Overview of BIOPTIC B1/B2/B3 Scientific Modelling Layer presented at the Stanford Drug Discovery Symposium 2025.
What We Showed?
An AI Engine for Every Step of Early Drug Discovery. Bioptic’s platform is built around three core models:
1. B1: Ligand-Based Hit ID
Using only a known active compound as input, our ultra-high-throughput engine screens 40B+ molecules in under 2.5 minutes — with 100% recall.
- It’s powered by a RoBERTa-based transformer trained on 160M+ SMILES, optimized using techniques from search engines and recommender systems.
- In a recent case study with Oncobox, we discovered VEGFR-3 inhibitors for solid tumors — with the top molecule showing IC50 = 420 nM and 4x faster hit-to-lead conversion.
2. B2: Sequence-Based Hit ID
Don’t have a known binder? No problem. Our Flamingo-inspired architecture can go directly from a protein sequence to hit compounds.
- Screens massive libraries in under a day
- Finds chemically diverse and novel molecules (< 0.4 Tanimoto similarity to known binders), making it ideal for IP generation and first-in-class drug programs
3. B3: ADME Predictions with Gemini
Early pharmacokinetics (PK) and safety profiles are built in. Our ADME module:
- Won the ASAP Antiviral ADMET challenge
- Uses few-shot learning to adapt to sparse or heterogeneous data
- Predicts key ADMET properties even in unexplored chemical space
