Why Heavy–Light Chain Pairing Matters in Antibody Language Models

CD ComputaBio provides software-based computational services to support research and development. We do not offer free software packages.

Overview

Why One Chain Can Set the Pair-Level Signal

A therapeutic antibody is not adequately described by two independent labels. A high-scoring VH cannot compensate for a low-scoring VL if the operational requirement is that both chains resemble the human repertoire. Marks and colleagues made this explicit: a therapeutic was called human only when its VH and VL scores each exceeded the Youden J statistic threshold selected for the corresponding classifier. Among 176 therapeutics annotated as human, all but one passed this joint rule; the exception had a VL score of 0.850, just below its 0.856 threshold. All 14 mouse therapeutics were classified as non-human.

For R&D teams, the decision rule is as important as the individual score. Independent chain models can still support pair-level triage when results are retained separately and combined conservatively. Collapsing VH and VL too early into an average can conceal a problematic chain. Conversely, rejecting a pair from a single score does not prove biological failure: the score is a repertoire-derived classifier output, not an assay of T-cell activation, exposure, affinity, or clinical immunogenicity.

What Hu-mAb Models—and What It Does Not

The authors trained random forests rather than a neural language model. Models were specific to chain type and human V-gene family, with human sequences used as positives and more than 13 million mouse, rat, and rhesus sequences as negatives. Each classifier used an 80/10/10 training, validation, and test split and 200 trees. For light chains, a separate classifier first distinguished kappa from lambda; the paper reports perfect classification on the evaluated VL dataset. These design choices make the output interpretable as gene-conditioned humanness, not sequence generation probability.

The study also compared the random forests with positive-only long short-term memory models. Across the 22 evaluated models, the random forests achieved higher area under the receiver-operating-characteristic curve and higher Youden J statistics. This is an empirical result for the study's classifier setup, not evidence that random forests generally outperform current protein language models. Modern paired models optimize different objectives and can learn cross-chain context, but they still require external evaluation against therapeutics, repertoire data, structural constraints, and laboratory measurements.

Training Chain- and Gene-Specific Classifiers

Gene specificity reduces a major confounder in humanness scoring: a sequence should not be penalized merely because it belongs to a different legitimate human V-gene family. During humanization, Hu-mAb keeps the complementarity-determining regions (CDRs) fixed and greedily searches framework substitutions. At each iteration it evaluates all possible single framework mutations, selects the mutation producing the greatest increase in the relevant classifier score, and repeats until the target score is reached.

This algorithm is constrained rather than generative. It proposes a path through framework sequence space while preserving the starting CDRs. The model nevertheless used some CDR positions for classification, and the effect of a mutation depended on the rest of the sequence. Marks and colleagues interpreted this context dependence as evidence that the classifiers captured residue interactions. For a paired language model, the analogous requirement is broader: conditional effects should be learned not only within one chain but also across the VH–VL interface, and the training data must contain authentic pairings to support that inference.

Joint Classification and the ADA Relationship

For 217 therapeutics with ADA information, the authors related the minimum of the VH and VL humanness scores to reported immunogenicity. The linear relationship was weak but stronger than a previously reported sequence-based correlation: the paper reports an R² of 0.31, compared with 0.18 in the earlier analysis. The minimum-score formulation is central—whichever chain was less human-like determined the pair-level value.

In the high-score group (minimum chain score above 0.9; n=85), 94.12% of therapeutics had reported ADA below 10%. The corresponding proportions were 84.21% for positive scores at or below 0.9 (n=57) and 48.00% for the negative group (n=75). High ADA above 50% was reported for 1.75% of the middle group and 14.67% of the negative group; none is shown for the high-score group. These are retrospective associations assembled from heterogeneous therapeutic records. They do not establish that increasing a model score will reduce ADA in a prospective program.

Humanization Mutations Compared with Experiment

The authors evaluated 25 therapeutics for which both precursor and experimentally humanized sequences were available. Hu-mAb proposed 59% of the framework mutations found in the experimental VH sequences and 58% of those in the VL sequences. When comparison was made at mutated positions, 68% of proposed VH residues and 77% of proposed VL residues exactly matched the experimentally derived residues. Allowing amino acids of similar type increased agreement to 77% and 85%, respectively. The paper estimated random exact overlap near 2% and type-adjusted overlap near 5%.

A negative-control experiment reinforced the importance of gene context. When humanization was driven by an incorrect V-gene classifier, 19 heavy and eight light chains could not reach the target score. For chains that did reach it, the wrong model required, on average, 12 additional VH mutations or 14 additional VL mutations. The result supports matching the scoring model to the intended germline context; it does not show that fewer mutations always preserve affinity or function.

The VH–VL Interface Is a Functional Boundary

Framework residues at the chain interface deserve separate attention because they affect VH–VL orientation and can indirectly reshape the paratope. At interface positions, Hu-mAb's exact mutation overlap with experimentally derived sequences was 74% for VH and 96% for VL, compared with averages of 68% and 77% across all evaluated framework mutations. However, Hu-mAb proposed fewer interface changes: averages of 0.8 mutations for each chain, versus 1.6 in experimental VH sequences and 1.8 in experimental VL sequences.

This pattern is informative but not a direct structural validation. The high agreement suggests that repertoire-derived chain models can recover many humanization choices at interface sites, while the lower mutation count shows that a humanness threshold does not fully encode the experimental engineer's objectives. Pair-aware modeling should therefore evaluate interface packing, orientation, buried polar groups, and potential clashes alongside sequence likelihood. Structural modeling can prioritize hypotheses, but expression, binding, stability, and functional assays remain necessary.

Lessons for Paired Antibody Language Models

The Hu-mAb study supports these principles indirectly. Because its VH and VL models are separate, it cannot determine whether a heavy sequence is compatible with one particular light sequence beyond a downstream combination rule. A genuinely paired language model should be able to condition one chain on the other and should be benchmarked against unpaired baselines. Useful evaluations include recovery of native partners among decoys, sensitivity to chain swaps, interface-contact prediction, pair-conditioned mutation ranking, and prospective measurement of expression and antigen binding.

Generative likelihood should also remain distinct from developability evidence. A sequence can be probable under a repertoire model yet carry liabilities, lose parental affinity, or pair poorly in a specific construct. Conversely, an unusual but functional therapeutic may receive a low repertoire score. Decision systems should retain multiple calibrated outputs—humanness, pair compatibility, structural confidence, specificity, and physicochemical risk—rather than presenting one composite number as biological truth.

Limitations and Next Validation Steps

The non-human training set was dominated by mouse, rat, and rhesus sequences. The authors noted weaker behavior for antibodies originating from less represented species, including a rabbit-derived example and a rat VL example. The workflow was developed for conventional antibodies and may not transfer directly to VHH antibodies, asymmetric formats, or engineered chains with unusual pairing constraints. CDRs were held fixed, so Hu-mAb did not attempt affinity maturation or repair CDR-localized liabilities.

The ADA analysis is observational and heterogeneous across molecules, indications, dosing regimens, assay formats, and clinical contexts. Humanness is only one contributor to immunogenicity, and the paper did not prospectively demonstrate reduced ADA, preserved affinity, or improved developability for generated designs. A practical validation plan would compare chain-only and paired models on held-out authentic pairs, challenge them with partner swaps, model interface geometry, and then test a deliberately diverse sequence panel for expression, monomer content, thermal stability, binding kinetics, specificity, and relevant cellular function.

Overview of What CD ComputaBio Can Provide

The study maps to several computational research needs. The services below describe related support without implying that CD ComputaBio reproduced Hu-mAb, trained on the same data, or can guarantee humanization, immunogenicity, or pairing outcomes.

References

1. Marks C, Hummer AM, Chin M, Deane CM. Humanization of antibodies using a machine learning approach on large-scale repertoire data. Bioinformatics. 2021;37(22):4041–4047. https://doi.org/10.1093/bioinformatics/btab434.
2. Olsen TH, Boyles F, Deane CM. Observed Antibody Space: A diverse database of cleaned, annotated, and translated unpaired and paired antibody sequences. Protein Science. 2022;31:141–146. https://doi.org/10.1002/pro.4205.
3. Raybould MIJ, Marks C, Krawczyk K, et al. Thera-SAbDab: the Therapeutic Structural Antibody Database. Nucleic Acids Research. 2020;48(D1):D383–D388. https://doi.org/10.1093/nar/gkz827.

Article and figure license. The primary study is Open Access under the Creative Commons Attribution 4.0 International License, which permits commercial reuse and adaptation with attribution. Figures 1 and 2 reproduce selected panels from Figures 2 and 3 of Marks et al.; the underlying data, sequence labels, and scientific annotations were not changed.

* For Research Use Only.

Related Services