🧪 Chemical Reaction Predictor
Predict product SMILES from reactant SMILES — GRU · Transformer · Ensemble — with molecular visualization, confidence scores, Tanimoto similarity, reaction type classification, atom mapping, retrosynthesis, 3D viewer, and more.
Enter SMILES strings. Supplying the actual product enables match metrics.
Browse predictions on the test set (50 samples).
Run over a batch from the built-in test set. Results include per-reaction-type breakdown and Tanimoto.
Upload a CSV with an input (reactant SMILES) column and optionally an output (actual product) column. Predictions run on every row.
Run the same input through all three models simultaneously and compare their predictions side by side.
All predictions made this session. ⭐ = starred via Manual input tab.
Standalone molecule drawing, property analysis, and validation.
Tanimoto similarity between two molecules
Atom Mapping
Visualise which atoms in the reactant correspond to atoms in the product using Maximum Common Substructure (MCS) matching. Matching atoms are shown in the same colour.
💡 Tip: After running a prediction in the Manual input tab, paste the predicted product SMILES here to see atom mapping.
Retrosynthetic Analysis
Enter a target molecule (product SMILES) to find possible precursor pairs using SMARTS-based disconnection rules.
Supports: ester hydrolysis, amide bond, Williamson ether, imine, Heck, Suzuki, retro-hydration.
⚠️ These are heuristic SMARTS disconnections for common reaction classes, not deep-learning retrosynthesis. For advanced retrosynthesis, consider tools like AiZynthFinder or ASKCOS.
3D Conformer Viewer
Generate and visualise a 3D conformer of any molecule using RDKit ETKDGv3 embedding + UFF force-field minimisation, displayed interactively via 3Dmol.js.
Drag to rotate · Scroll to zoom · Right-drag to pan
💡 Style buttons: Stick (default) · Sphere (space-filling) · Wire (wireframe)
Reagent & Solvent Browser
Search common reagents, solvents, and coupling agents by name or SMILES fragment. Copy the SMILES string into the Manual input tab to use as a reactant.
Confidence Calibration Plot
After running a Batch evaluate, click below to see how well the model's structural similarity (Tanimoto, used as a confidence proxy) correlates with character-level accuracy.
A well-calibrated model's points should cluster near the diagonal.
- Green dots = individual predictions
- Orange dots = decile bin averages
- Blue dashed = perfect calibration line
Note: Tanimoto similarity is used as a proxy for confidence since raw model log-probabilities are not exposed through the current predictor API.
Error Analysis
After running a Batch evaluate, filter results to show only failures (non-exact-matches), sorted by Levenshtein error (worst first).
Use the threshold slider to focus on the most severe errors.
💡 LevErr = 100% − Levenshtein accuracy. Higher = more different from actual. Tan = Tanimoto similarity (structural closeness).