Year: 2023
Type: dissertation
Abstract: In recent years, de novo molecular design using machine learning has made significant technical progress, but practical applications still face challenges.The primary barrier to practical applications of computationally designed molecules is the cost and technical difficulty of synthesizing them.As a solution to this problem, various methods for synthetic route design using deep neural networks have been studied.However, designing molecules and their synthetic routes simultaneously has received little attention.In this study, we formulate the problem of designing molecules and their synthetic routes within the framework of Bayesian inference.The design variables in our framework consist of a set of reactants in a reaction network and its network topology.The design space is vast, consisting of all combinations of purchasable reactants, which can number in the millions or more.In addition, the designed reaction networks can adopt any topology beyond simple multistep linear reaction routes.Thus, this is a hard combinatorial problem that requires a powerful algorithm to solve.To address this problem, we present a powerful sequential Monte Carlo algorithm that recursively designs a synthetic reaction network by sequentially building up single-step reactions.Our algorithm is highly efficient, and in a case study of designing drug-like molecules based on commercially available compounds, it outperformed heuristic combinatorial search methods in terms of computational efficiency, coverage, and novelty with respect to existing compounds.Our framework has practical implications in the design of new molecules for various applications, such as drug design and materials science.Moreover, it offers a significant advantage in terms of computational efficiency, which is crucial for large-scale molecular design.Our study also provides the Python library "Seq-Stack-Reaction" with an illustrative example of designing highly viscous lubricant molecules.This library can be used to design molecules with desired properties and synthetic routes, making it a valuable tool for researchers in various fields.i v 5.1 Number of design molecules reaching the target property range at each step for scenario 1 where the number of commercially available reactants is limited to 10K.The correspondence between the seven methods and their labels is shown inTable 5.1. . . . . . . . . . . . .38 5.2 Cumulative number of hit molecules in scenario 1 where the number of commercially available reactants is limited to 10K.The left and right plots show the cumulative number of hits as a function of the total number of molecules generated and the execution time (CPU time), respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .38 5.3 Number of design molecules reaching the target property range at each step for scenario 2 where all 150K commercial compounds were used in the design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39 5.4 Cumulative number of hit molecules in scenario 2 where all 150K commercial compounds were used in the design.The left and right plots show the cumulative number of hits as a function of the total number of molecules generated and the execution time (CPU time), respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .40 5.5 1-coverage (horizontal axis) and novelty (vertical axis) of the set of designed virtual molecules with respect to existing molecules in ChM-BLE, which is drawn as a function of varying thresholds of Tanimoto similarity.The circle represents the coverage and novelty when the similarity threshold is set to 0.7. . . . . . . . . . . . . . . . . . . . . .41 5.6 Histograms of SA scores of hit compounds obtained from the proposed method and existing molecules in ChEMBL. . . . . . . . . . . . . . .
Authors: Zhang, Qi, チョウ, キ, Zhang, Qi
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7029911743