This workflow supports turning a trained compound-property prediction model into interpretable domain knowledge about which structural fragments drive a target property (e.g., potency against a biological target). Starting from a set of compounds represented by structural fingerprints and a measured property value, a compact subset of informative binary fragment features is selected, and a regression model (e.g., a random forest) is trained on a training split of the compounds. The trained model is converted into an explicit set of decision rules (root-to-leaf paths), each specifying a combination of fragment-presence conditions and a predicted property value.
Predicted values are discretized into ordered intervals (e.g., low/medium/high), and an overview heatmap shows, per interval, how frequently each fragment feature and value (present/absent) occurs among rule conditions.
Analysts iteratively filter this rule set by feature conditions to test hypotheses about individual and combined fragment effects, revealing candidate synergistic combinations.
To move beyond pairwise inspection, rules are encoded as "documents" over feature-value "terms," and Non-negative Matrix Factorization (NMF) topic modeling is applied, with the number of topics selected through complementary criteria — coverage of a shared 2D topic embedding, the strength of each rule's dominant-topic association, and how well candidate topic counts separate rules of the target class from the rest — followed by selecting a topic count balancing these criteria. The resulting topic-term matrix and a topic-weight parallel-coordinates plot (colored by property class) reveal a topic strongly associated with the target class.
Projecting rules by their topic-weight vectors into 2D and spatially selecting the resulting target-class cluster isolates a focused rule subset, which is then de-duplicated (removing rules subsumed by more general ones) and re-analyzed with a smaller topic model to pinpoint the specific conditions distinguishing target-class from other rules within the cluster.
Guided by these topic-defining conditions, analysts iteratively add discriminating feature conditions to isolate specific combinations that consistently yield target-class predictions.
These combinations are mapped, via the known fingerprint-bit-to-substructure encoding, to concrete molecular fragments, visualized as node-link diagrams, and validated by screening held-out test compounds. Finally, the matched fragment combinations are highlighted directly on example compound structures, and the overall findings are synthesized into testable structure–property hypotheses about which fragment constellations mechanistically contribute to the target property.
The workflow is illustrated here on a case study of compounds acting on a biological target, predicting potency from structural fragment features, but is intended as a reusable template applicable to other compound classes, targets, properties, and model types convertible into rules.