On August 13, 2026 Communications AI and Computing, a Nature portfolio journal, published “Preferential CDR masking in paired antibody language models improves binding affinity prediction” from a Boston University-led team. Therapeutic antibodies are among the leading classes of biologics, but their architecture is awkward for general protein language models: each antibody pairs a heavy and a light variable domain, with conserved framework regions that hold the structure together and hypervariable complementarity-determining regions, the CDRs, that actually contact the antigen. Existing protein language models mask and learn across all regions uniformly, which spends capacity on the scaffolding.
The team built a region-aware pretraining strategy for paired variable-domain sequences and compared three masking approaches: uniform whole-chain masking, CDR-focused masking, and a hybrid. They ran it on two backbones, a 3-billion-parameter ESM2 model and a compact 600-million-parameter ESM C model. Final models were trained on more than 1.6 million paired antibody sequences and evaluated on binding affinity datasets covering over 90,000 antibody variants across six antigens, including single-mutant panels and combinatorial libraries.
CDR-focused training produced embeddings with better predictive performance for antibody-antigen binding, reaching up to 27 percent improvement over benchmarked antibody models. Two secondary findings are arguably more consequential than the headline number. Training exclusively on paired sequences was sufficient: pretraining on billions of unpaired sequences provided no measurable benefit. And the compact 600-million-parameter model matched or exceeded larger antibody-specific baselines.
The pattern is a counterexample to scale-first assumptions in applied biology modeling. Here the win came from matching the masking objective to the part of the molecule that carries the function, and from insisting on the right data type rather than the largest corpus. For a biotech or pharma team, that translates into a smaller model that is cheaper to train and to run at screening scale, and into a data strategy that prioritizes paired heavy-light sequencing over volume. The affinity-prediction step this improves sits early in antibody engineering, where better ranking of candidate variants directly reduces the number of wet-lab assays needed per program.