AntiFold

AntiFold webserver overview

AntiFold 0.3.1 performs antibody-specific inverse folding on a fixed variable-domain structure. It supports paired heavy and light variable domains, nanobodies, optional antigen context, IMGT region masks, probability analysis, sequence sampling, and native per-residue embeddings.

Each run returns AntiFold's residue-level amino-acid distribution and run files. Sampling runs also return designed antibody sequences with source-native scores. The model is described by Høie et al. (2025). The sections below cover choosing a region mask, running the probability analysis before sampling, temperature and seed choices, common run problems, and validation of designed sequences.

Pricing

Job	Credits	Assumptions
One antibody or nanobody structure	30	One .pdb, .ent, or .cif input. The current quote is unchanged by sequence count, temperature, region selection, antigen context, or embedding extraction.

The exact credit quote is calculated before submission. A workflow that submits multiple structures creates one quoted AntiFold job per structure.

Inputs

Input	Accepted data	Limit	Notes
Antibody Structure	.pdb, .ent, .cif, or an RCSB PDB fetch	50 MiB	Paired mode requires at least two protein chains. Nanobody mode accepts a single VHH chain.

The structure should contain antibody variable domains. Region-specific sampling uses the structure's IMGT residue numbering. AntiFold reports a warning when the numbering does not appear compatible instead of silently renumbering the structure.

Chain IDs can use any valid labels. In paired mode, blank heavy and light chain fields select the first and second non-antigen chains in file order. In nanobody mode, a single available chain can be selected automatically. A declared antigen chain is used as structural context and cannot also be assigned an antibody role.

AntiFold 0.3.1 reports only the final character of each selected chain label in its native results. Multi-character CIF labels can be associated with the submitted chains when every selected antibody and antigen label has a different final character. Labels such as HA, LB, and AG meet this condition. Colliding labels can fail in AntiFold or produce ambiguous native results. Successful ambiguous runs retain their native files and show a warning without assigning sampled sequences or mutations to submitted chains.

Settings

Core settings

Parameter	Type	Default	Description
Input mode	enum	Paired antibody (VH/VL)	Selects paired heavy/light or Nanobody / VHH processing.
Heavy chain ID	string	automatic	Optional paired-mode heavy-chain label. Blank selects the first non-antigen chain.
Light chain ID	string	automatic	Optional paired-mode light-chain label. Blank selects the second non-antigen chain.
Nanobody chain ID	string	automatic	Optional nanobody-mode VHH chain label. A single available chain is selected automatically.
Antigen chain ID	string	optional	Includes one antigen chain as structural context. Sampling remains limited to the selected antibody chain or chains.
Number of sequences	integer	0	Generates 0 to 100 variants. 0 runs the source-default probability analysis without sequence sampling.
Sampling temperature	number	0.2	Accepts 0 to 1.5. Lower values concentrate sampling on high-probability amino acids; higher values increase diversity.

IMGT region selection

Parameter	Type	Default	Description
IMGT regions to design	enum	CDRs only (CDR1, CDR2, CDR3)	Selects the source-native region mask used for sequence sampling.

Available aggregate masks are:

• cdrs: all CDR1, CDR2, and CDR3 positions
• all: all antibody variable-domain regions
• heavy_cdrs and light_cdrs: all CDRs on one chain
• CDR1, CDR2, and CDR3: the named CDR on both applicable chains
• frameworks: all framework regions
• heavy_all and light_all: all regions on one chain

Chain-specific masks are CDRH1, CDRH2, CDRH3, CDRL1, CDRL2, CDRL3, FWH1, FWH2, FWH3, FWH4, FWL1, FWL2, FWL3, and FWL4. Light-chain masks are unavailable in nanobody mode.

Choose Custom region list to combine masks using single spaces, for example CDRH1 CDRL3 or CDRH3 FW1. Accepted tokens are CDR1, CDR2, CDR3, CDRH, CDRL, all, allH, allL, the chain-specific masks above, and FW1, FW2, FW3, FW4. Nanobody lists must include at least one mask that applies to the VHH chain. Standalone FWH and FWL and integer position lists are not supported.

Advanced options

Parameter	Type	Default	Description
Random seed	integer	42	Accepts 0 to 4,294,967,295 and controls sequence sampling. Seeded FASTA designs are stable on the reviewed runtime, while final GPU probability digits may vary.
Extract per-residue embeddings	boolean	false	Returns AntiFold's native per-residue .npy representation. AntiFold 0.3.1 requires Number of sequences to be 0 when this option is enabled.

Choosing a region mask

Sampling changes only the residues inside the selected mask, so the mask defines the experiment. A narrow mask gives a library that is easier to interpret. CDRH3 explores the heavy-chain third loop alone, and heavy_cdrs redesigns all three heavy-chain loops while the light chain and every framework stay fixed.

CDR1, CDR2, and CDR3 select the same loop on both chains, frameworks selects every framework region, and all leaves no antibody region fixed. Broad masks change many residues at once. Use them only when the validation plan can separate structural compatibility from unwanted effects on binding, humanness, or developability.

The mask depends on IMGT residue labels. If the structure is not IMGT numbered, AntiFold warns and a region such as CDRH3 may not cover the intended residues. Renumber the structure to IMGT before regional sampling, and read the run log warnings after every run.

Running probability analysis first

Keep Number of sequences at 0 for the first run. The job returns the residue probabilities CSV and perplexity for every position without sampling, which shows which positions have a sharply preferred residue and which admit several alternatives. It also shows whether the model's top residue agrees with the submitted sequence inside the intended design region.

This run is also the only one that can extract per-residue embeddings, because AntiFold 0.3.1 does not combine embeddings with sampling. Inspect the probabilities and the run log before choosing a mask and temperature for the sampling run.

Temperature, sample count, and seed

The default temperature of 0.2 is conservative. The slider tooltip describes 0.1 to 0.3 as conservative designs, 0.4 to 0.6 as a balance of diversity and confidence, and 0.7 to 1.5 as a broad search; 0 always picks the most probable amino acid. Change the temperature or the mask between runs, not both at once.

If designs are nearly identical, raise the temperature in small steps. If they mutate too many positions, lower it or narrow the mask. A larger Number of sequences shows more draws, not a more accurate draw.

The Random seed defaults to 42. Identical settings reproduce the sampled FASTA designs on the reviewed runtime, so record the seed with the run. Probability and embedding values can differ in their final digits between equivalent GPU runs.

Outputs

Output	Format	When returned	Contents
Residue probabilities	.csv	Every successful run	Structure chain and position, insertion code, original and top residue, IMGT region, perplexity, and log probabilities for all 20 standard amino acids.
Designed sequences	.fasta	Number of sequences is greater than 0	Original and sampled antibody sequences with temperature, sample number, score, global score, sequence recovery, and mutation count in the native headers.
Per-residue embeddings	.npy	Embedding extraction is enabled	Source-native float32 latent representation, preserved as binary NumPy data.
Structured results	.json	Every successful run	Parsed residue rows, sampled sequence metadata, mutations, summary values, and source provenance.
Run log	.txt	Every successful run	AntiFold's native execution log and scientific warnings.

The Results view compares sampled sequences with the submitted sequence. The Residue probabilities view presents the native CSV as a spreadsheet. Every native file remains downloadable from the Files view and can be routed through workflows where its format is supported.

Understanding results

• top_res is the highest-probability amino acid at a residue position.
• The amino-acid columns contain source-native log probabilities. A larger value, meaning closer to zero, indicates greater model support at that position.
• perplexity is the effective breadth of the residue distribution. Lower values indicate a concentrated distribution; higher values indicate more alternatives. Compare positions within the same run rather than applying a universal pass or fail threshold.
• score is the sampled sequence's mean negative log-likelihood over the selected design region. global_score uses all selected antibody residues. Lower values indicate stronger model support, but neither value is an experimental affinity or stability measurement.
• seq_recovery is the fraction of sampled antibody positions that match the submitted sequence. Mutation records retain the native CSV position and insertion-code text. For uniquely associated multi-character labels, structured results distinguish the submitted chain label from the native chain label. Native CSV and FASTA files remain unchanged.

AntiFold assumes a fixed input backbone. A sequence with a favorable score still requires independent structural, developability, binding, and experimental validation for its intended use.

Troubleshooting

Symptom	Likely cause	Fix
Paired mode rejects the file	The structure has fewer than two protein chains	Use Nanobody / VHH for a single VHH chain, or upload the paired variable domains.
Wrong chains are treated as heavy and light	Blank chain fields follow file order	Set Heavy chain ID, Light chain ID, and Antigen chain ID explicitly.
Numbering warning, or a mask covers unexpected residues	Residue labels are not IMGT numbered	Renumber the structure to IMGT outside ProteinIQ and resubmit.
Embedding extraction conflicts with the settings	Number of sequences is greater than 0	Set Number of sequences to 0 for the embedding run.
Light-chain mask is unavailable or rejected	Nanobody mode has only one chain	Use a heavy-chain mask, or a custom list with at least one mask that applies to the VHH.
Sampled sequences are not assigned to chains	Selected chain labels share the same final character	Rename chains so every selected antibody and antigen label ends in a different character.
Designs are too similar or mutate too many positions	Temperature or mask is too low or too broad for the goal	Change the temperature in small steps, or narrow the mask, one change at a time.

Validating designed sequences

A favorable score shows model support on a fixed backbone, not binding or stability. Refold shortlisted sequences with an antibody structure predictor such as ABodyBuilder3 and align each model to the design backbone. Inspect the variable-domain orientation and every mutated loop, not only the whole-Fv summary.

Then confirm that mutations sit at the intended IMGT positions and check for new cysteines, glycosylation motifs, and surface hydrophobic or charged patches. BioPhi scores humanness separately, and that result should stay separate from the AntiFold score. Test the parent sequence alongside the designs in expression, stability, affinity, and specificity assays.

Sources

• Høie MH, et al. "AntiFold: improved structure-based antibody design using inverse folding." Bioinformatics Advances 5(1), vbae202 (2025). doi:10.1093/bioadv/vbae202