A chemistry engine that predicts reactions no one has run yet
Chem0 started as a research idea. We joined as their product, design, and AI team for two years. What if a researcher could see a reaction's risk, stability, and outcome before touching a single compound?
- Client
Chem0- Focus
- Product Strategy, Brand, UX/UI Design, Engineering, Machine Learning, Infrastructure
- Year
- 2024–2025
- Website
- chem0.com ↗
The brief
Give chemists a way to test a reaction before they run it, in one platform that works for both labs and pharma.
Chem0 came to us with a vision, not a product: modernise chemistry research by putting molecular intelligence, machine learning, and visualisation in one place.
They needed a long‑term partner to build it from scratch, from the brand and the interface to the AI models and the cloud they run on, for pharma companies, R&D labs, and universities alike.
The models
Instead of one model that guesses everything, we built three engines, each trained for a single scientific question.
The hardest part was trust. Chemists won't act on a black box, so every engine shows its inputs, from pH and humidity to functional groups and pKa, and returns a score a researcher can reason about.
Underneath, Graph Neural Networks learn how structure, physicochemical properties, and conditions interact. That lets Chem0 predict chemical interactions it has never seen, where rule‑based systems stop.
Chemical Stability Engine
Scores how stable a combination stays across pH, temperature, and humidity. About 92% accurate.
Reaction Risk Engine
Rates mixing compounds as low, medium, or high risk. Trained on 50,000+ combinations, up to 98% accurate.
Reaction Classification Engine
Predicts the most likely reaction type, including no reaction at all. About 94% accurate.
Molecular intelligence engine
The GNN core: descriptors, LogP, TPSA, pKa, and hydrogen bonding, feeding every model.
Research Mode
Can two drugs be formulated together safely? Chem0 answers in about 90 seconds.
A scientist drags two active ingredients onto the reaction canvas, sets real conditions such as 25°C, pH 7, and 60% humidity, and runs Research.
All three models run at once: a stability score from 0 to 100, a low, medium, or high risk for every functional‑group pair, and the likely reaction types, from salt formation to nucleophilic attack.
LabMate, Chem0's AI assistant, watches the experiment. Click any risky pair and it explains, in plain terms for a formulator, why it matters under those conditions.
81.26
Stability score for Avatrombopag + Azithromycin in the demo
86%
Model confidence for that result
1
High-risk functional-group pair flagged: ketone and ketone
The product
Around the models, a full product: the brand, the marketing site, and a research dashboard built for chemists.
Under the hood, a gateway on AWS Lambda runs the app's API and LabMate, powered by Gemini, while a separate service serves the Graph Neural Network models.
- Prediction of previously unseen chemical interactions
- AI chemistry assistant for scientific exploration
- Interactive 2D and 3D molecular visualisation
- Batch prediction APIs for large-scale workflows
- Multi-model inference architecture
- Enterprise-ready APIs with interactive documentation
- Research journeys designed for labs and institutions
- Complete brand, marketing site, and dashboard
Result
Chem0 grew from a research idea into an investor‑ready platform, earned a Letter of Intent from Strausswell Organics after beta, and drew acquisition offers while still in development.
- peak prediction accuracy
- 94.84%peak prediction accuracy
- research and industry institutions live
- 3research and industry institutions live
- validated chemical combinations
- 50,000+validated chemical combinations
- specialists on the dedicated team
- 7specialists on the dedicated team
“We came to Keizerworks with a research idea and left with a platform chemists actually trust. They built the models, the product, and the brand as one team, and we had a Letter of Intent before we'd even finished building.”