0.81
Leave-program-out AUC
AINN-P1 on VHH affinity maturation, vs 0.68 for finetuned ESM2, parity-plus at 3.9× fewer parameters.
Ainnocence designs binders from sequence alone — no crystal structure, no docking, no predicted-fold error to inherit. The same property lets us do what structure-based pipelines cannot: score every candidate against the surface and circulating human proteome for antibodies, and the full proteome for small molecules, before a single molecule is made. Specificity becomes a design constraint, not a late-stage surprise.
Validated targets with sub-nM leads
Single-digit picomolar campaigns
Molecules experimentally tested in wet lab
Parameter protein foundation model (AINN-P1)
Most AI drug-discovery platforms stop at a demo. We go further: our models are trained on industry-grade proprietary data and validated at every stage by wet-lab experiments, so what we deliver to pharma partners is already de-risked.
The result is a different kind of handoff, not a screening score, but a program: patentable, validated, licensable, and ready for your development pipeline.
Built on AINN-P1, SentinusAI combines sequence modeling, affinity prediction, and generative design into a single pipeline. Trained on proprietary assay data and continuously benchmarked against wet-lab outcomes.
IgG, Fab, scFv, VHH, bispecifics. Picomolar leads on IL-13, IL-25, IL-36R, TSLP.
De novo small molecules, PROTAC design, scaffold hopping. CRBN neosubstrate selectivity panels.
GLP-1 analogs, cyclic peptides, stapled therapeutics. 3.2M-peptide screens with 90%+ accuracy.
CAR construct optimization, activation prediction, persistence modeling.
Most AI protein design begins with a 3D structure. That dependency sets a ceiling: targets without solvable structures are out of reach, and every predicted fold carries error that propagates silently into every downstream decision.
Our AINN-P1 foundation model works from sequence. This removes the ceiling — multi-pass membrane proteins, GPCRs, and novel antigens with no structural data become tractable — and it makes proteome-scale screening computationally possible. For antibodies, we evaluate candidates against the surface and circulating human proteome during design, matched to where antibodies actually encounter their off-targets. For small molecules, CarbonAI screens against the full human proteome, where intracellular off-targets drive most late-stage attrition. Either way, off-target liabilities surface in silico rather than months later in a cell-based assay. Predictions close the loop against our wet-lab binding platforms, so every campaign returns validation data to the model.
The result is a design cycle where affinity and specificity are optimized together, from the first iteration.
AINN-P1 on VHH affinity maturation, vs 0.68 for finetuned ESM2, parity-plus at 3.9× fewer parameters.
Methods, benchmarks, and validation studies published across antibody, protein-language-model, and chemistry venues.
Founders bring twenty-plus years of structure-free affinity modeling and protein engineering to the company founded in 2021, not a ChatGPT-era pivot.
A selection of in-house antibody programs with wet-lab-confirmed picomolar binding. Full data room available under CDA.
| Target | Modality | Best Kd | Stage |
|---|---|---|---|
| IL-13 · atopic dermatitis, asthma | VHH / scFv | ~1 pM | Lead optimization |
| IL-25 · atopic & T2 inflammation | IgG | ~1 pM | Lead optimization |
| IL-36R · generalized pustular psoriasis | IgG | ~1 pM | Lead optimization |
| TSLP · severe asthma | IgG / Fab | ~1 pM | Lead optimization |
| IL-15 × IL-21 bispecific · oncology | Bispecific IgG | single-digit pM | Affinity maturation |
| TNF-α × OX40L bispecific · autoimmune | Bispecific IgG | single-digit pM | Affinity maturation |
| TROP2 · oncology ADC payload target | IgG / Fab | sub-nM | Wet-lab validated |
| B7-H3 · solid tumors | IgG | sub-nM | Wet-lab validated |
All affinities measured by SPR or BLI from AINN-P1-designed campaigns. Figures from Antibody Model Experimental Validation Statistics (Sep 2026), 40-target dataset, available under CDA.
We work with partners whose capabilities complete what a sequence-first model cannot do alone: human biological validation on one side, clinical-outcome prediction on the other. These are the two active, publicly announced collaborations today.
Partnership to combine Ainnocence AI-driven discovery with Obatala's human adipose and metabolic tissue platforms, so that de-novo molecules can be tested in physiologically relevant human biology before any animal study.
Read announcement →Industry-first collaboration integrating Ainnocence's discovery platform with Phase Advance's preclinical and clinical trial modeling, so that drug performance can be predicted years earlier than traditional development allows.
Read announcement →Decades of combined experience in protein engineering, cheminformatics, machine learning, and pharmaceutical commercialization.
Our data room is open under CDA. Pick a target, we'll share wet-lab data and freedom-to-operate analysis within one week.
Request the data room →