Active Learning answers "which of these candidates should we make next?". You give it the compounds you have measured and a pool of candidates, such as an enumerated library, Free-Wilson proposals or matched-pair analogs. A surrogate model predicts each candidate's value and how uncertain that prediction is. An acquisition function turns the two into a score, and a diversity step spreads the batch across chemical space. Each design–make–test–learn round is a new run: add the new measurements and run it again.
The surrogate
- Gaussian process (default). Fitted to your measurements every round, with a Tanimoto (MinMax) kernel on Morgan count fingerprints. It reports both a prediction and a standard deviation, and the round summary includes its cross-validated accuracy on your data.
- A saved model. An MPNN or QSAR model from MPNN Models or QSAR Models is used as trained. Acquisitions other than greedy need a model that reports uncertainty: an ensemble, an uncertainty head, a random forest, kNN or Gaussian process.
Acquisition
| Acquisition | Picks |
|---|---|
| Expected improvement (EI) | The largest expected gain over your best measurement |
| Upper confidence bound (UCB) | The best prediction (closest to a target range) plus a multiple of its uncertainty |
| Probability of improvement (PI) | The best chance of beating your best measurement, or of landing in a target range |
| Thompson sampling | The best values in a draw from the surrogate |
| Greedy | The highest prediction, ignoring uncertainty |
You can maximize, minimize, or aim for a target range. Compounds you have already measured are never suggested again, matched by InChIKey, even when their value is censored and left out of the fit.
Diversity
- MaxMin: among the best-scoring candidates, start from the best and keep adding the one farthest from every pick so far.
- Cluster: cluster the best-scoring candidates, then take each cluster's best in turn.
- Kriging believer (Gaussian process only): take the best, pretend it has been measured at its predicted value, rescore, repeat.
- None: the top scores.
Outputs
A ranked table of suggestions with each pick's predicted value and uncertainty, acquisition score, nearest measured compound and cluster. You also get the picks as SMILES for docking or other workflow steps, every candidate's score, and a round summary with the surrogate's accuracy and your best value so far.
Limits
- One objective per run. Apply constraints such as ADMET thresholds with a filter step before the candidates reach this tool.
- The Gaussian process uses at most a few thousand measurements, and the candidate pool is capped. The exact limits are in the reference below.
- Suggestions are a ranking from a model, not measurements.
Related tools
- Thompson Sampling searches combinatorial reagent spaces too large to enumerate, with a fixed evaluator. Use Active Learning when you already have the candidate list and want the model to learn from each round's results.
- Free Wilson Analysis proposes R-group combinations you have not made. The ready-made workflow Propose R-group combinations with Free-Wilson, then pick the next batch to make chains it into Active Learning.