AntiFold is an inverse folding model specialized for antibody variable domains. Given an antibody structure, it predicts amino acid sequences that would fold into that structure—the reverse of traditional structure prediction.
Developed by the Oxford Protein Informatics Group, AntiFold builds on Meta's ESM-IF1 model but is fine-tuned specifically on antibody structures from SAbDab (Structural Antibody Database) and predicted structures from OAS (Observed Antibody Space). This antibody-specific training dramatically improves sequence recovery for antibody CDR loops compared to general-purpose inverse folding tools.
The model uses IMGT numbering, the international standard for immunoglobulin sequences, enabling precise targeting of specific antibody regions during design.
AntiFold takes a 3D antibody structure and predicts the probability distribution of amino acids at each position. The model learns structural constraints—which residues are compatible with a given backbone geometry—from millions of antibody structures.
During fine-tuning from ESM-IF1, several strategies improved CDR3 recovery: span masking to learn regional context, weighted masking that emphasizes CDR residues (3:1 ratio over frameworks), layer-wise learning rate decay, and augmentation with OAS predicted structures. These modifications increased CDRH3 amino acid recovery from 43% (ESM-IF1 baseline) to 60%.
For each position, the model outputs:
From these probabilities, AntiFold samples sequences using temperature-controlled multinomial sampling. Lower temperatures produce conservative sequences closer to the probability maximum; higher temperatures explore more diverse sequence space.
ProteinIQ provides GPU-accelerated AntiFold without installation. The tool automatically handles IMGT renumbering via ANARCI.
| Input | Description |
|---|---|
Antibody Structure | PDB or mmCIF file containing the antibody variable domains. Structures can be uploaded directly or fetched by PDB ID. |
The structure should contain paired heavy and light chains (VH/VL) or a single-domain antibody (VHH/nanobody). An optional antigen chain may be included but should be kept small for optimal performance.
| Setting | Description |
|---|---|
Heavy chain ID | Chain identifier for the heavy chain (default: H). |
Light chain ID | Chain identifier for the light chain (default: L). |
Number of sequences | Sequence variants to generate per structure (1–100, default: 10). |
Sampling temperature | Controls sequence diversity (0.0–1.5, default: 0.2). Values 0.1–0.3 produce conservative designs; 0.7–1.5 explores diverse sequence space. |
| Setting | Description |
|---|---|
IMGT regions to design | Target specific regions: All regions, CDRs only, individual CDRs (CDR1, CDR2, CDR3), or Frameworks only. |
Complementarity-determining regions (CDRs) form the antigen binding site. CDR1, CDR2, and CDR3 loops on both heavy and light chains determine specificity. Framework regions provide structural scaffolding.
| Setting | Description |
|---|---|
Random seed | For reproducible results (default: 42). |
Include per-residue scores | Output log-likelihoods and perplexity at each position. |
Results include designed sequences in FASTA format with associated metrics:
| Column | Description |
|---|---|
score | Average log-likelihood over the designed region |
global_score | Average log-likelihood over all residues |
seq_recovery | Fraction of positions matching the original sequence |
When per-residue scoring is enabled, a CSV file provides position-by-position analysis including perplexity values and individual amino acid log-likelihoods.
Perplexity reflects how many amino acids are structurally compatible at a position:
| Perplexity | Interpretation |
|---|---|
| 1–3 | Highly constrained; few substitutions tolerated |
| 4–8 | Moderately flexible; several alternatives possible |
| 10–14 | Structurally tolerant; typical for exposed CDR positions |
| > 14 | Very permissive; likely surface-exposed or disordered |
Lower perplexity at a position suggests that the backbone geometry strongly constrains the amino acid identity—useful for identifying structurally critical residues.
ProteinMPNN is a general-purpose inverse folding model trained on diverse protein structures. While powerful, it occasionally produces artifacts problematic for antibodies: chain reordering, gaps in IMGT-numbered structures, and suboptimal CDR3 predictions.
AntiFold addresses these issues with antibody-specific training. On experimental structures, it achieves 60% CDRH3 amino acid recovery versus 43% for ESM-IF1 and 56% for the antibody-adapted AbMPNN. For the notoriously variable CDRH3 loop, AntiFold constrains perplexity to 4–8 amino acids (versus 6–10 for AbMPNN), producing more focused designs that better preserve backbone geometry.
