Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models.

Daniil Polykovskiy,Oktai Tatanov,Alán Aspuru-Guzik,Vladimir Aladinskiy,Mark Veselov,Sergey Nikolenko,Benjamin Sanchez-Lengeling,Rauf Kurbanov,Alex Zhavoronkov,Sergey Golovanov,Artur Kadurin,Simon Johansson,Hongming Chen,Alexander Zhebrak,Aleksey Artamonov,Stanislav Belyaev

doi:10.3389/fphar.2020.565644

Abstract

Generative models are becoming a tool of choice for exploring the molecular space. These models learn on a large training dataset and produce novel molecular structures with similar properties. Generated structures can be utilized for virtual screening or training semi-supervized predictive models in the downstream tasks. While there are plenty of generative models, it is unclear how to compare and rank them. In this work, we introduce a benchmarking platform called Molecular Sets (MOSES) to standardize training and comparison of molecular generative models. MOSES provides training and testing datasets, and a set of metrics to evaluate the quality and diversity of generated structures. We have implemented and compared several molecular generation models and suggest to use our results as reference points for further advancements in generative chemistry research. The platform and source code are available at https://github.com/molecularsets/moses.

Highlights

The discovery of new molecules for drugs and materials can bring enormous societal and technological progress, potentially curing rare diseases and providing a pathway for personalized precision medicine (Lee et al, 2018)
We focus on standardizing metrics and data for the distribution learning problem that we introduce below
We provide a benchmark suite—Molecular Sets (MOSES)—for molecular generation: a standardized dataset, data preprocessing utilities, evaluation metrics, and molecular generation models

Summary

Introduction

The discovery of new molecules for drugs and materials can bring enormous societal and technological progress, potentially curing rare diseases and providing a pathway for personalized precision medicine (Lee et al, 2018). Complete exploration of the huge space of potential chemicals is computationally intractable; it has been estimated that the number of pharmacologically-sensible molecules is in the order of 1023 to 1080 compounds (Kirkpatrick and Ellis, 2004; Reymond, 2015). Often, this search is constrained based on already discovered structures and desired qualities such as solubility or toxicity. Recent works demonstrated that machine learning methods can produce new small molecules (Merk et al, 2018a; Merk et al, 2018b; Polykovskiy et al, 2018b; Zhavoronkov et al, 2019a) and peptides (Grisoni et al, 2018) showing biological activity.

Objectives

Results

Conclusion