How IgBLAST results compare to the bench

You have an IgBLAST result. Here is what it supports.

• A V-gene call is the part you can lean on. Against simulated repertoires with known ground truth, IgBLAST's V-gene accuracy is 98.6% and its gene-level mishit frequency is the lowest of the three commonly used annotators. That is the strongest published result on this page.
• Do not trust a D-gene call. On the same ground-truth benchmark IgBLAST called the correct D gene in 52.1% of sequences. This is a property of the problem, not only of the tool — the best method tested reached 82.8% — but it means a D call is a coin flip dressed as an annotation. On this platform there is no D call at all.
• Every figure below is a gene-level figure. No published study benchmarks IgBLAST at allele resolution, and the one benchmark that compared it head-to-head says allele-level work "would be necessary." If your decision turns on *01 versus *02, nothing here supports it.
• Confirm a call with Sanger sequencing of the rearrangement interpreted under the ERIC guidelines, which require V, D and J genes and alleles, productivity, and germline identity to two decimal places. A clinical-grade answer needs nucleotide input; this platform takes amino acids.
• Somatic hypermutation here is a framework mismatch count, not a germline identity percentage. v_identity_pct is read off the alignment-summary Total row. It is not the IMGT-reported figure that the 98% U-CLL/M-CLL cut-off is defined against, and it must not be substituted for it.

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench.

How in silico is used currently

IgBLAST is not used to decide a germline assignment so much as to produce one at scale, and then — in any setting where the answer matters — to have it confirmed by a second method. Three patterns dominate.

• Repertoire-scale annotation. You have 10⁴–10⁶ sequences and you need every one assigned to a V, D and J gene before you can cluster clonotypes, count CDR3s, or measure hypermutation. This is the use the tool was built for and the use it is strongest at. Dedicated wrappers exist purely to run it over billions of sequences.
• As the arbiter for someone else's measurement. A large share of the papers that mention IgBLAST are not grading it — they are holding it fixed and grading something else. A 2025 comparison of nanopore and Illumina clonotyping in 13 CLL patients ran both platforms, then "confirmed" the clonotypes from each with IgBLAST and compared the two sequencing technologies to each other. IgBLAST was the ruler, not the thing being measured.
• Lineage and humanness triage. A V-gene call plus a germline-identity figure tells you how far a therapeutic candidate has drifted from germline, which is the input to humanness scoring. This is the use the platform's chain is built around, and it is the one that survives having only a V call.

What IgBLAST is not used for, by anyone reporting their validation carefully, is a clinically reportable germline assignment on its own. The ERIC recommendations for IGHV analysis in CLL — the one place in this literature with a guideline, a defined report, and a threshold with treatment consequences — name IMGT/V-QUEST and ARResT/AssignSubsets. They do not name IgBLAST.

Hansen MH, Gulmir R, Cédile O, Dahlmann SK, Veyhe SR, Stougaard M, et al. Performance of rapid clonotyping of chronic lymphocytic leukemia using Flongle flow-cell nanopore sequencing of IGH rearrangements. European Journal of Haematology. 2025;115(6):584-590. doi: 10.1111/ejh.70030

Soto C, Finn JA, Willis JR, Day SB, Sinkovits RS, Jones T, et al. PyIR: a scalable wrapper for processing billions of immunoglobulin and T cell receptor sequences using IgBLAST. BMC Bioinformatics. 2020;21(1):314. doi: 10.1186/s12859-020-03649-5

Against other in silico methods

Head-to-head, IgBLAST's standing depends entirely on which segment you ask about.

On V genes it is at the top of the general-purpose field. 98.6% gene-level accuracy on simulated ground truth against IMGT's 98.8%; the lowest average mishit frequency of the three most-used annotators (0.004 against MiXCR's 0.020); correct 100% of the time on the IGHV3-66/IGHV3-53 pair that MiXCR miscalled at up to 0.76. The million-sequence benchmark's own scorecard rates IgBLAST highest of the three on accuracy, on reference-germline flexibility, and on reproducibility with other tools, and concludes that "IgBLAST would be best if accuracy is of utmost importance."

On D genes it came last of the six methods tested. 52.1% against partis at 82.8%, ighutil at 71.4%, IMGT at 57.4% and even iHMMune-align at 53.0%. The modelling-based methods win here because D assignment is an inference problem under junctional trimming and insertion, not a similarity-search problem, and a local-alignment tool has little to work with on a segment roughly 10–20 bases long.

The D-gene weakness is partly a default, and the authors documented it. In IgBLAST's own evaluation, on a set of 101 clonally related sequences whose dominant rearrangement was known to use IGHD6-6*01, IgBLAST identified that D gene in only 48 of 101 sequences with its default D-gene mismatch penalty of −4. Lowering the penalty to −1 raised it to 87 of 101. The authors state directly that the default "is not optimal for identifying D genes with more mismatches." ⊘ This platform exposes no D-gene mismatch penalty parameter, and no num_alignments_d either — see inputs. On the nucleotide path the default would stand; on the protein path the question is moot, because there is no D database.

On J genes it sits mid-field: 87.2% against partis at 99.8%, IMGT at 93.5% and ighutil at 96.1%.

On throughput MiXCR wins and IgBLAST is second. MiXCR annotated every dataset fastest; IgBLAST and IMGT/HighV-QUEST traded places depending on data type.

One comparative property is worth more than any accuracy figure. IgBLAST and IMGT/HighV-QUEST, unlike MiXCR, do not report a statistical summary of how many reads were annotated — a point the benchmark raises explicitly as a hazard for low-quality data. A tool that does not tell you how much of your input it handled will let a bad run look like a good one. That is the same structural problem as this platform's SUCCESS-with-null-fields behaviour, one layer down.

In silico vs. wet lab

The wet-lab axis here is different in kind from a docking tool's, and the difference is the most important framing on this page. IgBLAST does not predict a physical quantity. It performs a classification: given a sequence, which germline segments recombined to make it? There is no co-crystal to compare against and no RMSD threshold. The comparator is repertoire sequencing plus expert-curated V/D/J assignment, and agreement is measured as gene-call concordance, CDR3 junction accuracy, and productive/unproductive call agreement against either a curated clinical reference or a simulated repertoire with known ground truth.

The standard approach for the question IgBLAST answers is Sanger sequencing of the clonotypic IGHV-IGHD-IGHJ rearrangement, interpreted under the ERIC guidelines. That protocol specifies leader primers, requires the report to carry V, D and J genes and alleles, requires somatic-hypermutation status to be determined only for productive rearrangements with the reason stated when a rearrangement is unproductive, and requires germline nucleotide identity to two decimal points as reported by IMGT — classified U-CLL ≥ 98%, M-CLL < 98%, borderline 97–97.99%.

Note what that list is: it is almost exactly the set of fields this platform returns as null.

Read the accuracy column by cell, never down the column. Sanger's "accuracy" is definitional — it is what the others are graded against. A mishit frequency is an error rate over gene calls. A per-segment percentage is the share of sequences assigned the correct gene regardless of allele. The 3–4% failure rate in the first row is a different quantity again: a rate of no answer, not of wrong answer.

What the table does support is time and cost at a stated accuracy, and on that reading the honest position for this chain is narrow: seconds instead of weeks, for a V-gene call at roughly 98.6% gene-level accuracy and nothing else. Everything the clinical standard requires beyond the V gene — the D and J calls, the alleles, the productivity verdict, the two-decimal germline identity — this platform does not produce.

Two asymmetries matter in the other direction. Sanger fails on 3–4% of cases and no amount of effort recovers them, which is part of why computational annotation exists. And the wet-lab comparator is itself only as good as the germline database it is read against, which is the subject of the whole next section.

Agreement with wet lab

There is no bench ground truth in this literature, and that is the central fact on this page. No study has taken a set of antibody sequences whose true germline origins were established experimentally and measured IgBLAST against them. There is no experiment that recovers which germline segments recombined in a mature, hypermutated B cell — the rearrangement deleted the evidence. So the field's ground truth comes from two substitutes, and both are one step removed:

1. Simulated repertoires, where the generating V, D and J genes are known by construction. This is the only place with true per-sequence ground truth, and every quantitative figure below comes from it.
2. A curated clinical reference, where an expert assignment under a guideline stands in for truth.

⊘ Simulation ground truth is not bench ground truth. A simulator samples germline segments from a reference database, trims them, inserts junctional nucleotides and applies a mutation model. Every one of those steps is a modelling choice, and a tool is rewarded for sharing the simulator's assumptions. The sharpest illustration is in the benchmark itself: the simulator's germline set held 55 V genes, IgBLAST's held 104, MiXCR's held 97, and all three shared only 32% (44 of 137) of them. Part of any measured mishit rate is therefore database mismatch rather than algorithm error. Read these numbers as the best available, not as the truth.

Gene-level accuracy against simulated ground truth

On 30,000 simulated sequences (independent; mutation frequency deliberately doubled to make inference harder), the fraction of sequences for which each method called the correct gene regardless of allele:

Three things to take from this table. IgBLAST's V-gene accuracy lands within 0.2 percentage points of IMGT's and about one point below the best HMM method — on V genes, every method here is good. Its D-gene accuracy is 52.1%, the lowest figure in that column, against 82.8% for the best method — and the authors note that the failures, which hit 10% of IgBLAST's inputs, are "usually due to poor D matches." And IgBLAST failed outright on 10% of sequences where partis and ighutil failed on essentially none: declining to answer is better behaviour than guessing, but 10% is a large share of a repertoire to lose silently.

On two 1-million-sequence simulated human IGH repertoires (independent, IgBLAST v1.12, one diverse and one polarized composition), the same question was asked as a mishit frequency — the share of first-reported gene assignments that were wrong. IgBLAST had the lowest average mishit frequency of the three tools tested, 0.004, against 0.020 for MiXCR. The errors are not uniformly spread: medians sat at or near zero and the averages were pulled up by outlier genes.

The outliers are the usable part. On IGHD7-27, the shortest D gene, IgBLAST aligned correctly at a frequency of only 0.89 (diverse) and 0.76 (polarized) — still far better than IMGT/HighV-QUEST at 0.25/0.40 and MiXCR at 0.46/0.63, but a 1-in-4 error rate on a single gene. On the near-identical pair IGHV3-66 / IGHV3-53, IgBLAST and IMGT/HighV-QUEST were correct 100% of the time in both repertoires where MiXCR miscalled at 0.49 and 0.76. For J genes, no tool exceeded a mishit frequency of 0.01.

CDR3 junction accuracy

There is no published figure for IgBLAST's CDR3 accuracy against ground truth. What exists is reproducibility between tools, which is a lower bound on how settled the answer is — and it is startlingly low. Across the simulated and experimental datasets, the share of unique CDR3 amino-acid sequences shared between the three tools ran from 10.4% to 31.9% on raw annotated output, 4.3% to 77.6% after preprocessing, and 31.9% to 77.5% among the 100 most frequent CDR3s. IgBLAST and IMGT/HighV-QUEST, which were run against the same germline set, overlapped on only 30.7% of annotated CDR3s, 12% after preprocessing, and 33% of the top 100.

The authors decomposed that disagreement, and the decomposition matters more than the headline. Of 759,505 input sequences, 6.82% (51,791) produced an IgBLAST CDR3 with no counterpart in IMGT/HighV-QUEST's output, against 1.95% (14,844) the other way. Of those, 1% (7,735) had the same V/D/J gene assignment but a different CDR3 annotation, and only 0.16% (1,248) had a different gene assignment. In other words: most CDR3 disagreement is not a germline-calling disagreement at all. It is a disagreement about where the junction boundaries are, between tools that agreed on the genes.

Productive / unproductive call agreement

No published figure. The question is not benchmarked. What is published is the downstream effect: filtering to productive, functional, unique CDR3s discards wildly different fractions of each tool's output. On one Illumina MiSeq dataset, preprocessing removed 52.8% of IgBLAST's V-gene alignments, 86.9% of IMGT/HighV-QUEST's and 16.3% of MiXCR's; on a second MiSeq dataset built with modern protocols, IgBLAST lost 37%. On an Ion Torrent dataset IgBLAST was the only tool able to align a meaningful number of sequences at all, and then 99.3% of them were filtered out, leaving 0.7%. A productivity call is doing a great deal of work, and nothing grades it.

⊘ On this platform, productive and stop_codon are always null. Both are nucleotide-level properties — an out-of-frame junction, a premature stop, a frameshifting indel — and they are undefined for amino-acid input.

Allele-level accuracy

No published figure, and the gap is explicit. Every number above is gene-level. The million-sequence benchmark states plainly that "since MiXCR does not annotate at the allele level, the analysis was done at the gene level," credits IgBLAST and IMGT/HighV-QUEST with allele-level granularity, and then concludes that "an allele-level benchmarking of these two tools would be necessary." It has not been done.

The one quantitative result that bears on allele resolution does not flatter the approach IgBLAST takes. Aligning every sequence to its closest match in the full IMGT allele set — which is what IgBLAST does by default, and what this platform does — was measured against simulated repertoires with known allele sets, 10 samples of 50,000 sequences each:

Against roughly 50 true alleles per sample, the closest-match approach inferred 53 spurious alleles at low hypermutation and 86 at high — alleles that are not in the sample's true germline set. The reason is structural: the IMGT set catalogues alleles across the whole species, while any one individual carries roughly 50 of 250 V genes, 25 of 35 D and 6 of 12 J. Aligning to all of them guarantees spurious assignments. As the authors put it, the approach "rarely gets the correct gene" but "almost never has a gene that is very far from the correct one" — which is exactly the right thing to know. The call will be approximately right and specifically wrong.

Ye J, Ma N, Madden TL, Ostell JM. IgBLAST: an immunoglobulin variable domain sequence analysis tool. Nucleic Acids Research. 2013;41(W1):W34-W40. doi: 10.1093/nar/gkt382

Smakaj E, Babrak L, Ohlin M, Shugay M, Briney B, Tosoni D, et al. Benchmarking immunoinformatic tools for the analysis of antibody repertoire sequences. Bioinformatics. 2020;36(6):1731-1739. doi: 10.1093/bioinformatics/btz845

Ralph DK, Matsen FA 4th. Consistency of VDJ rearrangement and substitution parameters enables accurate B cell receptor sequence annotation. PLoS Computational Biology. 2016;12(1):e1004409. doi: 10.1371/journal.pcbi.1004409

Ralph DK, Matsen FA 4th. Per-sample immunoglobulin germline inference from B cell receptor deep sequencing data. PLoS Computational Biology. 2019;15(7):e1007133. doi: 10.1371/journal.pcbi.1007133

Agathangelidis A, Chatzidimitriou A, Chatzikonstantinou T, Tresoldi C, Davis Z, Giudicelli V, et al. Immunoglobulin gene sequence analysis in chronic lymphocytic leukemia: the 2022 update of the recommendations by ERIC, the European Research Initiative on CLL. Leukemia. 2022;36(8):1961-1968. doi: 10.1038/s41375-022-01604-2

Davi F, Langerak AW, de Septenville AL, Kolijn PM, Hengeveld PJ, Chatzidimitriou A, et al. Immunoglobulin gene analysis in chronic lymphocytic leukemia in the era of next generation sequencing. Leukemia. 2020;34(10):2545-2551. doi: 10.1038/s41375-020-0923-9

What the figure doesn't cover

• ⊘ This platform computes a V-gene call and nothing else, and says SUCCESS regardless. igblast_annotate accepts amino-acid variable-region sequences and runs igblastp against the distribution's protein V-gene database only — the container ships no D or J germline database, and nucleotide mode is gated behind an operator-supplied database directory that is not configured. The job's terminal status is set to SUCCESS whenever at least one V call is parsed; per-sequence failures are logged but do not change it, and operation_successful is always true. So top_d_gene, top_j_gene, j_gene_hits, cdr3_sequence, cdr3_length, productive and stop_codon arrive as null with nothing marking them as not-computable. The fields that are real: sequence_name, chain_type (derived from the V gene name), v_gene_hits, top_v_gene, v_identity_pct and framework_mutations. Treat the documented output table in outputs as the schema, not as a promise about any particular run.
• ⊘ The historical wrong-answer defect, and why old results need re-checking. A parsing bug routed IgBLAST's format-7 section data rows through the wrong handler, so alignment-summary region coordinates were recorded as gene names and percent identities — production observed top_v_gene as the string "1", with top_d_gene, top_j_gene, chain_type, productive, v_identity_pct and framework_mutations affected alongside it. Because the hit array was non-empty (full of coordinate garbage), the parser's only sanity check never fired and the job reported SUCCESS. The fix is merged and recorded as deployed, and a plausibility check now stamps a recurrence of the coordinate-as-gene-name shape. What was not done is retracting the roughly 255 historical SUCCESS jobs that carry wrong germline calls. If you are holding an IgBLAST result from before the fix, re-run it.
• ⊘ The automated validation gate for this tool asserts nothing. Its reference-data manifest contains zero examples, so the gate skips. A green CI run is not evidence that the germline calls are right. Separately, the validator that would run expects CDR region fields and a J-gene family that this tool never emits.
• No bench ground truth exists for anything on this page. Every quantitative figure is against a simulated repertoire. There is no experiment that recovers the true germline origin of a hypermutated sequence, so there is no concordance study to want. Treat the simulation figures as the ceiling of what is knowable, and treat the gap between simulator and biology as unquantified.
• Every figure is gene-level; allele-level is unbenchmarked. The benchmark that measured gene-level accuracy says allele-level benchmarking "would be necessary" and did not do it. Meanwhile the closest-match-to-full-IMGT approach was measured to infer 53–86 spurious alleles per sample against ~50 true ones. An allele suffix on an IgBLAST call carries no validated accuracy.
• The germline database is a moving part, and it is not versioned in the result. Over ten years the human IMGT reference accumulated 221 changes — 9 gene removals, 37 gene additions, 98 allele additions, 54 sequence changes and 23 metadata changes, updating on average every 13 weeks. Two runs of the same sequence against different database vintages can return different gene names. The result object records the organism and domain system but not the germline database version.
• Version matters and the published figures are old. The independent benchmarks tested IgBLAST v1.12 (million-sequence benchmark) and the 2016-era release (30,000-sequence benchmark). This platform runs 1.22.0. No independent evaluation of a 1.2x release has been published, so the figures on this page describe software several minor versions behind what runs here.
• The organism parameter is the single highest-leverage input and it is silently permissive. Aligning a mouse antibody against the human database mislabels every gene, and nothing in the result flags a species mismatch. The container ships a protein V database for five organisms, but the parameter schema accepts only four — so one shipped database is unreachable. See inputs.
• v_identity_pct is not the ERIC germline-identity figure. It is read from the alignment-summary Total row, which includes the partial CDR3 region. The clinical 98% U-CLL/M-CLL threshold is defined against the IMGT-reported nucleotide identity to two decimal places, derived from Sanger data, and the ERIC update notes that the cut-off's value "has not been evaluated to the same degree in the NGS context." Do not carry this platform's percentage across to that threshold.
• D-gene assignment is intrinsically hard, independently of any tool. The ERIC recommendations state that IGHD identification "may be difficult" because of excessive exonuclease trimming of the D gene and hypermutation within the VH CDR3, both of which erase the evidence. A 52% D-call accuracy is partly the problem's difficulty, not only the tool's.
• CDR3 boundary definitions are not standardised between tools. The decomposition above found that most CDR3 disagreement came from sequences where the tools agreed on all three genes. Comparing CDR3s across tools without normalising the boundary convention produces false novelty. This platform reports no CDR1/CDR2 or framework-region boundaries at all — framework_mutations is a mismatch count summed over framework rows, not a region annotation.
• Nothing here reports how much of your input was handled. IgBLAST publishes no annotation-rate summary, and on this platform success_count / fail_count count sequences that produced a V call, not sequences that produced a complete annotation.

Go / no-go

Go — act on the result when:

• You want a V-gene germline call and a germline-identity figure to feed humanness scoring or lineage triage, and the V gene is genuinely all you need. This is the one claim the evidence supports well.
• You are annotating at repertoire scale and will cluster or count downstream, where a 0.004 gene-level mishit frequency is acceptable noise and individual calls are not load-bearing.
• You are using the call as one input among several — the pattern the published work actually uses, with a second annotator or a nucleotide-level method confirming anything consequential.
• You have checked that the organism you passed matches your sequence's species, and you are reading the ranked v_gene_hits rather than only top_v_gene, so you can see how close the runner-up was.

No-go — do not act on the result when:

• ⊘ You need a D-gene or J-gene call, a CDR3, or a productivity verdict. On this platform these are null. Even on a correctly configured nucleotide run, the published D-gene accuracy is 52.1%.
• ⊘ You are relying on the allele suffix. *01 versus *02 is unbenchmarked for this tool, and the closest-match-to-full-IMGT approach it uses was measured to produce more spurious alleles than true ones.
• ⊘ The decision is clinical or regulatory. The ERIC-guideline report requires V, D and J genes and alleles, a productivity determination, and IMGT-reported germline identity to two decimal places. This platform supplies one of those four. Use Sanger sequencing with IMGT/V-QUEST interpretation.
• ⊘ You are holding a result from before the parsing fix, or any result you have not sanity-checked. A gene name that is a bare number, a coordinate, or implausibly short is the signature of the historical defect. Re-verify any germline call you act on against IMGT/V-QUEST or IMGT/HighV-QUEST — and treat SUCCESS as meaning "at least one V call was parsed," not "the annotation is complete or correct."
• You are comparing CDR3s or germline calls across tools or across time without pinning the germline database version and normalising the CDR3 boundary convention. Between-tool CDR3 overlap as low as 4.3% is the documented consequence.
• Your sequence is from a species whose germline set is poorly catalogued. The published caution is that non-human IMGT sets are incomplete enough that novel alleles get misread as hypermutation indels.

Caveats on reading this page: every figure above is quoted from the cited source with its original metric and sample, read from the body text rather than the abstract. Figures from different studies are not comparable to each other — the simulators, germline sets, tool versions and filtering rules differ between them. All accuracy figures are gene-level, not allele-level, and all come from simulated data.