Analogical methods for molecular generation and property prediction.
Analogical generation
Mines transformation rules from matched molecular pairs across chemical series to generate analogues.
Counterfactual prediction
Causal-Chemprop augments an existing Chemprop model with counterfactual reasoning on measured anchors.
1: Mine the rules.
2: Apply the rules.
3: Score the candidates.
Causal-Chemprop retrieves analogues of the candidate to make a counterfactual property prediction.
measured truth
smooth model
activity cliff
counterfactual leap
structure →
4: Evaluate the uncertainty.
Variation across exemplars allows uncertainty quantification, guiding an acquisition function to balance exploration and exploitation of rules and compounds.
Two new methods.
Analogical generation turns the transformations in your library into new molecules.
Counterfactual prediction grounds every estimate to a measured anchor, and applies to any Chemprop encoder you bring.
Neopoly Ltd · 2026