Active Learning

Recommends which compounds to make next from what you have measured and a pool of candidates. A surrogate model predicts each candidate's value and uncertainty, an acquisition function balances exploiting good predictions against exploring uncertain ones, and a diversity step spreads the batch across chemical space. For combinatorial libraries too large to enumerate, use Thompson Sampling instead.
0 measured compounds · 0 candidates
Estimated minimum cost: $—

Examples

Surrogate model

The surrogate predicts each candidate's value and how uncertain that prediction is. A Gaussian process with a Tanimoto kernel on Morgan count fingerprints is fitted to your measurements each round. A saved MPNN or QSAR model is used as trained; acquisitions other than greedy need one that reports uncertainty (an ensemble, an uncertainty head, a random forest, kNN or Gaussian process).

Measurements

Compounds you have already made and measured, with the value you are optimizing. Repeated compounds are averaged and never suggested again. Paste SMILES and values one per line, or upload a CSV / TSV such as a curated bioactivity table. At most 5,000 compounds.

or drop a file here

Candidates

The pool to choose from, such as an enumerated library, Free-Wilson proposals or matched molecular pair analogs. Paste SMILES one per line, or upload a CSV / TSV whose other columns (ids, docking scores) are kept next to each suggestion. At most 10,000 candidates.

or drop a file here

Strategy

The acquisition function trades predicted value against uncertainty: greedy exploits the prediction alone, UCB adds a multiple of the uncertainty, EI and PI score the expected gain or the chance of beating your best measurement, and Thompson sampling ranks by a draw from the surrogate. Diversity keeps the batch from crowding into one chemical series.