ImmunoMatch

An antibody is an assembly of two different polypeptides — a heavy chain and a light chain — and not every combination folds into a stable molecule. In vivo, developing B cells solve this problem through successive rounds of rearrangement and selection in the bone marrow, discarding pairs that do not assemble. In vitro, antibody engineers face the same constraint from the opposite direction: humanisation, chain shuffling, and library design all mix and match chains, and a pair that fails to assemble wastes an expression campaign. Most antibody developability work has focused on aggregation, immunogenicity, and expression yield; the compatibility of the two chains themselves has been comparatively under-explored.

ImmunoMatch, from the Fraternali lab at University College London with collaborators at the University of Surrey, King's College London, and Birkbeck, University of London, addresses that gap directly. It is a classifier fine-tuned on AntiBERTa2 — the successor to AntiBERTa, an antibody-specific protein language model — that takes a heavy and a light chain variable domain and scores how likely they are to be a cognate pair. Training data comes from single human B cells, where the natural pairing is known; negatives are constructed by shuffling chains between cells.

The authors' central finding is a biological one as much as a methodological one: training separate models for the two human light chain isotypes, κ and λ, improves prediction, mirroring the sequential mechanism by which B cells select light chains during development. The work was posted as a preprint in February 2025 and published in Nature Methods later that year.

#Key Features

• Cognate-pair scoring: The model outputs a compatibility score for an arbitrary heavy-light pair, turning chain pairing into a quantity that can be ranked across a design library.
• Isotype-specific checkpoints: Three published models — a general one trained on mixed light chains, plus κ-specific and λ-specific variants — reflect the distinct selection regimes the two isotypes go through in the bone marrow.
• Interface sensitivity: Predictions respond to sequence differences at the heavy-light interface, indicating the model has learned residue-level determinants of assembly rather than a coarse repertoire prior.
• Maturation as a readout: Applied across B cell developmental stages, ImmunoMatch scores rise with maturity in both healthy and disease repertoires, so the model doubles as a probe of the selection process itself.
• Pairing recovery from unpaired data: The published models reconstruct paired antibodies from spatial VDJ sequencing data, where the physical link between the two chains is not preserved.

#Technical Details

ImmunoMatch is a sequence-pair text-classification head fine-tuned from alchemab/antiberta2-cssp, a RoFormer-based antibody language model, and implemented in PyTorch with the Transformers library. It consumes the paired VH and VL domain sequences and emits a binary cognate/non-cognate call with an associated score. Positive examples are the natural heavy-light pairs observed in single human B cells; negatives are generated by randomly pairing chains drawn from different cells. Three checkpoints are distributed on Hugging Face — fraternalilab/immunomatch for mixed light chains, plus -kappa and -lambda — all under CC-BY-NC-4.0, which restricts commercial use.

Inference requires no retraining: a Google Colab notebook scores individual pairs or batch annotates sequences from a CSV, and the framework is also distributed as a PyPI package. The GitHub repository carries a custom licence file rather than a standard open-source licence, so the code terms should be read alongside the non-commercial model licence.

#Applications

For therapeutic antibody discovery, ImmunoMatch offers a cheap in silico filter ahead of expression: ranking candidate chain combinations in a shuffled or humanised library so that pairs unlikely to assemble are dropped before they consume cloning and expression capacity. For repertoire immunology, it provides a way to reconstruct plausible heavy-light pairs from bulk or spatially resolved sequencing where the pairing was lost, and to quantify how pairing preferences shift across B cell development, vaccination, or autoimmune and malignant states.

#Impact

ImmunoMatch stakes out heavy-light chain compatibility as a developability property in its own right, and gives the field a concrete tool for it. The κ/λ result is the more interesting scientific claim, connecting a modelling choice to the ordered mechanism of light chain rearrangement in the bone marrow. Publication in Nature Methods and three ready-to-run checkpoints lower the barrier to adoption considerably, though the CC-BY-NC-4.0 model licence puts commercial antibody engineering — the most obvious application — outside the released terms without a separate arrangement. Validation is computational, on repertoire and spatial sequencing datasets, rather than on expression yields measured at the bench.