SyntheMol
Monte Carlo Tree Search over 132,000 molecular building blocks and 13 validated chemical reactions, guided by an antibacterial activity predictor and wet-lab assays.
The loop
- Proposes
Tree expansion combines REAL-space molecular building blocks under validated chemical reaction templates.
- Judged by
Reaction compatibility rules verify synthetic tractability at every node, while a Chemprop GNN/MLP oracle scores antibacterial activity against Acinetobacter baumannii.
- Steers
MCTS upper-confidence rollouts focus combinatorial search on high-scoring, Tanimoto-diverse reaction pathways before wet-lab synthesis and MIC testing.
Headline result
Searched a combinatorial space of 30 billion synthesizable molecules to discover 6 structurally novel compounds with potent wet-lab antibacterial activity against A. baumannii.
What this does not prove
Search space is restricted to 1-2 step combinatorial reactions from the Enamine REAL building-block catalog.
How it generalizes the harness
- The algorithm is the solver
- Inverted control
- A domain authority judges
- The output outlives the model
- Search over a population
- The generator is not an LLM
- Multi-fidelity screening ladder
Monte Carlo Tree Search over formal reaction templates calls a GNN property predictor and validates top candidates in wet-lab bacterial growth assays.
Classification
| What it proposes | Molecule |
|---|---|
| Generative model | Chemprop GNN value oracle + reaction-template MCTS |
| Algorithmic authority | SMARTS chemical reaction validator, MCTS, and wet-lab MIC assay |
| Feedback | Score |
| What gets updated | Search tree |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |