Scoring Antibody Humanness: Gene-Specific CNNs vs Germline Identity

Every antibody humanness score answers the same question about a different object. Percent identity to the closest human germline V-gene compares your sequence to germline theory. A gene-specific classifier such as Humatch compares it to what a neural network learned about individual human V-genes, and tells you which gene it looks like. A repertoire peptide search such as

OASis, the metric published with BioPhi, breaks the variable domain into 9-mers and asks how often each one appears in sequenced human repertoires. The three numbers are not interchangeable, they disagree in informative ways, and none of them is a prediction of immunogenicity. On .bloom* you can run the gene-specific score as a job (humatch_score) and the Sapiens rewrite that acts on it (biophi_humanize); OASis percentiles are not computed on .bloom*, so if that metric is what you need, run the open-source BioPhi package yourself.

This article is about choosing deliberately between the three, and about the claim none of them supports.

At a Glance#

Germline Identity: The Baseline You Already Have#

Percent identity to the closest human germline V-gene is the oldest humanness measure and still the most widely quoted. Its virtues are real: it needs no model, it is reproducible to the character, every reviewer understands it, and the comparison is auditable because you can print the alignment.

Its weaknesses are structural rather than fixable.

• It is position-blind. A single number cannot distinguish ten conservative framework substitutions from three substitutions buried in CDR-H3. Those two candidates have very different humanization paths and may have the same identity score.
• It penalises being a real antibody. Human antibodies are somatically mutated away from germline as a matter of course. A score anchored on germline treats ordinary affinity maturation as a deficit, which is exactly the objection that motivated repertoire-based scoring.
• It names a neighbour, not a target. The nearest germline gene is not necessarily the gene you should humanize toward; that choice also depends on what pairs well, what expresses well, and what you have manufacturing experience with.
• It says nothing about the pair. Identity is computed per chain. Two chains can each be close to a human gene and still form a combination that does not occur in human repertoires.

Keep germline identity as the control that catches gross errors — a sequence that is 62% identical to its nearest human gene needs no sophisticated scoring to tell you it is murine. Then ask a method that answers one of the four questions above.

Gene-Specific Classifiers: What Humatch Measures#

Humatch, published by the Oxford Protein Informatics Group in 2024, is built from three lightweight convolutional neural networks: one trained to identify human heavy V-genes, one for light V-genes, and one for well-paired antibody sequences. The authors report the networks classify these with near-perfect accuracy, and that the CNNs together with germline similarity support fast humanization that aligns well with known experimental data.

The design choice that matters for interpretation is multiclass output. A Humatch score is not just "human or not"; the heavy and light models score a sequence across individual V-gene classes, so the answer carries a gene assignment. The paper's humanization procedure uses that to steer a sequence

toward a specific target gene and away from others, and the authors give a concrete reason: guidance of that kind keeps final designs from sitting "between" genes, which is a trait not naturally observed. Their argument is that optimising toward a specific gene with good VH/VL pairing improves the chances a design is stable and expresses well, and reduces the chance of immunogenic epitopes forming between the two chains.

This is a different and narrower claim than "this sequence is 94% human", and it is more actionable. If a classifier tells you a chain reads as a blend of two human V-genes, the remedy is to pick one and commit to it.

Humatch's lineage matters too. The same group previously published Hu-mAb, a set of machine-learning classifiers that discriminate human from non-human variable-domain sequences using large-scale repertoire data. The Hu-mAb paper reports two findings worth carrying forward: the classifiers consistently outperformed the then-best-in-class model for separating human from murine sequences, and their output scores showed a negative relationship with the experimental immunogenicity of existing antibody therapeutics. The second finding is the strongest honest statement available for a classifier-style humanness score — a relationship across known therapeutics, not a prediction for your molecule.

Repertoire Peptide Search: What OASis Measures#

OASis is the humanness metric published alongside BioPhi in 2022, and it takes the opposite approach: instead of scoring a chain against gene models, it decomposes the variable domain into overlapping 9-mer peptides and asks how often each peptide appears in human antibody repertoires collected in the

Observed Antibody Space (OAS).

OAS itself is worth knowing about independently. It was created in 2018 to offer cleaned, annotated and translated B-cell receptor repertoire data in place of raw sequencing files, and the 2021 update added nucleotides and amino acids for every sequence, annotation sufficient for MiAIRR compliance, comments on potentially problematic sequences, and paired VH/VL data. A repertoire-based humanness score is therefore a statement about a specific corpus of observed human sequences — which is both its strength and the thing to disclose when you quote one.

Three properties follow from the 9-mer design:

• It decomposes. Because the unit is a peptide, the score localises. You can say which stretch of sequence is responsible rather than reporting one opaque number for the chain.
• It is interpretable in immunological units. A 9-mer is roughly the length scale of a peptide presented to a T-cell receptor, which is a more defensible bridge toward immunogenicity than global identity.
• It credits maturation. The reference is observed rather than germline, so a residue that is rare in germline but common in matured human antibodies is not treated as a defect.

The BioPhi authors state their validation precisely, and it is worth quoting precisely: OASis "separated human and non-human sequences with high accuracy, and correlated with clinical immunogenicity". Separately, and about a different component, they report that on an in silico humanization benchmark of 177 antibodies their Sapiens model "produced sequences at scale while achieving results comparable to that of human experts". Those are two claims about two tools that happen to ship together — a scoring claim and a rewriting claim — and conflating them is the most common misreading of that paper.

OASis percentiles are not computed on .bloom*. The platform runs BioPhi's Sapiens rewrite, not its OASis metric, so if an OASis percentile is a required field in your filing or your internal scorecard, run the open-source BioPhi package in your own environment and keep that number as its own artifact.

Why the Three Methods Disagree#

The disagreements are systematic, not noise, and each pattern has a diagnosis.

A chain scores well on the gene classifier and badly on peptide search. The chain sits close to one human V-gene overall, but contains a short stretch — very often in or adjacent to a CDR — whose exact 9-mers are rare in observed repertoires. The classifier averages over the chain; the peptide score does not. Diagnosis: a local liability in an otherwise human framework.

A chain scores well on peptide search and ambiguously on the gene classifier. Every peptide is individually unremarkable, but the chain as a whole does not read cleanly as any single human gene — the "between genes" shape Humatch's authors explicitly design against. Diagnosis: a design or graft that mixed frameworks.

Germline identity is high and both model-based scores are mediocre. The sequence is close to a germline gene, but close in the uninformative way: the residues that differ are the ones that matter, or the chain is germline-like and still an unnatural combination. Diagnosis: read the alignment by hand.

Germline identity is mediocre and the model-based scores are fine. This is the expected signature of a genuinely human, somatically matured antibody, and it is the case that most argues against using identity alone as a gate.

The practical rule: use two methods with different references, and treat a split as a localisation instruction rather than a tie to be broken by averaging.

What None of Them Predicts#

Immunogenicity. The honest summary of the literature is that these scores correlate with clinical immunogenicity across sets of known therapeutics — reported for OASis, and as a negative relationship with experimental immunogenicity for the Hu-mAb classifiers. A correlation across a reference set does not become an anti-drug-antibody rate for your candidate. Worse, the residual risk may live where scoring cannot help: an analysis of CD4+ helper T-cell epitopes across eight humanized antibodies found epitopes only in CDR-containing regions, and the authors proposed that some immunogenicity will always be present because of the nature of the antigen-combining site. They also showed that one or two amino-acid modifications within an epitope could reduce immunogenic potential while retaining full biological function — which is a statement about careful epitope-level engineering, not about raising a humanness score.

Binding. Humanness scoring is blind to affinity. The residues that most depress a humanness score are frequently the ones carrying specificity. Any humanization worklist has to be intersected with what you know about the binding site before it becomes a mutation plan.

Developability. A humanised sequence can still aggregate, carry an unworkable charge profile or express poorly. That is a separate axis with its own computational guidelines — the five developability metrics proposed for therapeutic antibody profiling are the usual starting point — and humanization can move all of them, because changing framework residues changes surface chemistry.

Where the threshold should sit. Every one of these scores is continuous and every pass/fail line drawn across one is a convention. A scoring job's default threshold is a starting convention, not a regulatory criterion; decide yours against your own historical candidates and record it with the score.

How to Run It on .bloom*#

The platform runs humanness scoring and humanization as asynchronous jobs, not as an interactive notebook session.

Scoring with the gene-specific classifier — humatch_score. You supply one to 50 variable-region amino-acid sequences, all of the same chain type, and the chain type is load-bearing: H selects the heavy-chain model and L the light-chain model, and mislabelling a chain produces a confident, meaningless number. You can also name each sequence, choose the numbering scheme used for alignment (IMGT, Chothia or Kabat), and set the score threshold that decides the human / not-human call. The full parameter list is in the Humatch inputs reference.

Per sequence, the job returns the overall humanness score, the pass/fail call against your threshold, the chain type it scored, scores across the V-gene classes, the closest human germline V-gene, and a normalised germline-likeness score for that gene. At job level you get the scheme and threshold that were actually used and a count of sequences that failed to score. Two consequences for planning: because one job scores one chain type, a heavy-and-light read is two jobs, and because the output is gene-level rather than per-residue, you localise by comparing against the germline gene the job names, not by reading a residue heat map out of the result.

Rewriting with Sapiens — biophi_humanize. This is the act-on-it step, and it is a genuinely different tool rather than a mode of the scorer. You give it a heavy-chain variable region (and optionally the light chain), or the name of a known therapeutic antibody, which the platform resolves to VH/VL sequences for you out of the Therapeutic Structural Antibody Database (Thera-SAbDab). Sequences are variable region only, roughly 90 to 130 residues — do not paste the constant region. You can set the target humanness to rewrite toward. What comes back is the humanized VH/VL and the list of mutations the model made, position by position. Parameters are in the BioPhi inputs reference. Because .bloom* does not compute OASis percentiles, a Sapiens run here reports the rewrite and its mutations and does not report an OASis-derived immunogenicity-reduction figure — which is the correct behaviour: that number cannot be derived from a mutation count.

A workable order of operations. Score the candidate with humatch_score and note the gene assignment and the germline-likeness number. Decide which gene you are humanizing toward. Rewrite with biophi_humanize, holding your own list of binding-critical positions outside the tool as the constraint you check the returned mutations against. Re-score the rewrite with humatch_score — the point of scoring with a method that did not produce the sequence — and keep both scores, both thresholds and the mutation list together as the record of the decision. If a filing needs an OASis percentile, produce it from the open-source package and attach it alongside.

Access. Both jobs run on the free tier and are on the platform's guest-allowed list, so you can run them without an account; a signed-in run is what keeps the record you can come back to and chain onward. Which tool? maps a goal to a job if you are not sure where to start.

For neighbouring reading on how computational predictions should and should not stand in for measurements, see ESMFold vs AlphaFold on structure prediction accuracy, and the open-source computational drug discovery tool survey for where humanness scoring sits in a wider stack.

Key Takeaways#

• The reference set is the method. Germline theory, learned gene classes and observed repertoires are three different yardsticks, and a score is only interpretable if you say which one produced it.
• Germline identity is a control, not a gate. It is position-blind and it penalises ordinary somatic maturation.
• A gene-specific classifier answers "which gene", which is actionable. Humatch's multiclass output exists so a design can be steered toward one human gene instead of landing between two.
• A peptide-level score localises. That is what OASis buys you; on .bloom* it is not computed, so run the open-source package when you need the percentile.
• Scoring and rewriting are separate steps. Score with a method that did not generate the sequence, or you are grading your own homework.
• No humanness score predicts immunogenicity. The published relationships are correlations over known therapeutics, and residual T-cell epitopes have been found precisely in CDR-containing regions.
• Intersect any worklist with binding and developability before it becomes a mutation plan.

Pick two references rather than one, write the threshold down next to the score, and keep the mutation list with both. Then run the humanness score and the Sapiens rewrite on your own sequences on the .bloom* platform.

References#

1. Chinery, L., Jeliazkov, J.R. & Deane, C.M. "Humatch - fast, gene-specific joint humanisation of antibody heavy and light chains." mAbs 16(1), 2434121 (2024). doi:10.1080/19420862.2024.2434121 · code: oxpig/Humatch
2. Marks, C., Hummer, A.M., Chin, M. & Deane, C.M. "Humanization of antibodies using a machine learning approach on large-scale repertoire data." Bioinformatics 37(22), 4041–4047 (2021). doi:10.1093/bioinformatics/btab434
3. Prihoda, D., Maamary, J., Waight, A., Juan, V., Fayadat-Dilman, L., Svozil, D. & Bitton, D.A. "BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning." mAbs 14(1), 2020203 (2022). doi:10.1080/19420862.2021.2020203 · code: Merck/BioPhi
4. Olsen, T.H., Boyles, F. & Deane, C.M. "Observed Antibody Space: A diverse database of cleaned, annotated, and translated unpaired and paired antibody sequences." Protein Science 31(1), 141–146 (2021). doi:10.1002/pro.4205
5. Kovaltsuk, A., Leem, J., Kelm, S., Snowden, J., Deane, C.M. & Krawczyk, K. "Observed Antibody Space: A Resource for Data Mining Next-Generation Sequencing of Antibody Repertoires." The Journal of Immunology 201(8), 2502–2509 (2018). doi:10.4049/jimmunol.1800708
6. Harding, F.A., Stickler, M.M., Razo, J. & DuBridge, R.B. "The immunogenicity of humanized and fully human antibodies: residual immunogenicity resides in the CDR regions." mAbs 2(3), 256–265 (2010). doi:10.4161/mabs.2.3.11641
7. Jain, T., Sun, T., Durand, S. et al. "Five computational developability guidelines for therapeutic antibody profiling." Proceedings of the National Academy of Sciences 116(4), 4025–4030 (2019). doi:10.1073/pnas.1810576116
8. Raybould, M.I.J., Marks, C., Lewis, A.P. et al. "Thera-SAbDab: the Therapeutic Structural Antibody Database." Nucleic Acids Research 48, D383–D388 (2019). doi:10.1093/nar/gkz827