AIKIUM
AI-DRIVEN NANOBODY DISCOVERY
Data Experience
Data Experience
Aiki
AI-driven nanobody generation that orchestrates Aikium's proprietary in-house models alongside every external tool cleared for commercial use. The recipe for wiring them together β target intake from UniProt or PDB, epitope specification in 3D, ensemble design, structural validation, and developability filtering β is what turns sequences into high-affinity, selective, developable binders.
One front-end. Ranked candidates exported as CSV, FASTA, and PDB. Designed to grow with each new model we add.
Note: ranked candidates are designs, not predicted binders β wet-lab validation remains required (typical hit rates β€5%). Aiki-Foundry runs that screening for you.
Aiki-Foundry
Our wet-lab approach has evolved. Instead of one trillion-scale library, we now deploy tens of focused, target-specific nanobody libraries in parallel β some drawn from an off-the-shelf collection of known binders and CDRs, others designed end-to-end by Aiki. Proprietary protocols compress every step: library design, assembly, Yotta-display screening, and high-throughput-ELISA validation.
Many targeted libraries beat one mega-library: deeper sampling of the relevant sequence space, shorter cycles, and parallel programs without competing for screening capacity.
Specialized in-house models, available standalone
Aiki-XP
Ranks a protein within its host's expression distribution by integrating genome context, operon architecture, codon usage, and biophysical features. Trained on 492K genes across 385 species; supports 1,831 host genomes. Tier D reaches Spearman 0.59.
Try Aiki-XP βAiki-GeNano
Three-stage language-model pipeline that generates ten ranked VHH candidates against linear peptides, IDRs, or soluble domains up to 244 aa. Returns developability scores, MSAs to the NBv1 scaffold, and ESMFold structures. +6.6 Β°C mean ΞTm over supervised baselines.
Try Aiki-GeNano βAiki-Sol
Predicts soluble vs. insoluble across five experimental regimes β four in-vivo centrifugation speeds (3Kβ100K Γg) plus cell-free β returning calibrated per-protocol probabilities. Strips common purification tags before scoring; available as web UI, REST, pip, FASTA batch, or Docker.
Try Aiki-Sol βThe three tools cover successive stages of a nanobody campaign β predicting expression in a chosen host, designing CDR variants against a target, and filtering candidates by developability before committing to wet-lab work.
The platform behind the models
Aiki and the specialized in-house models are trained on proprietary data from Yotta ML2, our mRNA-display platform engineered over two years to screen full-length protein scaffolds (VHH nanobodies, scFv). Yotta supports up to 1012 physical diversity per library β about 1,000Γ larger than phage display β and is specifically optimized for membrane proteins and intrinsically disordered regions. The platform now powers Aiki-Foundry β the focused-library wet-lab side of the offering.
For the comparison against phage display, screening timeline, and platform architecture, see the Yotta ML2 Technology section in the Technical tab.
Founded by scientists from Harvard Medical School, NVIDIA, and 10x Genomics.
Engage
Four ways to work with us, from low-commitment to deep. Terms are flexible β service program, partnership, or licensing β and we shape the engagement around the program, not around a SKU.
Try the live products
No commitmentAiki (end-to-end nanobody design) and the three standalone models β Aiki-XP, Aiki-GeNano, Aiki-Sol β run as hosted web apps with REST and batch endpoints. Aiki-Foundry's wet-lab campaign details are also on the landing tab. Fastest way to size up what we do before talking commercials.
Go to the live products βLicense data & pre-trained models
~2β4 weeks Β· IP-free for internal use90M+ sequenced binders across 90 epitopes and 50+ targets, with sequence, enrichment, and target metadata; plus the pre-trained weights behind Aiki-XP, Aiki-GeNano, and Aiki-Sol if you want to fine-tune in-house.
Discuss licensing βRun an Aiki-Foundry campaign
~8β16 weeks to validated hitsOne or more focused-library campaigns against your targets: AI-designed or off-the-shelf libraries, Yotta-display screening, HT-ELISA validation, end-to-end. Parallel campaigns across multiple targets are routine. Available as a service program or as a partnership β engagement terms negotiable, including IP structure.
Discuss a campaign βCo-develop a program
Strategic Β· milestone-based Β· shared IPFor programs with a clear path to IND: hit discovery through lead optimization, affinity maturation, and developability engineering, with shared IP and milestone-based structure. The deepest form of partnership we offer.
Discuss co-development βNote: our primary focus and validated data are protein scaffolds (VHH nanobodies, scFv). Yotta ML2 has also been tested with peptide libraries β if you're working on peptide therapeutics, ask.
Technical specifications for evaluating platform capabilities, library architectures, target coverage, experimental validation data, and AI model performance.
Yotta ML2 is a proprietary mRNA display variant optimized for screening protein scaffolds at 10ΒΉΒ² scale. The platform supports diverse scaffolds ranging from short peptides to full-length nanobodies (VHH, ~15 kDa) and engineered single-chain proteins (up to ~25 kDa).
Library assembly protocols accommodate extensive degeneracy across distant sequence regions. For example, SeqR-v3 libraries contain 46 degenerate positions distributed across 160 amino acids, enabling broad sequence space exploration. Custom library design services are available for partner-specified targets, with typical lead times of 8 weeks.
Yotta ML2 vs. Phage Display
- Β· ~1 billion variants
- Β· Challenging with membrane targets
- Β· 24+ week timelines
- Β· Generic libraries
- Β· Trillion-scale diversity
- Β· Optimized for membrane targets and IDRs
- Β· 8-week timelines to validated hits
- Β· AI-designed, target-specific libraries
Campaign timelines: 4 weeks from screening initiation to NGS-based hit identification using in-house libraries. Add 4 weeks for preliminary biophysical validation via ELISA.
Nanobody-v1
Nanobody-v2
SeqR-v3
Screening campaigns have been conducted against approximately 50 distinct protein targets, representing 90 unique epitopes. The portfolio includes GPCRs, single-pass and multipass membrane proteins, and proteins containing intrinsically disordered regions.
Target selection prioritizes membrane proteins and other historically challenging protein classes. Indications represented include oncology, immunology, neurology, and rare diseases.
The platform supports diverse antigen presentation modes that maintain native epitope structure and enable selection under physiologically relevantΒ conditions.
Screening campaigns generate pools of enriched binders that undergo staged validation. Initial enrichment is monitored via next-generation sequencing. Top-enriched candidates are evaluated by ELISA to confirm binding. Selected hits advance to biophysical characterization and functional assays as appropriate.
Observed rates: Mean of 9 validated binders per target across 20+ completed nanobody campaigns (range: 2-20+ per target).
Typical affinities: Validated hits show Kd values ranging from low nanomolar to sub-micromolar.
Validation workflow: Selection campaigns proceed through 6-12 rounds of alternating negative and positive selection. Enrichment is monitored by NGS. Top 100-500 candidates per target undergo ELISA screening. Confirmed binders advance to SPR or BLI for affinity determination, followed by functional assays when applicable.
Experimental data from screening campaigns has been used to train machine learning models for binding prediction and library design. External validation using published benchmarks indicates that Aikium data improves model performance on held-out test sets.
Testing methodology: A published nanobody-antigen predictor architecture was retrained using incremental additions of Aikium screening data. Performance was evaluated on the original authors' held-out test set without modification of model architecture or hyperparameters.
Reference: Model architecture from Deng et al., Nature Machine Intelligence (2024) . Evaluation performed on authors' test set.
Observation: Generalization error decreases systematically with training set size, suggesting transferable binding patterns across epitopes.
Aiki (end-to-end nanobody design), Aiki-Foundry (focused-library wet-lab campaigns), and the three standalone models (Aiki-XP for bacterial expression, Aiki-GeNano for nanobody design, Aiki-Sol for solubility) are all featured on the landing tab.
Go to landing tab βSelection Data & Analysis
Actual selection data from completed experimental screening campaigns β amino-acid enrichment patterns, binding distributions, and hit diversity across library-target combinations.
Identifiers are anonymized to protect IP. Starting from >10ΒΉΒ²-member libraries, each campaign typically yields over one million binders per target across nanobody (VHH, ~15 kDa) and proprietary SeqR (~25 kDa) scaffolds, spanning ~90 epitopes including GPCRs, membrane proteins, and IDRs. Experimentally validated binders are starred in the UMAP plots.
Note: interactive visualizations display a representative 0.1β1% sample (100kβ1M data points) for browser performance. Full datasets available under licensing.
Target Selection
Library
Binding Strength
Validation Status
Target Organism
Target Class
Target Category
Visualization Options
Size Options
Legend
Statistics
| Binder ID | Target ID | Library | Pseudo-Kd (nM) | Binding Strength | Validated |
|---|