Language-model numbering for antibodies, T cell receptors, and VNAR/VHH domains Learn more
ANARCII webserver overview
ANARCII is a suite of sequence-to-sequence transformer language models for numbering antibody, T cell receptor (TCR), and shark VNAR or VHH domains.
It predicts IMGT residue labels directly from amino acid sequence without aligning the query to a germline reference. The result can then be converted to Kabat, Chothia, Martin, or AHo numbering.
The ProteinIQ webserver runs ANARCII 2.0.8 on pasted sequences, FASTA or PIR datasets, and PDB, mmCIF, or mmJSON structures. Results include chain classification, model scores, domain boundaries, residue-level numbering, structured exports, and a downloadable structure artifact for structure inputs.
The Oxford Protein Informatics Group developed ANARCII as a language-model alternative to alignment-based antigen receptor numbering.
In the published challenging test sets, it matched existing methods at more than 99.99% of conserved residues and more than 99.94% of complete CDR regions.
The model is designed to generalize to rare sequence types and species that may not be represented well by a fixed germline reference set.
How to use ANARCII online
Paste an antibody, TCR, VNAR, or VHH sequence, upload FASTA, PIR, PDB, mmCIF, or mmJSON data, or fetch a structure from RCSB. Choose the model family and numbering scheme. ANARCII detects the input format and returns residue assignments, chain calls, model scores, downloadable data, and an IMGT-renumbered structure when applicable.
Inputs
A run accepts either sequence data or one structure through the same input. Paste sequence text, choose any supported sequence or structure file, or fetch a structure from RCSB.
| Input | Accepted formats | Limits and behavior |
|---|---|---|
Antibody or TCR input | Raw sequence or FASTA text; .fasta, .fas, .fsa, .fa, .faa, .mpfa, .pir, .nbrf, .ali, .txt, .pdb, .ent, .cif, .mmcif, .json, or .mmjson; RCSB PDB ID | One sequence dataset or one structure per run. FASTA and PIR files can contain multiple named records. Sequence files support gzip compression, and FASTA or PIR formats also support .z compression. Structure files support gzip compression. ANARCII detects the input type from the format and writes a PDB, mmCIF, or mmJSON output. Residue identifiers change only on chains that pass ANARCII's numbering checks. Files are limited to 50 MB. |
Job name | Text | Optional label used to identify the run. |
Sequence text must use uppercase one-letter amino acid codes. X, B, O, J, U, and Z are accepted; ANARCII tokenizes B, O, J, U, and Z as X.
A forward slash, backslash, or hyphen can separate paired sequence parts. Lowercase residues and unsupported punctuation are rejected before the run.
Per-job sequence limits are 50 records for guests, 500 for Free, 2,000 for Starter, and unrestricted by record count for Pro. The upload-size limit still applies.
Settings
The standard run uses the antibody accuracy model, IMGT numbering, and no advanced outputs.
| Setting | Default | Description |
|---|---|---|
Sequence family | Antibody | Chooses Antibody, T cell receptor, Detect antibody or TCR, or Shark VNAR / VHH. Automatic detection classifies each sequence before routing it to the antibody or TCR model. |
Model mode | Accuracy | Chooses the accuracy or speed weights. Accuracy is recommended for rare, divergent, or unusual receptor sequences. Speed improves throughput with a small reduction in agreement on difficult cases. Shark VNAR or VHH numbering uses its single available model for either selection. |
Numbering scheme | IMGT | Returns IMGT, Kabat, Chothia, Martin, or AHo numbering. ANARCII first predicts IMGT labels and then converts the tabular result to the selected scheme. |
Discover scFv domains | Off | Searches long sequence inputs for multiple receptor domains, such as the linked heavy and light variable domains in an scFv. This option is available only for sequence input. |
Report low-score warnings | Off | Adds warnings for successfully numbered sequences whose scores fall below the expected range for the selected model. It cannot be combined with automatic antibody or TCR detection in ANARCII 2.0.8. |
Return model logits | Off | Writes per-token model logits to anarcii_logits.json. These values are intended for diagnostics and can make downloads substantially larger. |
Outputs
Each detected domain or failed record receives a result row. With scFv discovery, one input can produce multiple domain rows.
A sequence that cannot be numbered remains in the table with chain type F and an explanation in Error. Other records in the same run can still succeed.
Result columns
| Column | Description |
|---|---|
Query | FASTA or PIR record name, generated sequence name, or structure model and chain identifier. |
Chain | Predicted receptor chain: H, K, or L for antibodies; A, B, G, or D for TCRs; F when numbering failed. |
Score | Mean model logit across valid non-insertion numbering labels. This is a model confidence signal, not a probability or experimental measurement. |
Query Start | Zero-based, inclusive index of the first numbered residue in the submitted sequence. |
Query End | Zero-based, inclusive index of the last numbered residue in the submitted sequence. |
Scheme | Numbering scheme used in the table and structured numbering outputs. |
Error | ANARCII failure detail for records that could not be numbered. Successful records leave this field blank. |
Numbered Sequence | Residue assignments in position:residue form, including insertion letters such as 111A:P. |
Result files
| File | When returned | What it contains |
|---|---|---|
anarcii_results.csv | Every run | ANARCII's horizontal numbering table in the selected scheme, including the complete ordered residue columns. |
anarcii_results.msgpack | Every run unless logits prevent serialization | Binary ANARCII results in the selected scheme for efficient programmatic use. |
anarcii_results.json | Every run | Structured result records with chain type, score, query boundaries, errors, and residue numbering. |
anarcii_legacy.json | Sequence input | ANARCII's compatibility representation for workflows that consume the earlier ANARCI-style result structure. |
anarcii_logits.json | When Return model logits is enabled | Per-token IMGT model logits captured before any alternate-scheme conversion. |
anarcii_renumbered.pdb, .cif, or .json | Structure input | Downloadable structure output. Qualifying receptor chains use IMGT residue identifiers; chains that fail numbering retain their existing identifiers. |
run.log | Every run | Selected settings, record counts, warnings, artifact decisions, and captured ANARCII messages. |
ANARCII 2.0.8 cannot write its MessagePack artifact when tensor logits are retained. With Return model logits enabled, numbering still completes and the CSV, JSON, logits, and log files are returned, but anarcii_results.msgpack may be omitted with an explicit warning.
Understanding ANARCII results
The Score is the average of the strongest predicted numbering-token logits over non-insertion positions. Higher values mean the selected model assigned its numbering labels more decisively.
Scores from different sequence families or model modes do not share one calibrated scale and should not be compared directly.
ANARCII 2.0.8 uses family-specific failure thresholds:
| Sequence family | Failure threshold | Result below threshold |
|---|---|---|
| Antibody | 15 | Chain F with Score less than cut off. |
| T cell receptor | 25 | Chain F with Score less than cut off. |
| Shark VNAR / VHH | 24 | Chain F with Score less than cut off. |
A sequence with fewer than 50 predicted non-insertion residues receives a score of 0 and the error Less than 50 non insertion residues numbered. This is the expected result for short random sequences, fragments without a complete variable domain, and many non-receptor proteins.
When low-score warnings are enabled, ANARCII flags successful results that sit above the failure threshold but at or below the typical range for that model:
| Model | Warning range |
|---|---|
| Antibody accuracy | Greater than 15 and at most 25 |
| Antibody speed | Greater than 15 and at most 22.5 |
| TCR accuracy or speed | Greater than 25 and at most 32.5 |
| Shark VNAR / VHH | Greater than 24 and at most 24.5 |
A low score can reflect a divergent receptor, an incomplete domain, sequencing errors, the wrong sequence-family model, or a non-receptor sequence.
The numbered residues, conserved framework positions, domain boundaries, and input quality provide the context needed to interpret that score.
How ANARCII works
ANARCII treats numbering as a sequence-to-sequence translation problem. A transformer encoder reads the amino acid sequence. An autoregressive decoder then predicts a chain token followed by IMGT position tokens.
Insertions are emitted as special tokens and converted into labels such as 111A or 112B. Missing integer positions become gaps, keeping receptor domains in a consistent coordinate system.
Five trained model variants cover the supported receptor families:
- Antibody accuracy: Predicts heavy, kappa, or lambda chain type and IMGT numbering.
- Antibody speed: Uses a lighter model for higher-throughput antibody numbering.
- TCR accuracy: Predicts alpha, beta, gamma, or delta chain type and numbering.
- TCR speed: Provides higher-throughput TCR numbering.
- Shark VNAR / VHH: Uses one accuracy model tuned for shark VNAR-like domains, including their characteristic CDR2 gap pattern.
Long sequences pass through a window-selection stage that identifies likely receptor domains before numbering. Automatic family detection classifies records as antibody or TCR.
scFv discovery can return multiple domains from one linked sequence. Alternate numbering schemes are applied after the model has produced IMGT numbering.
For detailed V, D, and J annotation, IgBLAST provides germline-alignment results rather than language-model numbering scores.
ANARCII vs ANARCI
ANARCII and ANARCI both assign standardized residue numbers to antigen receptor variable domains. Their methods and supporting outputs differ.
ANARCII predicts numbering tokens with a sequence-to-sequence transformer. ANARCI aligns each query against germline-derived profile HMMs, then uses the best match to classify and number the domain.
| Comparison | ANARCII | ANARCI |
|---|---|---|
| Core method | Alignment-free transformer language model | HMMER alignment against species and chain-type profile HMMs |
| Best fit | High-throughput numbering, rare or divergent receptor sequences, shark VNAR or VHH domains, and structure renumbering | Reference-aligned receptor classification, HMM statistics, species matching, and optional closest V and J germline assignment |
| ProteinIQ input | Pasted sequences, FASTA or PIR datasets, PDB, mmCIF, and mmJSON structures | Pasted sequences and FASTA datasets |
| Receptor models | Antibody heavy, kappa, and lambda chains; TCR alpha, beta, gamma, and delta chains; shark VNAR or VHH | Antibody heavy, kappa, and lambda chains; TCR alpha, beta, gamma, and delta chains |
| Numbering schemes | IMGT, Kabat, Chothia, Martin, and AHo | IMGT, Kabat, Chothia, Martin, AHo, and Wolfguy |
| Confidence evidence | Model score, failure reason, and optional per-token logits | HMM bit score, E-value, best species and chain hit, and complete HMM hit table |
| Species and germline output | Not reported | Best HMM species match and optional closest V and J genes with sequence identity |
| Structure output | Downloads a PDB, mmCIF, or mmJSON artifact with qualifying chains renumbered to IMGT | No structure input or renumbered structure output in the ProteinIQ tool |
Choosing between ANARCII and ANARCI
ANARCII is the more direct choice for large numbering jobs, structure files, VNAR or VHH sequences, and receptors that may be poorly represented by known germlines.
ANARCI is preferable when the analysis needs alignment statistics, Wolfguy numbering, a species-similarity result, or a closest germline V and J assignment. Its species field reflects the best HMM match, not definitive biological origin.
When the methods disagree under the same numbering scheme, compare the detected domain boundaries and insertion placement. A strong ANARCI hit supplies inspectable alignment evidence, while ANARCII is designed to generalize beyond a fixed reference set.
Important limitations
- ANARCII assigns numbering and receptor chain type. It does not infer antibody species, germline genes, clonotypes, somatic mutations, or antigen specificity.
- Model scores are useful within the selected family and mode, but they are not probabilities and do not validate biological function.
- Short fragments, non-receptor proteins, heavily corrupted reads, and domains with fewer than 50 numbered non-insertion residues can return chain
F. - Automatic antibody or TCR detection cannot be combined with low-score warnings in ANARCII 2.0.8.
- scFv domain discovery is available only for sequence input.
- Renumbered PDB, mmCIF, and mmJSON structures always use IMGT residue identifiers. When another scheme is selected, the table, CSV, MessagePack, and structured JSON use that scheme while the structure remains IMGT-numbered.
- Structure renumbering changes residue identifiers. It does not repair missing atoms, rebuild loops, change coordinates, or confirm that every chain is an antigen receptor.
- Numbering-scheme conversions change residue labels and CDR boundaries, not the underlying sequence. Results should be compared only when the same scheme is used.
