Why Antibody Structure Prediction Matters
Antibody therapeutics are among the fastest-growing drug classes, with over 100 approved worldwide. Designing and optimizing these molecules requires accurate 3D structures, but experimental determination via X-ray crystallography or cryo-EM is slow and expensive. AI structure prediction fills this gap: given just the amino acid sequence of heavy and light chains, these tools predict the full antibody structure in seconds to minutes.
The critical challenge is predicting CDR loop conformations — the six hypervariable loops (H1, H2, H3, L1, L2, L3) that form the antigen-binding surface. CDR-H3, the most variable loop, is the hardest to model and the most important for binding specificity. General-purpose protein folders like AlphaFold2 handle antibody frameworks well but often struggle with CDR loops because they lack antibody-specific training data.
Three specialized tools have emerged to solve this problem: ImmuneBuilder, ABodyBuilder2, and IgFold. Each takes a different architectural approach, and each has distinct strengths. This guide compares them head-to-head so you can choose the right tool for your project.
The Three Contenders
ImmuneBuilder
ImmuneBuilder, developed at the Oxford Protein Informatics Group, uses an equivariant transformer architecture trained specifically on antibody structures from SAbDab (the Structural Antibody Database). It predicts structures for paired antibodies (ABodyBuilder2 module), nanobodies/VHH domains (NanoBodyBuilder2), and T-cell receptors (TCRBuilder2) — making it the most versatile of the three tools.
The equivariant attention mechanism ensures that predictions are rotationally and translationally invariant, which is important for accurately capturing the geometry of CDR loops without relying on multiple sequence alignments. ImmuneBuilder takes roughly 20 seconds per prediction on a modern GPU.
ABodyBuilder2
ABodyBuilder2 adapts the AlphaFold2 architecture for antibodies. It leverages the proven Evoformer and structure module from AlphaFold2 but retrains them on antibody-specific data from SAbDab. This means it benefits from AlphaFold2's powerful geometric reasoning while being tuned to the unique structural patterns of immunoglobulins.
Because it inherits the full AlphaFold2 pipeline — including MSA processing and template search — ABodyBuilder2 tends to be the slowest of the three, typically taking several minutes per prediction. However, this extra computation often translates to strong accuracy, particularly on framework regions and canonical CDR loops.
IgFold
IgFold takes a fundamentally different approach by using pre-trained protein language model embeddings (from AntiBERTy and ESM-2) as input features instead of MSAs. This makes it by far the fastest option — predictions complete in just a few seconds. The language model embeddings capture evolutionary information implicitly, without requiring an explicit sequence search step.
IgFold's architecture feeds these embeddings through an IPA (Invariant Point Attention) structure module, similar to AlphaFold2's structure module but without the Evoformer. This lightweight design is the key to its speed advantage.
Architecture Comparison
- ImmuneBuilder: Equivariant transformer with attention over residue pairs. No MSA required. Trained on SAbDab antibody structures. Separate modules for antibodies, nanobodies, and TCRs.
- ABodyBuilder2: AlphaFold2-based architecture with Evoformer and structure module, retrained on antibody data. Uses MSAs and structural templates. Highest computational cost.
- IgFold: Language model embeddings (AntiBERTy/ESM-2) fed into an IPA structure module. No MSA, no templates. Fastest inference by a significant margin.
Accuracy: CDR Loop RMSD Benchmarks
The standard benchmark for antibody structure prediction is backbone RMSD (root-mean-square deviation) on CDR loops, measured against experimentally solved crystal structures. Lower RMSD means the predicted loop is closer to the true conformation.
Framework Regions
All three tools predict antibody framework regions with high accuracy, typically below 1.0 angstrom backbone RMSD. This is expected — framework regions are structurally conserved across antibodies, so the prediction task is relatively straightforward. For framework-only applications (like antibody numbering or humanization assessment), any of the three tools will work well.
CDR Loops H1, H2, L1, L2, L3
For the five canonical CDR loops (all except H3), ImmuneBuilder and ABodyBuilder2 perform similarly, with median backbone RMSD values typically in the 0.5 to 1.5 angstrom range depending on the loop. IgFold is competitive but shows slightly higher variance. These loops adopt a limited set of canonical conformations, so all three tools handle them reasonably well.
CDR-H3: The Critical Test
CDR-H3 is where the tools diverge most significantly. This loop varies widely in length (4 to 30+ residues), lacks canonical structure classes, and often forms complex conformations including beta-hairpins and kinked bases. Published benchmarks show:
- ImmuneBuilder: Median CDR-H3 backbone RMSD around 2.0 to 2.8 angstroms, with strong performance on shorter H3 loops
- ABodyBuilder2: Similar median CDR-H3 RMSD of 2.0 to 2.5 angstroms, benefits from template information when available
- IgFold: Median CDR-H3 RMSD around 2.5 to 3.5 angstroms, slightly less accurate but much faster
For long CDR-H3 loops (greater than 15 residues), all tools struggle. This remains an open challenge in computational antibody modeling. When CDR-H3 accuracy is paramount — for example, in docking-based affinity prediction — consider running multiple tools and comparing predictions.
Speed Comparison
Speed matters when you are screening hundreds of antibody variants or integrating structure prediction into an automated pipeline. The three tools span a wide range:
- IgFold: Roughly 2 to 5 seconds per prediction. The language model embedding approach avoids MSA search entirely, making it ideal for high-throughput screening.
- ImmuneBuilder: Roughly 15 to 30 seconds per prediction on GPU. A good balance between speed and accuracy for most workflows.
- ABodyBuilder2: Roughly 2 to 5 minutes per prediction. The MSA and template search steps add significant overhead, though accuracy is competitive.
Nanobody and TCR Support
Beyond conventional paired antibodies, researchers increasingly work with nanobodies (single-domain VHH antibodies from camelids) and T-cell receptors (TCRs). Tool support varies significantly:
- ImmuneBuilder: Full support for nanobodies (NanoBodyBuilder2 module) and TCRs (TCRBuilder2 module). The only tool with dedicated models for all three molecule types.
- IgFold: Can accept single heavy chain inputs, making it usable for nanobodies in practice, though not specifically trained on VHH structures. No TCR support.
- ABodyBuilder2: Designed for paired heavy-light chain antibodies. Not recommended for nanobodies or TCRs.
If your work involves nanobodies or TCRs, ImmuneBuilder is the clear choice. The growing importance of nanobodies in diagnostics, imaging, and therapeutics makes this a meaningful differentiator.
When to Use Each Tool
- Choose ImmuneBuilder when you need a reliable all-purpose antibody predictor, when working with nanobodies or TCRs, or when CDR accuracy matters more than raw speed
- Choose ABodyBuilder2 when maximum accuracy on paired antibodies is the priority and you can afford longer runtimes — particularly when structural templates are available for similar antibodies
- Choose IgFold when speed is the priority — screening large variant libraries, rapid prototyping, or integrating predictions into real-time pipelines
SciRouter: ImmuneBuilder + AntiFold in One Workflow
SciRouter provides ImmuneBuilder as a managed API endpoint — no GPU setup, no dependency management, no model downloads. Send your sequences, get back structures with per-residue confidence scores. Combined with AntiFold for CDR sequence design, SciRouter offers a complete antibody predict-then-design pipeline through a single API.
We chose ImmuneBuilder as our primary antibody predictor because of its balance of accuracy, speed, and versatility (antibodies, nanobodies, and TCRs in one model). Paired with AntiFold for inverse folding, you can go from sequence to optimized CDR variants without leaving the API.
Predict an Antibody Structure
import os, requests, time
API_KEY = os.environ["SCIROUTER_API_KEY"]
BASE = "https://api.scirouter.ai/v1"
HEADERS = {"Authorization": f"Bearer {API_KEY}"}
# Predict paired antibody structure with ImmuneBuilder
job = requests.post(f"{BASE}/antibodies/structure", headers=HEADERS, json={
"model": "immunebuilder",
"heavy_chain": "EVQLVESGGGLVQPGGSLRLSCAASGFTFSDYWMHWVRQAPGKGLVWVSRIN"
"SDGSSTSYADSVKGRFTISRDNAKNTLYLQMNSLRAEDTAVYYCAR",
"light_chain": "DIQMTQSPSSLSASVGDRVTITCRASQSISSYLNWYQQKPGKAPKLLIYAAS"
"SLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQSYSTPLT",
}).json()
# Poll until complete
while True:
result = requests.get(
f"{BASE}/antibodies/structure/{job['job_id']}", headers=HEADERS
).json()
if result["status"] == "completed":
break
if result["status"] == "failed":
raise RuntimeError(result.get("error"))
time.sleep(3)
print(f"Overall confidence: {result['confidence']:.2f}")
for cdr, score in result["cdr_scores"].items():
print(f" {cdr} pLDDT: {score:.1f}")
with open("antibody.pdb", "w") as f:
f.write(result["pdb"])
print("Structure saved to antibody.pdb")Design Optimized CDR Variants with AntiFold
# Use the predicted structure to design new CDR-H3 sequences
design_resp = requests.post(f"{BASE}/antibodies/design", headers=HEADERS, json={
"model": "antifold",
"pdb": result["pdb"], # Structure from ImmuneBuilder
"target_cdrs": ["H3"], # Focus on the critical CDR-H3 loop
"num_designs": 10, # Generate 10 variants
"temperature": 0.2, # Conservative diversity
}).json()
print(f"Generated {len(design_resp['designs'])} CDR-H3 variants:\n")
for i, d in enumerate(design_resp["designs"]):
print(f" Variant {i+1}: {d['cdr_h3_sequence']}")
print(f" Score: {d['score']:.3f} Mutations: {d['num_mutations']}")Comparison Summary
Each tool occupies a distinct niche. IgFold wins on speed and is ideal for high-throughput screening. ABodyBuilder2 leverages the proven AlphaFold2 architecture for maximum accuracy on paired antibodies when runtime is not a constraint. ImmuneBuilder offers the best balance of speed, accuracy, and versatility — and is the only tool with dedicated support for nanobodies and TCRs.
For most antibody engineering workflows, running ImmuneBuilder through SciRouter gives you production-quality predictions in about 20 seconds, with the option to immediately chain into AntiFold for CDR design. That predict-then-design loop — from sequence to optimized variants — is the core of modern computational antibody engineering.
Next Steps
To learn more about the full antibody design pipeline, read our AI Antibody Design tutorial which walks through structure prediction, CDR design, and validation end to end.
For protein structure prediction beyond antibodies, see our ESMFold vs AlphaFold2 vs Boltz-2 comparison. Or sign up for a free SciRouter API key and predict your first antibody structure in under a minute.