Ainnocence says sequence-only protein model boosts antibody ranking accuracy 28%

Ainnocence released research showing its AINN-P1 protein foundation model improved ranking of antibody-antigen binding affinity from sequence alone, with a 28% relative gain over a task-specific model in five-fold cross-validation. The company says the approach works without structures, alignments or fine-tuning and could speed antibody discovery workflows.

Why it matters: - Antibody affinity maturation depends on ranking many candidate variants with limited experimental data. - A sequence-only model can reduce reliance on co-crystal structures, docked models and multiple sequence alignments, which are often unavailable or noisy in early discovery. - Faster downstream training could make it easier to reuse the same foundation representations across targets and campaigns.

What happened: - Ainnocence Inc. said its AINN-P1 protein foundation model improved mean Spearman rank correlation for predicted change in binding free energy, or ΔΔG, from 0.42 to 0.53. - The result represents a relative gain of about 28% over a task-specific model trained from scratch on the same data. - The comparison used identical five-fold cross-validation. - The company released the findings in a preprint titled “AINN-P1: A Compact Sequence-Only Protein Language Model Achieves Competitive Fitness Prediction on ProteinGym.” - The paper lists Roger Wang, Kevin Jin and Lurong Pan as authors.

The details: - The study treated affinity maturation as a learning-to-rank problem, since decision-making depends on which antibody candidates advance to the wet lab. - Spearman rank correlation was the primary metric. - NDCG and AUC were tracked as secondary measures. - A simple linear model on frozen AINN-P1 embeddings reached a Spearman ρ of 0.457. - That linear probe outperformed a task-specific network trained end-to-end from scratch, which scored 0.417. - The result suggests the gain came from representation quality rather than added model capacity. - The model used sequence only and did not require co-crystal structure, docked structure or multiple sequence alignment at any stage. - Training the downstream head took seconds per fold with the foundation model held frozen. - The team fit feature normalization only on training folds to avoid information leakage. - The evaluation used a curated antibody-antigen ΔΔG dataset.

Between the lines: - The strongest signal in the study is that a frozen encoder beat a fully trained task-specific model on the same labels. - That makes the result a test of protein representations, not just model size or optimization. - For novel antibody-antigen pairs, the approach matters because structural ground truth often does not exist. - A sequence-first workflow can lower the cost of iterating on candidate ranking inside an active discovery program. - Dr. Lurong Pan said the linear probe result shows “the biology is already in the representation.”

What’s next: - Ainnocence plans task-adaptive fine-tuning of AINN-P1 on antibody-antigen affinity data. - The company also plans higher-capacity model variants and multi-objective heads that jointly optimize affinity, developability and specificity. - Closed-loop integration with experimental feedback is also planned. - Prospective wet-lab validation and broader benchmarking across additional targets are underway. - The capability is being integrated into SentinusAI®, Ainnocence’s biologics and antibody design platform.

The bottom line: - Ainnocence is positioning AINN-P1 as a practical sequence-only foundation model for antibody discovery, with early data suggesting it can rank binding affinity better than a bespoke model trained from scratch.