Ainnocence’s foundation model trained from scratch, requiring no structure, no multiple sequence alignment, no fine-tuning, training in hours rather than weeks.
Ainnocence Inc.,an AI-driven pharmaceutical & material IP-generating platform company,today released research showing that representations learned by its AINN-P1protein foundation model substantially improve a core task in antibodyengineering: ranking candidate antibody-antigen pairs by binding affinity.Using AINN-P1 as a frozen encoder, the company improved the mean Spearman rankcorrelation for predicted change in binding free energy (ΔΔG) from 0.42 to 0.53,a relative gain of approximately 28% over a task-specific model trained fromscratch on the same data and evaluated under an identical five-foldcross-validation protocol.
The findingsare described in a new preprint, “AINN-P1: A CompactSequence-Only Protein Language Model Achieves Competitive Fitness Prediction onProteinGym,” by RogerWang, Kevin Jin, and Lurong Pan.
Keyfindings
● 28% relativeimprovement in ΔΔGranking quality (Spearman ρ: 0.417 → 0.533) under identical five-foldcross-validation.
● A simple linearmodel on frozen AINN-P1 embeddings (ρ = 0.457) already beats atask-specific network trained end-to-end from scratch (ρ = 0.417) directevidence that the gain comes from representation quality, not from added modelcapacity.
● Sequence-only. No co-crystal structure, docked model,or multiple sequence alignment is required at any stage.
● Orders-of-magnitudelower training cost: secondsper fold rather than hours, with the foundation model held frozen.
Why it matters
Antibodyaffinity maturation requires ranking large panels of candidate variants undertight experimental-label budgets. Building a bespoke predictor for each newtarget is data-hungry, slow and hard to reproduce across campaigns.
Many of thestrongest published affinity and fitness predictors depend on structural input,an experimentally solved or computationally docked three-dimensional model ofthe complex or on deep multiple sequence alignments. For novel antibody-antigenpairs, which is precisely the regime that matters in discovery, co-crystalstructures usually do not exist, and docked models introduce error at the exactinterface the prediction depends on.
Ainnocence’sresult is achieved without either. AINN-P1 encodes antibody and antigensequences directly, and a lightweight supervised head ranks candidates from theresulting embeddings. Because the foundation model is never fine-tuned, thesame embeddings can be reused across targets and objectives, and eachdownstream head trains in seconds enabling rapid, low-data iteration inside alive discovery campaign.
“The mostinformative result in this study is the one that looks least impressive onpaper,” said Dr. Lurong Pan, Founder and CEO of Ainnocence. “A plain linearprobe on our frozen embeddings outperforms a full network trained end-to-end onthe same labels. That tells you the biology is already in the representation.We are not winning by building a bigger model on top, we are winning becauseAINN-P1 has learned something real about how proteins bind, from sequencealone.”
How it was measured
The teamframed affinity maturation as a learning-to-rank problem: because downstreamdecisions depend on which candidates advance to the wet lab, the relativeordering of candidates matters more than absolute ΔΔG values. Spearman rankcorrelation served as the primary metric, with NDCG and AUC tracked assecondary measures.
Threeconfigurations were evaluated on a curated antibody-antigen ΔΔG dataset. Allshared identical inputs, labels, and cross-validation folds, and featurenormalization were fit only on training folds to prevent information leakageisolating the effect of the representation itself.
Meanfive-fold cross-validated Spearman ρ for antibody-antigen ΔΔG ranking. Allconfigurations use identical data and folds. Source: Wanget al., 2026
“Becausewe keep the foundation model frozen, these numbers are a floor, not a ceiling,”said Roger Wang, co-author of the study. “We have not yet spent a singlegradient step specializing AINN-P1 to antibody-antigen data. That is the next experiment,and we expect the margin to widen.”
What comes next
Because thefoundation model is held frozen throughout, the reported accuracy represents aconservative lower bound. Ainnocence plans a task-adaptive fine-tuning stagethat specializes AINN-P1 to antibody-antigen affinity data, alongsidehigher-capacity model variants, multi-objective heads that optimize affinityjointly with developability and specificity and closed-loop integration ofexperimental feedback. Prospective wet-lab validation and broader benchmarkingacross additional targets are underway.
Thecapability is being integrated into SentinusAI®, Ainnocence’s biologics andantibody design platform.
Representativeoutcomes from recent antibody discovery campaigns run on SentinusAI® aresummarized in Table 1, spanning cytokines, receptors, immune checkpoints, andbispecific formats.
RepresentativeHigh-Yield Antibody Discovery Campaigns on SentinusAI®
Target
Hits/Tested
Hit Rate (%)
Best KD
TSLP
13/9
69.2
1pM
IL36R
20/11
55
1pM
GD2 X CD3
44 / 85
51.8
129 pM
IL15
34 / 74
45.9
392 pM
CD19
50 / 124
40.3
28 pM
Activin A X GDF8
46 / 116
39.7
18 pM
TSHR
41 / 121
33.9
N/A
FcRn
33 / 98
33.7
49 pM
C1q
26 / 84
31
540 pM
Availability
The preprintis available at AINN-P1:A Compact Sequence-Only Protein Language Model Achieves Competitive FitnessPrediction on ProteinGym.The AINN-P1 model and the curated antibody-antigen ΔΔG dataset are proprietaryto Ainnocence Inc. aggregate results are reported in full in the preprint.Inquiries regarding data or model access may be directed to the correspondingauthor.
AboutAinnocence, Inc.
Founded in2021 and headquartered in California, Ainnocence is a next-generationbiotechnology company transforming drug discovery and synthetic biology throughAI-based, sequence-first engineering. The company’s self-evolving platformevaluates up to 10 billion molecules spanning proteins, antibodies, smallmolecules, nucleic acids, and chemical formulations within hours to weeks,enabling rapid, multi-objective design across therapeutic, biological, andchemical systems. By reducing R&D timelines and costs while increasingsuccess rates, Ainnocence empowers industry and academic partners to pursuecomplex biological innovation with greater precision and control.
MediaContact
Dr. Lurong Pan,PhD, Founder and CEO
+1205-249-7424
Visitus on social media: LinkedIn | YouTube | X
Forward-lookingstatements. Thisrelease contains forward-looking statements regarding anticipated modelperformance, planned research directions, and platform development. Resultsdescribed reflect retrospective computational evaluation on a curated internaldataset and have not been prospectively validated in the laboratory. Actualresults may differ materially. This release does not constitute an offer to sell,or a solicitation of an offer to buy any securities.