AINN-Express: A Leakage-Aware, Sequence-Only Predictor of VHH Antibody Expression Built on the AINN-P1 Protein Foundation Model

Abstract

Expression -- whether an antibody can be produced at usable yield -- is one of the earliest and most expensive filters in therapeutic discovery. We present AINN-Express, a sequence-only predictor of VHH single-domain antibody (nanobody) expression built on AINN-P1, Ainnocences protein foundation model. AINN-Express encodes a VHH with a frozen AINN-P1 encoder and scores it with a lightweight gradient-boosted classifier: it takes only an amino-acid sequence, returns an expression probability, and needs no structure and no per-task model training. Under a leakage-safe, leave-program-out evaluation, AINN-Express reaches ROC-AUC 0.87 within known antibody programs and 0.81 on entirely new programs -- well above the majority baseline -- making it a practical tool for prioritizing candidates before wet-lab work. We further justify the encoder choice with a controlled, leakage-aware benchmark against general-purpose protein language models: on this task, AINN-P1 (167 M parameters) generalizes to unseen programs far better than a general-purpose ESM2 (650 M) -- 0.81 versus 0.68 new-program ROC-AUC -- and matches a domain-finetuned ESM2 (0.83) with no task-specific finetuning, at roughly one-quarter of the parameters. The gap is invisible under a random split, where all encoders score [~]0.88; only leave-program-out evaluation reveals that general-purpose embeddings largely encode program identity rather than transferable determinants of expression. Purpose-built representation quality, not parameter count, is what makes AINN-Express generalize.