Predict aqueous solubility of organic molecules using Random Forest and Neural Network models trained on the ESOL dataset. Enter a SMILES string or try one of the examples.
Browse recent predictions, see which model versions are active, and inspect how the latest outputs are distributed.
Solubility bands across the current history page.
Loading prediction distribution...
Recent molecules, timestamps, predictions, and model version tags.
Loading history...
Loading results...
How SolPredict works, from data to predictions.
The ESOL dataset (Delaney, 2004) contains experimentally measured aqueous solubility for 1,128 small organic molecules. Target variable is log(S) in mol/L. We use an 80/20 train/test split with random seed 42.
Each molecule is converted to a 2048-bit Morgan fingerprint (ECFP4, radius=2) using RDKit. Each bit encodes the presence of a circular substructure, producing a fixed-length binary vector that standard ML algorithms can consume directly.
Ensemble of 100 decision trees via bootstrap aggregation. Averages predictions to reduce variance.
Feed-forward MLP (2048 → 512 → 128 → 1) with ReLU and dropout (p=0.2). Trained with Adam, MSE loss, 100 epochs.
Models are evaluated with R\u00B2 (variance explained, higher is better), RMSE (root mean squared error, penalizes outliers), and MAE (mean absolute error, robust to outliers). All error metrics are in log mol/L units.