Reactions as signed ECFP differences give a fast prefilter for template applicability and reaction classification
- Reactions are represented as the product-minus-substrate difference of counted ECFPs; for ECFP-compatible templates, applying the reaction shifts the fingerprint by a constant, mapping graph transformations to translations in fingerprint space.
- Reaction-center ECFPs encode the local environment a template needs, giving a fast coordinate-wise prefilter for whether a template can be applied.
- On USPTO-50k and MetaNetX reactions across ECFP radii, both representations carry information for reaction-class and enzyme-class classification, and reaction-center ECFPs add information beyond reaction ECFPs.
💡 Points Based on the abstract
The practical value is the cheap applicability prefilter: template-based planners and pathway-search engines spend much of their time on templates that cannot match, and this prunes them before subgraph matching. USPTO-50k classification is a saturated benchmark, so I'd read little into it. Drop it into an existing template-based planner and measure the false-reject rate, because a prefilter that silently discards valid templates costs routes.