Expression — the ability to produce an antibody at usable yield — is an early and costly filter in therapeutic antibody discovery. The authors present AINN-Express, a sequence-only predictor designed to estimate the probability that a VHH single-domain antibody (nanobody) will express at useful levels. The model is intended as a practical triage tool to prioritize candidates before wet-lab experiments, minimizing experimental cost and time.
AINN-Express encodes an input VHH amino-acid sequence using a frozen encoder from the AINN-P1 protein foundation model developed by Ainnocence. The encoder outputs fixed embeddings that are scored by a lightweight gradient-boosted classifier to produce an expression probability. Key characteristics of the approach are that it is sequence-only (no structure required), uses a frozen foundation encoder (no per-task encoder training), and applies a simple downstream classifier, enabling rapid scoring without extensive task-specific training.
To assess generalization in a manner that reflects real-world discovery programs, the authors use a leakage-safe, leave-program-out evaluation. This protocol withholds entire antibody programs during testing so that the model must generalize to sequences from programs not seen during training. The study argues that this leakage-aware split is crucial because random splits can permit models to learn program identity signals rather than biologically transferable determinants of expression.
Under the leakage-aware, leave-program-out evaluation, AINN-Express achieved a ROC-AUC of 0.87 when evaluated within known antibody programs and 0.81 on entirely new programs. Both results are reported as substantially above the majority-class baseline and support the claim that AINN-Express can be useful for candidate prioritization prior to wet-lab validation.
The authors performed a controlled, leakage-aware benchmark comparing AINN-P1 to general-purpose protein language models. On the VHH expression task, AINN-P1 (167 million parameters) generalized to unseen programs with ROC-AUC 0.81, outperforming a general-purpose ESM2 model (650 million parameters), which achieved ROC-AUC 0.68 on new programs. A domain-finetuned ESM2 matched or slightly exceeded AINN-P1 on new-program ROC-AUC (0.83) but required task-specific fine-tuning. Notably, under a random-split evaluation, all encoders scored approximately 0.88 ROC-AUC, masking the differences observed under leave-program-out testing.
The comparative results are used to support the interpretation that purpose-built representation quality, rather than raw parameter count, drives transferability to unseen antibody programs for the expression task. The authors argue that general-purpose embeddings can encode program identity signals that inflate performance under naive random splits; only leakage-aware evaluation exposes whether embeddings capture transferable determinants of expression.
AINN-Express is positioned as a practical pre-wet-lab triage tool: it requires only an amino-acid sequence, produces an expression probability, and does not demand structure prediction or per-task encoder fine-tuning. Reported ROC-AUCs (0.87 within known programs; 0.81 on novel programs) suggest it can meaningfully rank candidates before investment in wet-lab expression studies. The model’s design — frozen foundation encoder plus a lightweight classifier — aims to balance predictive performance and operational simplicity.
All authors are affiliated with Ainnocence, Inc., which develops the AINN-P1 protein foundation model and funded the work; this affiliation is declared as a competing interest. The preprint was posted on bioRxiv on July 24, 2026, and is available with DOI https://doi.org/10.64898/2026.07.21.739256. The source indicates that the copyright holder is the author/funder and that reuse is not permitted without permission.
Details beyond those reported in the preprint — such as dataset size, exact classifier hyperparameters, per-program data characteristics, or further ablations — were not reported in the source and therefore are not described here. The authors emphasize the importance of leakage-aware evaluation when characterizing generalization for antibody discovery tasks.