How do chemists find the best molecule for a particular reaction? Chemical space, the collection of all possible molecules, both previously reported and hypothetical, is infinite. The subspace of only medium-sized organic molecules, molecules consisting of hydrogen, carbon, oxygen, nitrogen, and sulfur atoms, is estimated to be 10.60. This enormity requires powerful and reliable computational tools to identify the final list of candidates for subsequent experimental testing. The traditional solution to this challenge works as a funnel. The part of the chemical space containing the compounds considered most promising for a particular practical application is sifted through a set of specific criteria, such as optimal value of a particular chemical or physical property, ease, etc. Synthesis of candidate molecules in the laboratory, etc. Although powerful, these techniques are biased toward the types of compounds initially selected, leaving little room for discovering new molecules.Writing in progress natural computational scienceRenana Gershoni-Poranne, Alex Bronstein and colleagues demonstrate how to discover new and better-performing molecules through bottom-up design, combining diffusion models for structure generation and neural networks for property prediction. Masu.1.
