forestgump

Life is like a box of baselines.

Morgan-fingerprint baselines for cheminformatics models.

You trained a molecular property model for 3 days on 8×H100. Did it beat a random forest trained on Morgan fingerprints?

forestgump takes your molecules as SMILES, plus your labels, your cross-validation folds and your model's out-of-fold predictions. It fits three baselines on exactly the same folds: the training mean, Tanimoto kNN on Morgan fingerprints, and a random forest on Morgan fingerprints. Name a context column, such as an assay, a chemical series or a reaction's catalyst, and it adds a fourth: the context mean. Then it runs a paired test to decide whether your model beat the forest.

Your model, trained on 8×H100 for 3 days, lost to a random forest. The forest trained in 31 seconds on a CPU.
The forest has no GPU. The forest has no feelings. The forest won.

Install

pip install forestgump

Try it

forestgump demo                             # a model that predicts noise. The forest wins.
forestgump demo --model oracle              # knows the answer, beats the forest on all 5 folds, and only ties
forestgump demo --model oracle --repeats 3  # 15 folds is enough evidence. You win.

The demo molecules are 105 para-disubstituted benzenes. Their labels are RDKit's Crippen logP.

The tie is not a bug. Five folds is not much evidence, and the test is built not to call a winner without it.

Use it

You need one CSV with one row per molecule. Its columns are:

forestgump run my_results.csv
forestgump run my_results.csv --smiles SMILES --target y --fold cv_fold --pred my_model

For repeated cross-validation, add one row per (molecule, repeat) and pass --repeat repeat_column. The test set of (repeat r, fold f) is those rows. Its training set is every other row of repeat r.

To add the context mean, name one or more context columns. With several, a group is one combination of their values.

forestgump run my_results.csv --context catalyst base

From Python:

import forestgump

report = forestgump.run(df, smiles="smiles", target="target", fold="fold", pred="pred", context=["catalyst", "base"])
print(report.verdict)  # "win", "lose", "tie", or "refuse"
print(report.render())
report.to_json("forest.json")

The baselines

Each baseline is retrained on every training set, and its hyperparameters are picked from that training set only. Test rows are only ever predicted.

• The Mean has nothing to tune. It predicts the training-set mean for every test row, or for a two-class target the training share of the larger class.
• Retrieval-Augmented Mean appears only with --context, and it has nothing to tune either. For each test row it retrieves the training rows with the same context values and averages their labels. A group with no training rows gets the training-set mean, and the report says how many test rows that happened to.
• Large Lookup Model uses inner cross-validation, which takes milliseconds for kNN: 5 folds, or 3 below 50 training rows.
• The Forest fits each of its 12 configs once and keeps the one with the best out-of-bag score. Every tree skips about a third of the training rows, so the forest gets a validation score for free. If you'd rather tune it like everything else, --forest-inner-cv uses inner cross-validation instead. That fits each config once per inner fold and then refits the winner, so it is about 5× slower.

Both fingerprint baselines use 2048-bit Morgan fingerprints from RDKit.

How the verdict is decided

The primary metric is RMSE for numeric targets and ROC-AUC for two-class targets. The report also gives MAE, R² and Spearman ρ, or accuracy.

1. Every method is scored on every test set. The full per-test-set table is printed before any verdict.
2. Your model is compared with each baseline on the per-test-set differences, using the Nadeau–Bengio corrected resampled t-test. CV training sets overlap, so the differences are correlated and a plain paired t-test is overconfident. The correction inflates the variance by 1/J + n_test/n_train over J test sets (Nadeau & Bengio 2003; Bouckaert & Frank 2004 for repeated CV).
3. The p-values, three or four with --context, are Holm-adjusted together.
4. The verdict compares you with the forest at α = 0.05: win, lose, or tie. With fewer than 3 test sets the forest declines to be judged.

The report also gives the mean difference, its confidence interval and Cohen's d_z. Use --json to get every number, the hyperparameters chosen on each test set, and the baselines' predictions.

Flags

License

Apache-2.0.