The Pulse
Nona Builds a 366M-Parameter Model for Human Heavy-Chain Antibodies
Nona Biosciences says its HCAbLM model was trained on 31.8 million fully human heavy-chain-only antibody sequences. The company reports that the 366-million-parameter system outperformed larger general-purpose and antibody models on develop
AI.info Team ·
Nona Puts a Specialized Model Against Larger Protein Systems
Nona Biosciences says it has developed HCAbLM, a language model trained specifically on fully human heavy-chain-only antibodies rather than on broad protein or conventional antibody datasets. The Cambridge, Massachusetts, company announced the model on September 16, 2026, alongside a bioRxiv preprint titled A foundation model learns the sequence and functional grammar of fully human heavy-chain-only antibodies.
The claim sets up a direct challenge to the usual assumption that larger general-purpose models are automatically better suited to biological sequence problems. Nona says HCAbLM contains 366 million parameters, yet outperformed Meta’s 6-billion-parameter ESM-6B and antibody models including IgLM and AbLang on two public developability measures: size-exclusion chromatography purity and hydrophobic-interaction chromatography behavior.
The company did not provide the benchmark scores in its announcement. Nona’s release describes the comparisons as cross-project developability tests and says HCAbLM also shows transferability to thermal-stability measurements.
31.8 Million Sequences From 73 Transgenic Mice
HCAbLM was trained on 31.8 million fully human heavy-chain-only antibody sequences collected from 73 independently immunized HCAb transgenic mice. Nona says the dataset gives the model access to sequence patterns that general-purpose protein systems rarely encounter because their training data contain few heavy-chain-only antibodies.
Heavy-chain-only antibodies differ from conventional antibodies built from paired heavy and light chains. Nona’s Harbour Mice platform is designed to generate fully human antibodies in both conventional two-heavy, two-light-chain formats and heavy-chain-only formats, which can produce compact variable heavy-domain binders for use in therapeutic designs.
The company describes the sequence patterns learned by HCAbLM as a specialized “sequence grammar.” In practical terms, Nona says the model is intended to identify features associated not only with binding, but also with whether a candidate folds correctly, remains stable and can meet manufacturing requirements.
Developability, Not Binding, Is Nona’s Target
Nona frames the problem as a shift from finding antibodies that bind a target to finding candidates that can become usable drug molecules. Its release identifies manufacturing at scale, aggregation and stability as the main development concerns that HCAbLM is intended to help assess.
That distinction matters because a sequence can show useful binding while still failing later because it expresses poorly, aggregates or lacks sufficient thermal stability. Nona says the model’s representations correlate with experimentally measured properties, including SEC purity, HIC behavior and thermal stability, although the announcement does not disclose the size of the evaluation sets or the numerical results for each property.
“By learning the unique sequence characteristics of fully human HCAbs and demonstrating transferability to experimentally measured molecular properties, HCAbLM provides a new foundation for AI-enabled antibody discovery and development.”
Dr. Di Hong, chief executive officer, Nona Biosciences
Why the Model’s Smaller Size Matters
Nona’s comparison with ESM-6B puts model size at the center of its argument. HCAbLM uses 366 million parameters, while the ESM-6B system cited by the company uses 6 billion, but Nona says the smaller model achieved better results on the two reported developability metrics.
Parameter count alone does not determine performance on a narrow biological task. A model trained on a large, format-specific repertoire may capture constraints that a much larger system trained on broader data does not represent well. Nona’s announcement presents HCAbLM as evidence for that approach, though independent replication will be needed to establish how broadly the result holds.
HCAbLM Joins Nona’s Drug-Discovery Stack
Nona says it plans to integrate HCAbLM with its existing Harbour Mice technology, single-B-cell screening, NonaCarFx functional screening and Hu-mAtrIx AI platform. The company positions the model as a tool for evaluating and optimizing fully human heavy-chain-only antibodies, not as a standalone drug candidate or clinical product.
Its Harbour Mice platform supports applications that include bispecific and multispecific antibodies, CAR-T therapies, antibody-drug conjugates and mRNA-based therapeutics. Those uses remain part of Nona’s platform description; the September 16 announcement does not identify a specific therapeutic program advanced with HCAbLM.
The immediate result is a specialized model, a large antibody sequence collection and a company-reported benchmark advantage over larger systems. The preprint now provides the technical record behind those claims, while the next test is whether HCAbLM can improve experimentally validated antibody candidates beyond the benchmark setting.