AINN-Express

Whether an antibody can be produced at usable yield is one of the earliest and most expensive filters in therapeutic discovery: a candidate that binds well but expresses poorly still consumes cloning, transfection, and purification effort before it is discarded. AINN-Express, from Ainnocence, moves that filter in silico for VHH single-domain antibodies (nanobodies). It takes an amino-acid sequence and returns an expression probability — no structure, no antigen context, and no per-task model training.

Architecturally the model is a thin supervised head on a frozen foundation model. Sequences are encoded by AINN-P1, Ainnocence's 167M-parameter protein language model, and the resulting representation is scored by a lightweight gradient-boosted classifier. The encoder is never updated and the classifier is trained once, so new VHH sequences — and entire new antibody programs — are scored without retraining anything. The design choice is the argument: the authors contend that the useful signal already lives in a purpose-built representation, and that a heavier task-specific fine-tune is not what makes the predictor generalize.

The paper's most transferable contribution is its evaluation design. AINN-Express is assessed leave-program-out, on antibody programs never seen during training, rather than by a random sequence split. Under a random split, every encoder the authors tested scored roughly 0.88 ROC-AUC — an apparent tie that dissolves once whole programs are held out. Their reading is that general-purpose protein embeddings largely encode program identity rather than transferable determinants of expression, and that random splits quietly reward that leakage.

#Key Features

• Sequence-only prediction: The only input is an amino-acid sequence. No predicted or experimental structure, no multiple sequence alignment, and no antigen are required.
• Frozen encoder, no per-task training: A trained-once gradient-boosted head sits on the frozen AINN-P1 encoder, so novel sequences and novel programs are scored zero-shot with no fine-tuning step in the loop.
• Leakage-safe benchmarking: A leave-program-out protocol holds out complete antibody programs, separating genuine transfer from memorized program identity.
• Representation quality over parameter count: The frozen 167M-parameter encoder matches a domain-finetuned 650M-parameter baseline on unseen programs at roughly a quarter of the parameters.
• Probability output for triage: The classifier emits an expression probability rather than a bare label, so candidate panels can be ranked and thresholded to fit a team's throughput.

#Technical Details

Under leave-program-out evaluation, AINN-Express reaches ROC-AUC 0.87 within known antibody programs and 0.81 on entirely new programs, both well above the majority-class baseline. A controlled, leakage-aware encoder comparison isolates the contribution of the representation itself: swapping in a general-purpose ESM-2 encoder at 650M parameters drops new-program ROC-AUC to 0.68, while an ESM-2 fine-tuned on the antibody domain reaches 0.83 — a score the frozen 167M-parameter AINN-P1 encoder essentially matches at 0.81 with no task-specific fine-tuning. All encoders converge to about 0.88 under a random split, which is precisely the point the authors use to argue that random splits are the wrong measurement for this task.

#Applications

The intended use is prioritizing nanobody candidates before wet-lab expression work — triaging panels from immunization campaigns, synthetic libraries, or computational design so that scarce expression and purification capacity goes to sequences most likely to produce material. Because scoring requires nothing but the sequence, it can run early, in bulk, and alongside affinity and other developability predictors in a discovery funnel. As a proprietary model from a drug discovery company, its practical availability is through Ainnocence's platform rather than as a self-hosted tool.

#Impact

AINN-Express contributes to a growing insistence on leakage-controlled evaluation for protein property prediction, a concern it shares with work such as Aiki-XP on bacterial protein expression: the headline finding is less that one encoder wins than that the standard random split cannot tell encoders apart at all. Several constraints bound how far the result travels. The work is a bioRxiv preprint that has not been peer reviewed, and it is posted under an all-rights-reserved license with no accompanying code, weights, model card, or data card, so the numbers cannot be independently reproduced. The evaluation is single-lab and internal rather than an external community benchmark, and the model's scope is narrow by construction: VHH single-domain antibodies only, sequence in and expression probability out, with no handling of structure, antigens, or other developability properties.