Chemical space exploration: how genetic algorithms find the needle in the haystack

Abstract

We explain why search algorithms can find molecules with particular properties in an enormous chemical space (ca 1060 molecules) by considering only a tiny subset (typically 103−6 molecules). Using a very simple example, we show that the number of potential paths that the search algorithms can follow to the target is equally vast. Thus, the probability of randomly finding a molecule that is on one of these paths is quite high and from here a search algorithm can follow the path to the target molecule. A path is defined as a series of molecules that have some non-zero quantifiable similarity (score) with the target molecule and that are increasingly similar to the target molecule. The minimum path length from any point in chemical space to the target corresponds is on the order of 100 steps, where a step is the change of and atom- or bond-type. Thus, a perfect search algorithm should be able to locate a particular molecule in chemical space by screening on the order of 100s of molecules, provided the score changes incrementally. We show that the actual number for a genetic search algorithm is between 100 and several millions, and depending on the target property and its dependence on molecular changes, the molecular representation, and the number of solutions to the search problem.

Perspective and challenges in electrochemical approaches for reactive CO2 separations

“…With such descriptors as starting points for emerging genetic algorithms ( Henault et al., 2020 ; Jensen, 2019 ) or reinforcement learning methods ( Gómez-Bombarelli et al., 2018 ), the discovery of novel redox-active organic compounds with desired performance characteristics can be automated. Examples of the use of DFT-driven screening and discovery based on machine learning include the identification of organic chromophores for photovoltaic applications ( Hachmann et al., 2011 ).…”

Illuminating Elite Patches of Chemical Space

“…To study the difference in efficiency of GB-EPI and GB-GA, we make a statistical analysis of a representative rediscovery task (troglitazone). In line with earlier work 14,50 on the efficiency of GB-GA, we calculate the average number of tness function evaluations and CPU time needed for rediscovery, and the rediscovery success rate of both algorithms. As we learned from the median molecule task, starting from a randomised set of molecules elucidates the exploratory power of the algorithms more.…”

“…Therefore, we start this rediscovery task with the 100 topscoring molecules from 10 000 molecules randomly chosen from a 1.6 million ChEMBL subset, as constructed by Henault et al 50 In this subset all molecules with a bit-vector Tanimoto similarity to the target above 0.323 are removed. 13 Table 2 shows the results for 100 runs of GB-EPI and GB-GA (with settings taken from Henault et al 50 ), both with a maximum of 1000 generations per run.…”

Generating stable molecules using imitation and reinforcement learning

“…Instead an automated search for interesting candidates is desired. Examples of such search methods include evolutionary algorithms [1,2], basin-hopping [3] and particle swarm optimization [4].…”