Heavy2Light

Heavy2Light is a conditional generative model that writes an antibody light-chain variable sequence given a heavy chain. An antibody's binding site is formed jointly by both chains, but the public sequence record is badly lopsided: repertoire databases hold billions of unpaired heavy or light chains from bulk sequencing and only a few million natively paired examples, because pairing requires single-cell methods. That imbalance makes it hard to learn which light chains actually go with which heavy chains, a question that matters both for interpreting immune repertoires and for reconstituting therapeutic candidates from heavy-chain-only data.

The model was developed at the Institute of Biochemistry and Molecular Medicine at the University of Bern by Lea Brönnimann, Chiara Rodella, and Thomas Lemmin, and posted to bioRxiv in August 2025. Its strategy is to spend the abundant unpaired data first and the scarce paired data last. Two chain-specific language models are pretrained separately — HeavyBERTa, a RoBERTa-style masked language model, and LightGPT, a GPT-2-style causal language model — and only then joined into an encoder-decoder that treats heavy-to-light pairing as a machine translation task, with cross-attention and lightweight adapters carrying the small amount of paired supervision.

Within the catalog it sits alongside other antibody language models with different objectives: the CDR-masked paired antibody language model learns binding-affinity embeddings rather than generating chains, and efficient evolution uses general protein language model likelihoods to suggest affinity-maturing mutations. Heavy2Light is generative and specifically about inter-chain dependency.

#Key Features

• Two-stage training: Chain-specific pretraining on unpaired repertoire data, followed by adapter-based fine-tuning on paired sequences, sidestepping the scarcity of native pairs.
• Parameter-efficient coupling: Only bottleneck adapters and cross-attention parameters are trained during fine-tuning; both pretrained backbones stay frozen.
• Conditioning improves plausibility: Light chains generated with a heavy chain in context fold better than unconditioned samples, reaching mean pTM 0.939 and pLDDT 91.7 versus 0.775 and 86.5.
• Maturation-state classifiers: The same backbones fine-tune into naive-versus-memory B cell classifiers, reaching 92.31% accuracy on heavy chains.
• Recovers immunological structure: Memory-derived heavy chains generate light chains with restricted V gene usage, reproducing known maturation-dependent pairing constraints.

#Technical Details

HeavyBERTa was pretrained by masked language modeling on more than 99 million human heavy-chain sequences from the Observed Antibody Space, restricted to healthy donors with no recent vaccination; LightGPT was pretrained by autoregressive next-token prediction on more than 22 million light chains from the same source. Two HeavyBERTa configurations were trained: a small one (4 layers, hidden size 512, 13.15M parameters) and a larger one (12 layers, hidden size 768, 86.06M parameters), reaching 88.95% and 89.12% masked-residue accuracy. LightGPT uses a GPT-2 configuration with 12 layers and hidden size 768 (85.86M parameters) and reaches 86.35% accuracy. Both use a 25-token vocabulary. Heavy2Light combines the small HeavyBERTa as encoder with LightGPT as decoder through cross-attention and bottleneck adapters with reduction factor 16, fine-tuned on 588,388 paired sequences curated from OAS and the Patent and Literature Antibody Database under an 80/10/10 split, using nucleus sampling at top-p 0.85 and temperature 0.8 for generation.

Recovery of the native partner light chain is moderate and uneven across the domain: mean identity to the true sequence is 60.73% overall, with framework regions at 58.54-72.90% and CDRs at 33.40-40.29%. Structural quality is nonetheless preserved — sequences folded with Chai-1 superimpose on their native counterparts with correct immunoglobulin topology even at 30.97% identity. Generated light chains reach 93.10% germline identity and correlate with their input heavy chain's germline identity at Pearson r = 0.440, against 0.593 for native pairs and -0.002 for shuffled controls. Kappa chains show a trimodal identity distribution not explained by V gene usage or B cell state.

#Applications

Heavy2Light lets researchers propose plausible light-chain partners for heavy chains recovered from bulk repertoire sequencing, expanding single-chain datasets into candidate pairs for downstream folding, expression, or screening. The maturation classifiers annotate B cell state directly from sequence in datasets that lack sorting metadata. For therapeutic discovery, the conditional generator supplies diverse but germline-coherent light-chain candidates around a fixed heavy chain.

#Impact

The work establishes that heavy-light pairing carries learnable signal beyond V gene frequency statistics, and that a translation framing over frozen chain-specific language models can recover it from limited paired data. Its most cited caveat is the authors' own: sequence-level recovery is moderate, so generated chains are hypotheses requiring experimental validation, and the model is biased toward naive-like, germline-proximal light chains because public repertoires are dominated by them. The authors also show that interface pTM from a general folding model cannot distinguish correct from shuffled pairs, so structural scoring alone is not a validation shortcut. Code is on GitHub without a license file, all trained checkpoints are on HuggingFace, and the preprint has since been published in iScience.