Despite remarkable advances in computational chemistry, prediction of chemical reactions is still challenging,
because investigating possible reaction pathways is computationally prohibitive due to the high complexity of
chemical space. For instance, their brute-force sampling is too demanding because of their large degrees of
freedom. A stochastic sampling method inherently requires many trials no matter how effective it is, because it
cannot guarantee 100% probability of finding a designated target structure within a finite number of samplings.
A feasible strategy for efficient prediction is to utilize chemical heuristics and machine learning techniques. We
proposed a novel approach to search reaction paths in a fully automated fashion by combining chemical theory
and heuristics. A key idea of our method is to extract a minimal reaction network composed of only favorable
reaction pathways from the complex chemical space through molecular graph and reaction network analysis.
This can be done very efficiently by exploring the routes connecting reactants and products with minimum
dissociation and formation of bonds. Finally, the resulting minimal network is subjected to quantum chemical
calculations to determine kinetically the most favorable reaction path at the predictable accuracy. To further
accelerate the graph-based method, we introduce state-of-the art machine learning techniques. They can replace
chemical heuristics and expensive calculations with more systematic, unbiased computational rules. In this talk,
we show the recent progress in this project with several examples.