Extended Data Fig. 3: Computational complexity of the molecule search space.

(left) Maximum number of molecule states versus maximum molecule size for the search tree described in this study. Extrapolating from these results yields approximately 1.9 × 10⁹ valid molecules with 12 or fewer heavy atoms. (right) Computational time required to enumerate the search space as a function of maximum molecule size. In addition to requiring more evaluations, larger molecules require additional computational time to check for a valid 3D embedding and to enumerate possible stereoisomers, and the time required for larger molecules may grow faster than simply exponential. An exponential fit to the last four datapoints indicates that a full enumeration of the 12 or fewer heavy atom search space would require 17.25 days. In addition to the time required to enumerate all candidates, a high-throughput screening would require evaluation of the final molecules with the reward function. This would require three separate neural network evaluations (which could be called in parallel) to estimate radical stability, redox potential, and X-H bond strength, and would likely add several days to the overall computational cost.