QSAR Models trains a panel of classical models on molecular fingerprints and descriptors, compares them under one split by cross-validation, and saves the best one as a model you can predict with. You get the winner and an honest comparison: how every algorithm did in cross-validation and on a held-out test set. At the dataset sizes most projects have (tens to a few thousand compounds), these models are often as accurate as graph neural networks and train in seconds.
The panel
| Algorithm | What it is |
|---|---|
| Ridge / logistic regression | A linear model on standardized features |
| Random forest | An ensemble of decision trees |
| LightGBM | Gradient-boosted trees |
| SVM | A support vector machine with an RBF kernel |
| kNN (Tanimoto) | A similarity-weighted average of the nearest training molecules |
| Gaussian process (Tanimoto) | A probabilistic model with calibrated uncertainty (regression, up to 5,000 molecules) |
Untick any you do not need. Each algorithm tries a small fixed set of settings, so run time stays predictable.
How the best model is chosen
- A test set is held out first (
random,scaffoldorclustersplit, or your owntrain/val/testlabels). Scaffold and cluster splits keep related molecules together, so the test set measures prediction on new chemotypes. - Each algorithm is scored by k-fold cross-validation on the rest (5 folds by default; scaffold and cluster groups stay within one fold).
- The algorithm with the best cross-validated metric (RMSE by default for regression, ROC AUC for classification) is refitted on all non-test molecules and saved.
- Every algorithm is also refitted once and scored on the test set, alongside mean and nearest-neighbour baselines, so the comparison shows whether the winner really learned something.
Features
Morgan count fingerprints (radius and size are configurable), RDKit 2D descriptors, MACCS keys and PaDEL descriptors, alone or combined. The kNN and Gaussian process models always use Tanimoto similarity on Morgan counts.
Predictions
Each prediction comes with:
- a 90% prediction interval (regression; split-conformal, from cross-validation errors), and a standard deviation from the Gaussian process, random forest or kNN models;
- the similarity of the molecule to its nearest training molecule, so you can tell when a prediction is outside the model's experience;
- signed atom attributions for linear, random forest and LightGBM models on Morgan features: atoms that raise the prediction and atoms that lower it, from SHAP values.
Binary models predict the probability of the positive class. Train a binary model from 0 / 1 labels, or from values with a threshold (for example p_activity ≥ 6).
Inputs
SMILES with one value each, or a table such as a Bioactivity Dataset Curator curated table, including its split column. Up to 10,000 molecules.
Outputs
| File | Contents |
|---|---|
comparison.csv | Every algorithm's metric on every fold and on the test set |
test_predictions.csv | Observed values and each algorithm's test-set predictions (for parity and ROC plots) |
feature_importance.csv | The saved model's most influential features |
model/ | The saved model, with its model card (no pickled code) |
The saved model appears in My Models and can score whole libraries in a workflow with Saved Model Scoring.
Related tools
Curate the data first with Bioactivity Dataset Curator, or train a message-passing network on the same data with MPNN Models.