Required	Constraints	Description
`params`	AbodyBuilder3FoldRequestParams	no	default `{'plddt': False, 'seed': 42}`	Optional parameters controlling this action (defaults are used when omitted).
`items`	list[AbodyBuilder3FoldRequestItem]	yes	items 1–4	Batch of inputs to process in a single request. Up to 4 sequence pairs per request.

Nested types

AbodyBuilder3FoldRequestItem

Field	Type	Required	Constraints	Description
`heavy_chain`	string	yes	len 1–2048	Antibody heavy-chain amino-acid sequence.
`light_chain`	string	yes	len 1–2048	Antibody light-chain amino-acid sequence.

AbodyBuilder3FoldRequestParams

Field	Type	Required	Constraints	Description
`plddt`	boolean	no	default `False`	Whether to return per-residue pLDDT confidence scores in the response.
`seed`	integer \| null	no	default `42`	Random seed for reproducible sampling.

Raw JSON Schema

{
  "$defs": {
    "AbodyBuilder3FoldRequestItem": {
      "additionalProperties": false,
      "properties": {
        "heavy_chain": {
          "description": "Antibody heavy-chain amino-acid sequence.",
          "maxLength": 2048,
          "minLength": 1,
          "title": "Heavy Chain",
          "type": "string"
        },
        "light_chain": {
          "description": "Antibody light-chain amino-acid sequence.",
          "maxLength": 2048,
          "minLength": 1,
          "title": "Light Chain",
          "type": "string"
        }
      },
      "required": [
        "heavy_chain",
        "light_chain"
      ],
      "title": "AbodyBuilder3FoldRequestItem",
      "type": "object"
    },
    "AbodyBuilder3FoldRequestParams": {
      "additionalProperties": false,
      "properties": {
        "plddt": {
          "default": false,
          "description": "Whether to return per-residue pLDDT confidence scores in the response.",
          "title": "Plddt",
          "type": "boolean"
        },
        "seed": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": 42,
          "description": "Random seed for reproducible sampling.",
          "title": "Seed"
        }
      },
      "title": "AbodyBuilder3FoldRequestParams",
      "type": "object"
    }
  },
  "additionalProperties": false,
  "properties": {
    "params": {
      "$ref": "#/$defs/AbodyBuilder3FoldRequestParams",
      "default": {
        "plddt": false,
        "seed": 42
      },
      "description": "Optional parameters controlling this action (defaults are used when omitted)."
    },
    "items": {
      "description": "Batch of inputs to process in a single request. Up to 4 sequence pairs per request.",
      "items": {
        "$ref": "#/$defs/AbodyBuilder3FoldRequestItem"
      },
      "maxItems": 4,
      "minItems": 1,
      "title": "Items",
      "type": "array"
    }
  },
  "required": [
    "items"
  ],
  "title": "AbodyBuilder3FoldRequest",
  "type": "object"
}

Response — AbodyBuilder3FoldResponse

Field	Type	Required	Constraints	Description
`results`	list[AbodyBuilder3FoldResponseResult]	yes		Per-input results, returned in the same order as the request items.

Nested types

AbodyBuilder3FoldResponseResult

Field	Type	Required	Constraints	Description
`pdb`	string	yes		Predicted structure in PDB format.
`plddt`	list[number] \| null	no		Per-residue pLDDT confidence score (0–100; higher is more confident).

Raw JSON Schema

{
  "$defs": {
    "AbodyBuilder3FoldResponseResult": {
      "exclude_none": true,
      "exclude_unset": true,
      "properties": {
        "pdb": {
          "description": "Predicted structure in PDB format.",
          "title": "Pdb",
          "type": "string"
        },
        "plddt": {
          "anyOf": [
            {
              "items": {
                "type": "number"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Per-residue pLDDT confidence score (0\u2013100; higher is more confident).",
          "title": "Plddt"
        }
      },
      "required": [
        "pdb"
      ],
      "title": "AbodyBuilder3FoldResponseResult",
      "type": "object"
    }
  },
  "properties": {
    "results": {
      "description": "Per-input results, returned in the same order as the request items.",
      "items": {
        "$ref": "#/$defs/AbodyBuilder3FoldResponseResult"
      },
      "title": "Results",
      "type": "array"
    }
  },
  "required": [
    "results"
  ],
  "title": "AbodyBuilder3FoldResponse",
  "type": "object"
}

Usage¶

One-line summary: Antibody structure prediction model that predicts 3D Fv region coordinates from paired heavy/light chain sequences using a GNN architecture with optional ProtT5 language model embeddings.

Overview¶

AbodyBuilder3 is an antibody structure prediction model developed by Exscientia (Kenlay et al. 2024). Given paired heavy (H) and light (L) chain amino acid sequences, it predicts the 3D atomic coordinates of the antibody Fv region and outputs a PDB structure file. It offers two model variants: a "language" variant that incorporates ProtT5 protein language model embeddings for higher accuracy, and a "plddt" variant that provides faster inference with confidence estimation.

Architecture¶

Property	Value
Architecture	Graph Neural Network (GNN)
Language variant	GNN + ProtT5 language model embeddings (GPU required)
pLDDT variant	GNN with confidence estimation (CPU-only)
Framework	PyTorch Lightning (LitABB3)
Input	Paired H and L chain amino acid sequences
Output	PDB structure string, optional pLDDT scores

Capabilities & Limitations¶

CAN be used for: - Predicting 3D Fv structures from paired heavy/light chain sequences - Obtaining per-residue pLDDT confidence scores (via params.plddt=true) - Generating PDB files for downstream structure-based analysis - Batch processing up to 4 antibody sequence pairs per request - Deterministic predictions with configurable random seed

CANNOT be used for: - Single-chain inputs or nanobody (VHH) structure prediction - Antibody-antigen complex structure prediction - Constant region (Fc/CH/CL) structure prediction - General protein structure prediction (antibody-specific only) - Sequences containing non-standard amino acids

Usage Examples¶

Predict antibody structure¶

from models.abodybuilder3.schema import (
    AbodyBuilder3FoldRequest,
    AbodyBuilder3FoldRequestItem,
    AbodyBuilder3FoldRequestParams,
)

request = AbodyBuilder3FoldRequest(
    params=AbodyBuilder3FoldRequestParams(plddt=False, seed=42),
    items=[
        AbodyBuilder3FoldRequestItem(
            heavy_chain="EVQLVESGGGLVQPGGSLRLSCAASGFTFSSYAMSWVRQAPGKGLEWVSAISGSGGSTYYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCAR",
            light_chain="DIQMTQSPSSLSASVGDRVTITCRASQSISSYLNWYQQKPGKAPKLLIYAASSLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQSYSTPLT",
        )
    ],
)

Predict structure with pLDDT confidence scores¶

from models.abodybuilder3.schema import (
    AbodyBuilder3FoldRequest,
    AbodyBuilder3FoldRequestItem,
    AbodyBuilder3FoldRequestParams,
)

request = AbodyBuilder3FoldRequest(
    params=AbodyBuilder3FoldRequestParams(plddt=True, seed=42),
    items=[
        AbodyBuilder3FoldRequestItem(
            heavy_chain="QVQLQQSGPGLVKPSQTLSLTCAISGDSVSSNSAAWNWIRQSPSRGLEWLGRTYYRSKWYNDYAVSVKSRITINPDTSKNQFSLQLNSVTPEDTAVYYCAR",
            light_chain="EIVLTQSPGTLSLSPGERATLSCRASQSVSSSYLAWYQQKPGQAPRLLIYGASSRATGIPDRFSGSGSGTDFTLTISRLEPEDFAVYYCQQYGSSPRT",
        )
    ],
)

Architecture & training¶

Architecture¶

Model Type & Innovation¶

AbodyBuilder3 is an antibody structure prediction model developed by Exscientia (Kenlay et al. 2024). It predicts the 3D structure of antibody Fv regions (paired heavy and light chains) from sequence alone. The key innovation is a scalable architecture that achieves accuracy comparable to AlphaFold2-based methods while being significantly faster and more resource-efficient.

AbodyBuilder3 uses a graph neural network (GNN) architecture that processes antibody sequence features to predict backbone and side-chain atom coordinates. It operates in two variants: a "language" variant that incorporates ProtT5 protein language model embeddings for higher accuracy, and a "plddt" variant that uses a lighter-weight approach with confidence estimation (pLDDT scores).

Parameters & Layers¶

Component	Details
Architecture	Graph Neural Network (GNN)
Language variant	GNN + ProtT5 language model embeddings
pLDDT variant	GNN with confidence estimation (no language model)
Input	Paired `heavy_chain` and `light_chain` amino acid sequences
Output	PDB structure string, optional per-residue pLDDT scores
Framework	PyTorch Lightning (LitABB3 module)

Training Data¶

Property	Details
Source	Structural Antibody Database (SAbDab)
Structure type	Antibody Fv region crystal structures
Training approach	Two-stage training with checkpoint selection

Loss Function & Objective¶

AbodyBuilder3 uses a structure prediction loss that optimizes predicted atom positions against experimentally determined crystal structures. The two-stage training process refines the model progressively, with the best checkpoint from the second stage selected for deployment.

Tokenization / Input Processing¶

Input format: Paired heavy_chain and light_chain amino acid sequences (strings)
Validation: Extended amino acid alphabet
Maximum length: 2048 residues per chain
Language variant: Sequences are additionally processed through ProtT5 to obtain per-residue language model embeddings, which are concatenated with the GNN input features
Graph construction: Sequences are converted to graph representations using string_to_input utility from the AbodyBuilder3 library

Performance & Benchmarks¶

Published Benchmarks¶

From Kenlay et al., Bioinformatics (2024):

Model	CDR-H3 RMSD (A)	Overall RMSD (A)	Speed
AbodyBuilder3 (language)	Competitive	Competitive	Fast
AlphaFold2 (antibody mode)	Best	Best	Slow
IgFold	Moderate	Moderate	Moderate
ABlooper	CDR-only	N/A	Fast

See sources.yaml applied_literature for per-CDR backbone RMSD from published head-to-head benchmarks (e.g. Dreyer et al. mAbs 2025, BioGeometry GeoFlow-V2 2025).

BioLM Verification Results¶

Test Case	Action	Tolerance	Status
Standard antibody pair	fold	rel_tol 1e-3, cosine_distance < 0.02, pdb_rmsd < 0.05 A	PASS

Comparison to Alternatives¶

Model	Strength	When to prefer
AbodyBuilder3	Fast, accurate antibody structure prediction	High-throughput antibody modeling
AlphaFold2	Highest accuracy for general proteins	When maximum accuracy needed, speed not critical
IgFold	Good antibody structure prediction	Alternative antibody-specific predictor
ESMFold	Fast general protein structure	Non-antibody proteins
Boltz	State-of-the-art biomolecular structure	Complex biomolecular assemblies

Strengths & Limitations¶

Pros¶

Fast antibody structure prediction relative to AlphaFold2-based methods
Two variants allow accuracy/speed tradeoff (language vs pLDDT)
Outputs standard PDB format directly
Optional pLDDT confidence scores for per-residue quality assessment
Deterministic predictions (seeded for reproducibility)

Cons¶

Limited to antibody Fv regions (heavy and light chain pairs only)
Does not predict antigen binding or antibody-antigen complex structures
Language variant requires GPU (L40S) for ProtT5 embedding computation
Does not support nanobodies (single-chain antibodies)
Batch size limited to 4 per request

Known Failure Modes¶

Unusual CDR-H3 loop conformations (very long or highly constrained) may have lower prediction accuracy
Sequences with non-standard residues may cause processing errors
The pLDDT variant may produce lower-quality structures compared to the language variant

Implementation Details¶

Inference Pipeline¶

Request
  |-- 1. Validate paired H/L chain sequences
  |-- 2. Seed all random number generators for determinism
  |-- 3. Convert sequences to graph input (string_to_input)
  |-- 4. [language variant] Compute ProtT5 embeddings
  |-- 5. Batch input and transfer to device
  |-- 6. Forward pass through GNN model
  |-- 7. Add atom37 representation to output
  |-- 8. Convert output to PDB string
  |-- 9. [if plddt=True] Extract per-residue pLDDT scores
  |-- 10. Format and return response

Memory & Compute Profile¶

Resource	Language Variant	pLDDT Variant
GPU	L40S (48 GB VRAM)	None (CPU-only)
Memory	12 GB RAM	8 GB RAM
CPU	4.0 cores	2.0 cores
Batch size	4	4

Determinism & Reproducibility¶

Setting	Value
Torch manual seed	Yes (42 at model load, configurable per request)
CUDA manual seed	Yes (42)
NumPy seed	Yes (per request)
PyTorch Lightning seed	Yes (per request)
cuDNN deterministic	Yes
cuDNN benchmark	Disabled
Request-level seed	Configurable via `params.seed` (default: 42)

Caching Behavior¶

Response caching is handled by the serving layer, not within the model container itself: - Cloudflare R2 caching for persistence across requests - Cache keys determined by the full request payload (sequences + parameters)

Versions & Changelog¶

Version	Date	Changes
v1	2024	Initial implementation with language and plddt variants

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

AbodyBuilder3 is designed for antibody Fv (variable fragment) structures. It takes paired heavy chain (VH) and light chain (VL) amino acid sequences as input and predicts the 3D atomic coordinates of the Fv region.

Important coverage notes: - Requires both heavy and light chain sequences (paired input) - Predicts the Fv region only, not constant domains (CH1, CL, Fc) - Does not support nanobodies (VHH) or single-chain inputs - Does not predict antibody-antigen complex structures

Cross-Applicability¶

Molecule Type	Applicability	Evidence	Caveats
Conventional antibodies (VH/VL)	High	Primary training target (SAbDab structures)	Fv region only
Nanobodies (VHH)	Not applicable	Requires paired heavy/light input	Use alternative structure predictors
Fab fragments	Partial	Predicts Fv portion only	Constant domains not modeled
T-cell receptors (TCRs)	Low	Not trained on TCR structures	Structural similarity exists but untested
General proteins	Not applicable	Antibody-specific architecture	Use ESMFold, Chai-1, or AlphaFold2

Biological Problems Addressed¶

Antibody Structure Prediction (Published)¶

Biological context: Understanding the 3D structure of an antibody is critical for rational engineering of binding affinity, specificity, and developability. Experimental structure determination by X-ray crystallography or cryo-EM is expensive and slow (months per structure). Computational structure prediction enables rapid hypothesis generation and design iteration.

The most challenging aspect of antibody structure prediction is the CDR-H3 loop, which is the most diverse in both sequence and structure among all antibody regions. CDR-H3 is often the primary determinant of antigen specificity, and its conformation cannot be reliably predicted from sequence templates alone.

How AbodyBuilder3 helps: Given paired heavy and light chain sequences, AbodyBuilder3 predicts the full Fv backbone and side-chain coordinates. The language variant leverages ProtT5 protein language model embeddings for enhanced accuracy, particularly in the CDR regions. The output is a standard PDB file that can be directly used for downstream structure-based analysis and design.

Output interpretation: The PDB output contains predicted atom coordinates with standard chain and residue labeling. When pLDDT scores are requested, per-residue confidence values (0--100 scale) indicate the model's confidence in each position. Regions with pLDDT > 70 are generally reliable; regions below 50 should be interpreted with caution, particularly in the CDR-H3 loop.

Structure-Based Antibody Engineering (Published)¶

Biological context: Many antibody engineering tasks -- including CDR grafting, humanization, affinity maturation, and developability optimization -- benefit from structural information. Knowing the spatial arrangement of CDR loops relative to the framework scaffold guides the selection of mutations that improve function without destabilizing the fold.

How AbodyBuilder3 helps: Predicted structures serve as input for structure-based design tools (e.g., AntiFold for inverse folding, molecular docking for binding prediction). Having rapid access to antibody structures enables high-throughput structure-based screening of antibody libraries.

Virtual Antibody Library Characterization (Anticipated)¶

Biological context: Antibody discovery campaigns produce large sequence libraries from phage display, yeast display, or B-cell sequencing. Understanding the structural diversity of these libraries -- particularly the CDR loop conformations -- helps assess library quality and guide selection strategies.

How AbodyBuilder3 helps: The fast inference speed enables structure prediction for hundreds or thousands of antibody sequences, providing structural annotations at library scale. This can reveal CDR-H3 loop conformation clusters, identify structurally redundant candidates, and prioritize structurally diverse subsets for experimental characterization. However, this application has not yet been validated at scale.

Applied Use Cases¶

AbodyBuilder3 was published in 2024. The following applied literature benchmarks or builds on this model:

Dreyer et al. (mAbs, 2025) — Ibex (Prescient Design/Genentech) benchmarks ABodyBuilder3 head-to-head on the ImmuneBuilder test set; reports CDR H3 backbone RMSD of 2.86 Å, CDR L3 RMSD of 1.13 Å for ABodyBuilder3.
BioGeometry Team (bioRxiv, 2025) — GeoFlow-V2-ab benchmarks ABodyBuilder3 alongside AlphaFold-Multimer V2.3 on 205 antibody test structures; ABodyBuilder3 serves as the antibody-specific baseline.
Ali et al. (bioRxiv, 2025) — Adopts the ABodyBuilder3 architecture as the foundation for nanobody-specific structure predictors via self-distillation from unlabelled VHH sequences.
Ali et al. (bioRxiv, 2026) — NbForge, a nanobody folding model with blueprint- and disulphide-aware inductive biases, is explicitly described as "derived from ABodyBuilder3/AlphaFold2."

Predecessor Models¶

AbodyBuilder (Leem et al., 2016): The original antibody structure prediction tool using homology modeling with SAbDab templates. AbodyBuilder3 replaces the template-based approach with a learned GNN architecture.
AbodyBuilder2: Intermediate version with improved template selection and loop modeling.

Complementary Models¶

AbodyBuilder3 works well in combination with other models in this catalog:

AntiFold: For inverse folding on predicted structures. Pipeline: predict structure with AbodyBuilder3, then design sequences with AntiFold.
AbLEF: For sequence-level developability screening. Pipeline: predict structure for visualization, assess developability from sequence with AbLEF.
ESM2: For general protein representation. Embeddings from ESM2 can complement structural information from AbodyBuilder3.

Alternative Models¶

Alternative	Advantage over AbodyBuilder3	Disadvantage vs AbodyBuilder3
AlphaFold2 (antibody mode)	Higher accuracy, especially CDR-H3	Much slower, requires MSA
IgFold	Antibody-specific, fast	Different accuracy/speed tradeoff
ESMFold	General protein coverage	Not antibody-specialized
Boltz	State-of-the-art general structure	Slower, designed for complexes

Biological Background¶

Antibody Structure¶

Antibodies are Y-shaped proteins with two identical heavy chains and two identical light chains. The antigen-binding site is formed by the variable domains of the heavy chain (VH) and light chain (VL), collectively known as the Fv (variable fragment) region.

Variable domain architecture:

VH Domain
  |-- Framework 1 (FR1)
  |-- CDR-H1
  |-- Framework 2 (FR2)
  |-- CDR-H2
  |-- Framework 3 (FR3)
  |-- CDR-H3  <-- most variable, hardest to predict
  |-- Framework 4 (FR4)

VL Domain
  |-- Framework 1 (FR1)
  |-- CDR-L1
  |-- Framework 2 (FR2)
  |-- CDR-L2
  |-- Framework 3 (FR3)
  |-- CDR-L3
  |-- Framework 4 (FR4)

The framework regions form a conserved beta-sandwich scaffold (immunoglobulin fold), while the CDR loops protrude from this scaffold to form the antigen-binding surface (paratope).

CDR-H3 Loop Prediction Challenge¶

CDR-H3 is the most sequence-diverse and structurally variable of all six CDR loops. It is formed by V-D-J recombination with junctional diversity, producing loop lengths ranging from 1 to over 30 residues. Unlike CDR-H1, CDR-H2, and the light chain CDRs, CDR-H3 conformations cannot be reliably predicted from canonical class assignments. This makes CDR-H3 structure prediction the primary benchmark for antibody modeling methods.

pLDDT Confidence Scores¶

Predicted Local Distance Difference Test (pLDDT) is a per-residue confidence metric (0--100 scale) that estimates the accuracy of predicted atom positions. Originally introduced with AlphaFold2, pLDDT has become a standard measure for structure prediction confidence. For antibodies, framework regions typically have high pLDDT (>80), while CDR loops -- especially CDR-H3 -- often have lower pLDDT reflecting genuine structural uncertainty.

Sources & license¶

License: Apache-2.0 (text)

Papers

ABodyBuilder3: Improved and scalable antibody structure predictions — Bioinformatics, 2024 · DOI arXiv

Source repositories

github: https://github.com/Exscientia/abodybuilder3

Cite

@article{kenlay2024abodybuilder3,
  title={ABodyBuilder3: Improved and scalable antibody structure predictions},
  author={Kenlay, Henry and Dreyer, Frédéric A and Krawczyk, Konrad and Sherborne, Bryn and Deane, Charlotte M},
  journal={Bioinformatics},
  year={2024},
  doi={10.1093/bioinformatics/btae576}
}

ABodyBuilder3¶

At a glance¶

API & schema¶

fold¶

Usage¶

Overview¶

Architecture¶

Capabilities & Limitations¶

Usage Examples¶

Predict antibody structure¶

Predict structure with pLDDT confidence scores¶

Architecture & training¶

Architecture¶

Model Type & Innovation¶

Parameters & Layers¶

Training Data¶

Loss Function & Objective¶

Tokenization / Input Processing¶

Performance & Benchmarks¶

Published Benchmarks¶

BioLM Verification Results¶

Comparison to Alternatives¶

Strengths & Limitations¶

Pros¶

Cons¶

Known Failure Modes¶

Implementation Details¶

Inference Pipeline¶

Memory & Compute Profile¶

Determinism & Reproducibility¶

Caching Behavior¶

Versions & Changelog¶

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

Cross-Applicability¶

Biological Problems Addressed¶

Antibody Structure Prediction (Published)¶

Structure-Based Antibody Engineering (Published)¶

Virtual Antibody Library Characterization (Anticipated)¶

Applied Use Cases¶

Related Models¶

Predecessor Models¶

Complementary Models¶

Alternative Models¶

Biological Background¶

Antibody Structure¶

CDR-H3 Loop Prediction Challenge¶

pLDDT Confidence Scores¶

Sources & license¶

`fold`¶