ImmuneBuilder vs ABodyBuilder vs IgFold: Antibody Structure Prediction Compared

Why Antibody-Specific Structure Prediction Matters

Antibodies have a unique structural challenge that general-purpose protein structure prediction tools struggle with. The framework regions are highly conserved and easy to predict, but the complementarity-determining regions (CDRs) – especially CDR-H3 – are hypervariable loops that adopt diverse conformations. These loops are precisely where accuracy matters most, because they determine antigen binding specificity and affinity.

General tools like AlphaFold2 and ESMFold produce reasonable antibody framework predictions but often fail on CDR loop conformations. This gap motivated the development of antibody-specific predictors: ImmuneBuilder, ABodyBuilder2, and IgFold. Each takes a different approach to the problem, with distinct tradeoffs in accuracy, speed, and accessibility.

In this comparison, we evaluate all three tools head-to-head on accuracy benchmarks, CDR loop quality, inference speed, and practical API access in 2026.

The Three Contenders

ImmuneBuilder

ImmuneBuilder, developed at the University of Oxford, uses a deep learning architecture specifically designed for immunoglobulin prediction. It handles antibodies (ABodyBuilder module), nanobodies (NanoBodyBuilder module), and T-cell receptors (TCRBuilder module) through specialized sub-models.

• Architecture: Equivariant graph neural network with antibody-specific training
• Input: Heavy and light chain amino acid sequences (paired or unpaired)
• Training data: SAbDab (Structural Antibody Database) – the largest curated antibody structure database
• Speed: 1–3 minutes per antibody on GPU
• Special features: Confidence scores per residue, handles nanobodies and TCRs, outputs IMGT-numbered PDB
• License: BSD 3-Clause – open source, commercial use permitted

ABodyBuilder2

ABodyBuilder2, also from Oxford (the same group behind the original ABodyBuilder), combines MSA-based features with a structure module derived from AlphaFold2. It represents the most computationally intensive approach of the three, trading speed for accuracy on difficult CDR-H3 loops.

• Architecture: AlphaFold2-derived with antibody-specific Evoformer modifications
• Input: Heavy and light chain sequences + MSA (auto-generated)
• Training data: SAbDab + general PDB structures
• Speed: 5–15 minutes (MSA generation dominates)
• Special features: Highest accuracy on long CDR-H3 loops, pLDDT-style confidence, ensemble predictions
• License: Apache 2.0 – open source, commercial use permitted

IgFold

IgFold, from Johns Hopkins University, uses a protein language model (AntiBERTy) to predict antibody structures without requiring MSA generation. This makes it the fastest of the three and the most similar in philosophy to ESMFold – trading MSA-based accuracy for speed and simplicity.

• Architecture: AntiBERTy language model + IPA (Invariant Point Attention) structure module
• Input: Heavy and light chain sequences (no MSA required)
• Training data: Paired antibody sequences from the OAS (Observed Antibody Space) database
• Speed: 30–60 seconds per antibody (fastest of the three)
• Special features: No MSA needed, handles single chains well, lightweight model
• License: BSD 3-Clause – open source, commercial use permitted

Head-to-Head Accuracy Comparison

The following benchmarks are from published results on the RosettaAntibodyBenchmark2 dataset and independent evaluations on recent SAbDab entries. RMSD values are in Angstroms (lower is better).

Framework Region Accuracy

All three tools predict antibody framework regions with near-experimental accuracy. The frameworks are conserved beta-sheet structures that are well-represented in training data.

• ImmuneBuilder: Framework RMSD 0.4–0.6 Angstrom
• ABodyBuilder2: Framework RMSD 0.3–0.5 Angstrom
• IgFold: Framework RMSD 0.5–0.7 Angstrom

CDR Loop Accuracy (Excluding H3)

CDR-L1, L2, L3, H1, and H2 adopt canonical conformations that are well-characterized in the structural antibody databases. All three tools handle these loops well:

• ImmuneBuilder: Median CDR RMSD 0.7–1.2 Angstrom across L1/L2/L3/H1/H2
• ABodyBuilder2: Median CDR RMSD 0.6–1.1 Angstrom
• IgFold: Median CDR RMSD 0.8–1.3 Angstrom

CDR-H3: The Critical Test

CDR-H3 is the most variable loop in the antibody and the hardest to predict accurately. It has the fewest canonical conformations, the widest length distribution (3 to 30+ residues), and the greatest structural diversity. This is where tool choice matters most.

• ImmuneBuilder: Median CDR-H3 RMSD 1.8–2.5 Angstrom (length-dependent)
• ABodyBuilder2: Median CDR-H3 RMSD 1.9–2.6 Angstrom, best on long loops (above 15 residues)
• IgFold: Median CDR-H3 RMSD 2.2–3.0 Angstrom, less accurate on long H3 loops
• AlphaFold2 (for reference): Median CDR-H3 RMSD 3.0–4.5 Angstrom

The key takeaway: all three antibody-specific tools significantly outperform general-purpose predictors like AlphaFold2 on CDR-H3. ImmuneBuilder and ABodyBuilder2 are the most accurate, with IgFold trading some accuracy for speed.

VH-VL Orientation

The relative orientation of the heavy and light chain variable domains affects antigen binding geometry. Predicting this orientation accurately is important for downstream docking and engineering tasks.

• ImmuneBuilder: VH-VL orientation error 2.1 degrees (median)
• ABodyBuilder2: VH-VL orientation error 1.8 degrees (median)
• IgFold: VH-VL orientation error 2.5 degrees (median)

Speed and Infrastructure Comparison

Inference speed determines what is practical at scale. If you need to predict structures for thousands of antibody variants from a phage display campaign, the time per prediction directly impacts your project timeline.

• IgFold: ~45 seconds (no MSA, language model only) – fastest
• ImmuneBuilder: ~2 minutes (no MSA, graph neural network) – fast
• ABodyBuilder2: ~10 minutes (MSA generation + inference) – slowest

Throughput for 1,000 Antibody Variants

• IgFold: ~12.5 hours on a single GPU
• ImmuneBuilder: ~33 hours on a single GPU
• ABodyBuilder2: ~7 days on a single GPU (MSA bottleneck)

Running ImmuneBuilder Through SciRouter

SciRouter provides ImmuneBuilder through its antibody structure prediction API. Here is how to predict an antibody structure and assess its CDR loop quality:

import scirouter

client = scirouter.SciRouter()

# Trastuzumab (Herceptin) -- anti-HER2 therapeutic antibody
# Heavy chain variable region
heavy_chain = (
    "EVQLVESGGGLVQPGGSLRLSCAASGFNIKDTYIHWVRQAPGKGLEWVARI"
    "YPTNGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSRWGGD"
    "GFYAMDYWGQGTLVTVSS"
)

# Light chain variable region
light_chain = (
    "DIQMTQSPSSLSASVGDRVTITCRASQDVNTAVAWYQQKPGKAPKLLIYS"
    "ASFLYSGVPSRFSGSRSGTDFTLTISSLQPEDFATYYCQQHYTTPPTFGQG"
    "TKVEIK"
)

# Predict structure
job = client.antibodies.fold(
    heavy_chain=heavy_chain,
    light_chain=light_chain,
    model="immunebuilder",
)
result = client.antibodies.fold_result(job.job_id, poll=True)

print(f"Overall confidence: {result.confidence:.2f}")
print(f"\nCDR loop confidence scores:")
for cdr, score in result.cdr_scores.items():
    quality = "Excellent" if score > 85 else "Good" if score > 70 else "Uncertain"
    print(f"  {cdr}: {score:.1f} ({quality})")

# Save structure
with open("trastuzumab_fv.pdb", "w") as f:
    f.write(result.pdb)
print("\nStructure saved to trastuzumab_fv.pdb")

Comparing Multiple Antibodies

When evaluating a panel of antibody candidates, run structure prediction on all of them and compare CDR-H3 confidence scores. Higher confidence correlates with more reliable structures and easier downstream CDR design:

import scirouter

client = scirouter.SciRouter()

# Panel of anti-HER2 antibody candidates (variable regions)
candidates = [
    {
        "name": "Trastuzumab",
        "heavy": "EVQLVESGGGLVQPGGSLRLSCAASGFNIKDTYIHWVRQAPGKGLEWVARI"
                 "YPTNGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSRWGGD"
                 "GFYAMDYWGQGTLVTVSS",
        "light": "DIQMTQSPSSLSASVGDRVTITCRASQDVNTAVAWYQQKPGKAPKLLIYS"
                 "ASFLYSGVPSRFSGSRSGTDFTLTISSLQPEDFATYYCQQHYTTPPTFGQG"
                 "TKVEIK",
    },
    {
        "name": "Pertuzumab",
        "heavy": "EVQLVESGGGLVQPGGSLRLSCAASGFTISDAWIHWVRQAPGKGLEWVARI"
                 "SPAGGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCARDVGY"
                 "DLTYFNYWGQGTLVTVSS",
        "light": "DIQMTQSPSSLSASVGDRVTITCRASQSISSYLNWYQQKPGKAPKLLIYA"
                 "ASSLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQSYSTPLTFGQG"
                 "TKVEIK",
    },
    {
        "name": "Candidate-3",
        "heavy": "EVQLVESGGGLVQPGGSLRLSCAASGFTFSDYWMHWVRQAPGKGLVWVSR"
                 "INSDGSSTSYADSVKGRFTISRDNAKNTLYLQMNSLRAEDTAVYYCARGGD"
                 "YGYDYWGQGTLVTVSS",
        "light": "DIQMTQSPSSLSASVGDRVTITCRASQSISSYLNWYQQKPGKAPKLLIYA"
                 "ASSLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQSYSTPLTFGQG"
                 "TKVEIK",
    },
]

print(f"{'Antibody':<18} {'Confidence':>10} {'CDR-H3':>8} {'CDR-L3':>8}")
print("-" * 50)

for ab in candidates:
    job = client.antibodies.fold(
        heavy_chain=ab["heavy"],
        light_chain=ab["light"],
        model="immunebuilder",
    )
    result = client.antibodies.fold_result(job.job_id, poll=True)

    print(
        f"{ab['name']:<18} {result.confidence:>9.2f} "
        f"{result.cdr_scores.get('H3', 0):>7.1f} "
        f"{result.cdr_scores.get('L3', 0):>7.1f}"
    )

From Structure to Design: The Full Workflow

Structure prediction is typically the first step in a computational antibody engineering workflow. Once you have a predicted structure, you can use it for CDR redesign with AntiFold, docking against an antigen, or humanization scoring.

SciRouter's Antibody Design Studio chains ImmuneBuilder structure prediction with AntiFold CDR design into a single workflow:

import scirouter

client = scirouter.SciRouter()

# Step 1: Predict structure
fold_job = client.antibodies.fold(
    heavy_chain="EVQLVESGGGLVQPGGSLRLSCAASGFNIKDTYIHWVRQAPGKGLEWVARI"
                "YPTNGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSRWGGD"
                "GFYAMDYWGQGTLVTVSS",
    light_chain="DIQMTQSPSSLSASVGDRVTITCRASQDVNTAVAWYQQKPGKAPKLLIYS"
                "ASFLYSGVPSRFSGSRSGTDFTLTISSLQPEDFATYYCQQHYTTPPTFGQG"
                "TKVEIK",
    model="immunebuilder",
)
structure = client.antibodies.fold_result(fold_job.job_id, poll=True)
print(f"Structure predicted (confidence: {structure.confidence:.2f})")

# Step 2: Design CDR-H3 variants with AntiFold
design_result = client.antibodies.design(
    pdb=structure.pdb,
    target_cdrs=["H3"],
    num_designs=20,
    temperature=0.2,
    model="antifold",
)

print(f"\nGenerated {len(design_result.designs)} CDR-H3 variants:")
for i, d in enumerate(design_result.designs[:5]):
    print(f"  {i+1}. {d.cdr_h3_sequence} (score: {d.score:.3f})")

# Step 3: Validate top designs with ESMFold
print("\nValidating top 3 designs...")
for d in design_result.designs[:3]:
    val_job = client.proteins.fold(
        sequence=d.full_heavy_chain,
        model="esmfold",
    )
    val = client.proteins.fold_result(val_job.job_id, poll=True)
    print(f"  {d.cdr_h3_sequence}: pLDDT {val.mean_plddt:.1f}")

Decision Guide: Which Tool Should You Choose?

Choose ImmuneBuilder When:

• You need a good balance of accuracy and speed
• You want API access through SciRouter without managing infrastructure
• You are working with standard IgG antibodies (heavy + light chain)
• You need downstream CDR design with AntiFold (SciRouter chains them)
• You also need nanobody or TCR structure prediction (ImmuneBuilder handles all three)

Choose ABodyBuilder2 When:

• Maximum CDR-H3 accuracy is critical and speed is secondary
• You are predicting structures for a small number of high-value candidates (under 50)
• The CDR-H3 loop is unusually long (above 15 residues)
• You need ensemble predictions with confidence estimates

Choose IgFold When:

• You need to screen thousands of antibody sequences quickly
• Speed is the top priority and CDR-H3 accuracy can be approximate
• You are working with single-domain antibodies or nanobodies
• You prefer a lightweight tool with minimal dependencies

When to Use General Tools Instead

For antibody-antigen complex prediction (not just the antibody alone), consider Boltz-2 or Chai-1, which can predict multi-chain assemblies including the antigen-antibody interface. For quick single-chain screening without antibody-specific accuracy, ESMFold remains the fastest option.

Licensing and Access Summary

• ImmuneBuilder: BSD 3-Clause, available via SciRouter API and self-hosted
• ABodyBuilder2: Apache 2.0, available as a web server (opig.stats.ox.ac.uk) and self-hosted
• IgFold: BSD 3-Clause, available on GitHub and via pip install

All three tools are open source with commercial-friendly licenses. SciRouter provides managed API access to ImmuneBuilder and pairs it with AntiFold for a complete antibody structure-to-design pipeline – no GPU setup, no Docker containers, no database downloads.

Next Steps

Ready to predict antibody structures? Here is how to get started:

• Try Antibody Design Studio to run the full ImmuneBuilder + AntiFold pipeline in your browser
• Read the complete antibody design tutorial for a step-by-step walkthrough of the design workflow
• Explore stability prediction to score engineered antibody variants for thermostability

Sign up for a free SciRouter API key and predict your first antibody structure in under 3 minutes. 500 free credits per month, no credit card required.