<div>
<div>
<div>
<p>Automated reaction prediction has the potential to elucidate complex reaction networks
for applications ranging from combustion to materials degradation. Although substantial
progress has been made in predicting specific reaction pathways and resolving mechanisms,
the computational cost and inconsistent reaction coverage of automated prediction are still
obstacles to exploring deep reaction networks without using heuristics. Here we show that cost
can be reduced and reaction coverage can be increased simultaneously by relatively straight-
forward modifications of the reaction enumeration, geometry initialization, and transition
state convergence algorithms that are common to many emerging prediction methodologies.
These changes are implemented in the context of Yet Another Reaction Program (YARP), our
reaction prediction package, for which we report a head-to-head comparison with prevailing
methods for two benchmark reaction prediction tasks. In all cases, we observe near perfect
recapitulation of established reaction pathways and products by YARP, without the use of
heuristics or other domain knowledge to guide reaction selection. In addition, YARP also discovers many new kinetically relevant pathways and products reported here for the first
time. This is achieved while simultaneously reducing the cost of reaction characterization
nearly 100-fold and increasing transition state success rates and intended rates over 2-fold
and 10-fold, respectively, compared with recent benchmarks. This combination of ultra-low
cost and high reaction-coverage creates opportunities to explore the reactivity of larger sys-
tems and more complex reaction networks for applications like chemical degradation, where
approaches based on domain heuristics fail. </p>
</div>
</div>
</div>
Deep Retrosynthetic Reaction Prediction using Local Reactivity and Global Attention
