How does a 167M-parameter protein foundation model outperform a 650M general-purpose model on the tasks that matter? AINN-P1 is sequence-first, mLSTM-based, and designed from the ground up for antibody engineering, not borrowed from NLP.
Ainnocence Inc., an AI-driven drug discovery company, today announced new results for AINN-P1, a purpose-built protein foundation model designed to deliver efficient and transferable predictions for protein engineering and biologics discovery.
AINN-P1 takes a sequence-first approach, generating protein representations directly from amino acid sequences, without requiring multiple sequence alignments, structure prediction, or external functional annotations. Conventional pipelines rely on homolog search, which is slow and sensitive to the quality of the reference database. AINN-P1 removes that dependency.
Architecturally, the model is built around an mLSTM backbone. Transformer all-to-all attention requires n² connections; mLSTM recurrent state passing requires n steps, with a memory footprint that does not grow with sequence length. This makes the model fast on long sequences (full antibodies, scFvs, trispecifics) where attention-based models become expensive.
ProteinGym consolidates deep mutational scanning assays across four task categories, binding, activity, expression, and stability. AINN-P1 leads the stability track, which is the gating property for biologics developability: a candidate that is unstable in production or in serum is a candidate that cannot ship.
Antibody discovery data are organized into programs sharing large stretches of framework sequence. Under a random split, close relatives land on both sides of the split, a model can score well just by learning program identity. Under leave-program-out (LPO), only transferable biophysics can win.
Ainnocence evaluated VHH single-domain antibody expression under both protocols on proprietary data from real discovery programs:
| Model | Params | Random split AUC | LPO AUC | Finetuning |
|---|---|---|---|---|
| ESM2-base (frozen) | 650M | 0.78 | 0.66 | None |
| ESM2 (fine-tuned) | 650M | 0.86 | 0.816 | Task-specific |
| AINN-P1 (frozen) | 167M | 0.84 | 0.810 | None |
ROC-AUC for VHH expression classification (chance = 0.50); proprietary Ainnocence discovery-program data. ESM2-ft = ESM2-650M fine-tuned on the task; ESM2-base = frozen ESM2-650M embeddings; AINN-P1 = frozen embeddings, no task-specific fine-tuning.
AINN-P1 beat the 650M ESM2 baseline by +0.15 AUC on unseen programs (0.810 vs 0.660) with 3.9× fewer parameters, and matched the fine-tuned ESM2 within 0.006 AUC (0.810 vs 0.816), without any task-specific fine-tuning. The AUC drop from random split to new programs is the leakage signal: our architecture loses less to it.
Ablation across 7 classifiers × 3 encoders (21 configurations): ESM2-base was weakest at every classifier position, the effect is a property of the embeddings, not the head fitted on top.
Antibody–antigen affinity ranking, from sequence alone. Using AINN-P1 as a frozen encoder, mean five-fold cross-validated Spearman ρ for ΔΔG ranking rose from 0.417 (task-specific network trained from scratch) to 0.533, a ≈28% relative gain under identical data and framework.
This matters because affinity is where drug discovery lives or dies. A model that encodes sequence into representations that preserve binding geometry, without ever seeing a structure, is a model that can enter a program on day one and start ranking real candidates on day two.
Full technical details, benchmark protocols, and ablation studies are in the preprint. The model is one of several sequence-first foundation models powering the Ainnocence discovery stack, alongside CarbonAI (chemistry), PeptideAI (peptides), CellulaAI (cell therapy), SenseAI (RNA), and NatmolAI (natural products).
Ainnocence is an AI-native biotech company founded in 2021 in the Bay Area. Our platforms cover antibody, small molecule, peptide, cell therapy, natural product, and RNA modalities. Every lead is wet-lab validated. Every claim is published or produced on demand.
Pilot engagements typically scope in one week. We'll run a representative campaign against a target you specify and share the raw results.
Start a pilot →