Fig. 2

A scenario where a reaction from an observed pathway is not described by any transformation rule used for training. The observed pathway has compound A transformed to B then to C, with the reaction from \(B\rightarrow C\) described by a transformation rule ( \(r^{B\rightarrow C}\)) but none for \(A\rightarrow B\) ( \(r^{A\rightarrow B}\)). The Single-Generation evaluation approach would only evaluate \(B\rightarrow C\) (with a reward +) and ignore \(A\rightarrow B\), since no rule can be applied. The Multi-Generation evaluation approach would penalize both compounds B and C (−) for not being predicted