Every Ainnocence platform traces back to one architectural decision: treat molecules as sequences, learn from the data you actually have, and let 3D structure be an output, not an input.
Our platforms are not a product of the 2023 LLM wave. The architectural choices that make them work were made years earlier, on different problems, at different scales, and compound every time we train the next model.
Our founding architectural insight. CDNN modeled information flow across biological signaling systems using recurrent state-passing rather than attention. The core intuition, that long-range biological dependencies are better captured by sequential dynamics than all-to-all attention, reappears in every model we've built since.
CDNN applied to SMILES, selecting the first class of chemical descriptors that scaled to billions of compounds without 3D docking. Became the foundation for what would later become CarbonAI.
A simple, fast, efficient GNN framework for virtual screening, published and benchmarked externally. Replaced docking as the primary filter in CarbonAI pipelines, enabling 1010-compound overnight screens.
First production platform. Combined CDNN-style sequence encoders with downstream heads for antibody binding, immunogenicity, and expression. Launched with 1M+ antibody variants evaluated across early validation campaigns.
Our current flagship. 167M parameters, mLSTM backbone, no MSA/structure dependency. Leads ProteinGym on stability. On leave-program-out VHH expression, matches finetuned ESM2-650M (0.810 vs 0.816 AUC) at 3.9× fewer parameters. Powers all antibody programs in the pipeline.
Specialized engines across small molecule, peptide, cell therapy, RNA, and natural product. Each builds on the AINN-P1 foundation and its predecessors. All five are production-grade; all five publish benchmarks that make their own numbers harder to beat.
Not every architecture wins on every problem. For protein and chemistry sequences, the economics favor recurrent state passing, not because transformers are bad, but because biology has structure that recurrence captures natively.
Short sequences (10–200 tokens), highly local dependencies with occasional long-range ring closures. Transformer attention is often wasted here. Our chemistry stack uses specialized GNN encoders (ParaVS) and recurrent variants where sequence information dominates.
Long sequences (100–5000 tokens), hierarchical structure, strong local motifs. Transformer attention is O(n²) in memory and compute, expensive to serve, slow to train. mLSTM state passing is O(n) with a bounded memory footprint, and empirically preserves the long-range biophysics that matter for binding and stability.
Both branches share one principle: sequence first. We ask the structure to emerge from the model's representation, not the other way around. That is what makes it possible to run production-grade discovery on targets with no crystal structure, no cryo-EM, no AlphaFold model that resolves the binding interface, which, in antibody engineering, is most of them.
We publish benchmarks that make our own numbers harder to beat. The AINN-P1 preprint is public; the raw per-target wet-lab datasets are available to evaluation partners under CDA.
Pilot engagements scope in a week. We'll run a representative campaign and share the raw results, including the ones that didn't work.
Scope a pilot →