<p>In the field of computer-aided synthesis planning, retrosynthetic pathway planning stands as a core issue. It aims to leverage data-driven strategies to explore synthetic routes that are both efficient and feasible for the target molecule. Current retrosynthesis methods rely on two core components: (a) an effective search methods, such as Monte Carol tree search, heuristic A* search, depth-first search, or breadth-first search; and (b) a useful evaluation methods. However, the existing evaluation methods cannot accurately depict the synthetic cost of molecules. In our study, we introduce a novel strategy for the evaluation of intermediate molecules in multi-step synthesis, namely similarity estimate. This strategy gauges the synthesis complexity of molecules by estimating their prevalence. Experimental results on classical datasets substantiate that our approach can effectively reduce the complexity of path search and augment the interpretability of multi-step retrosynthesis algorithms.</p> Graphical abstract