MOLECULE OPTIMIZATION BY EXPLAINABLE EVOLUTION

Abstract

Optimizing molecules for desired properties is a fundamental yet challenging task in chemistry, material science, and drug discovery. This paper develops a novel algorithm for optimizing molecular properties via an Expectation-Maximization (EM) like explainable evolutionary process. The algorithm is designed to mimic human experts in the process of searching for desirable molecules and alternate between two stages: the first stage on explainable local search which identifies rationales, i.e., critical subgraph patterns accounting for desired molecular properties, and the second stage on molecule completion which explores the larger space of molecules containing good rationales. We test our approach against various baselines on a real-world multi-property optimization task where each method is given the same number of queries to the property oracle. We show that our evolution-by-explanation algorithm is 79% better than the best baseline in terms of a generic metric combining aspects such as success rate, novelty, and diversity. Human expert evaluation on optimized molecules shows that 60% of top molecules obtained from our methods are deemed successful.

1. INTRODUCTION

The space of organic molecules is vast, the size of which is exceeding 10 60 (Reymond et al., 2010) . Searching over this vast space for molecules of interest is a challenging task in chemistry, material science, and drug discovery, especially given that molecules are desired to meet multiple criteria, e.g., high potency and low toxicity in drug discovery. When human experts optimize molecules for better molecular properties, they will first come up with rationales within desirable molecules. Typically, the rationales are subgraphs in a molecule deemed to contribute primarily to certain desired molecular properties. Once rationales are identified, chemists will design new molecules on top of rationales hoping that, the desired properties of new molecules will be further enhanced due to the existence of rationale and changes of non-rationale parts. The cycle of identifying molecular rationales and redesigning new hypothetical molecules will be carried on until molecules that meet certain property criteria are discovered. In this paper, we develop a novel algorithm that mimics the process of molecule optimization by human experts. Our algorithm finds new molecules with better properties via an EM-like explainable evolutionary process (Figure 1 ). The algorithm alternates between two stages. During the first stage, we use an explainable local search method to identify rationales within high-quality molecules that account for their high property scores. During the second stage, we use a conditional generative model to explore the larger space of molecules containing useful rationales. Our method is novel in that we are using explainable models to help us exploit useful patterns in the molecules, yet leveraging generative models to help us explore the molecule landscape. Comparing to existing methods that directly learn a generative model using Reinforcement Learning or perform continuous optimization in the latent space of molecules (Olivecrona et al., 2017; You et al., 2018a; Dai et al., 2018b) , our method is more sample-efficient and can generate more novel and unique molecules that meet the criteria. We evaluate our algorithm against several state-of-the-art methods on a molecule optimization task involving multiple properties. Compared with baselines, our algorithm is able to increase the success In the E-step of time t, we draw samples from Q(p(s)|θ t-1 ) to approximate Q(θ|p t (s)) using rationales extracted from the seed molecules via an explainable model. Then in the M-step, we optimize Q(θ|p t (s)) w.r.t. θ, i.e. pushing the graph completion model p θ (•|s) towards generating higher scoring molecules conditioned on the rationale samples. rate by 50%, novelty by 14%, while having a competitive diversity. We further propose a new metric, QNU score, to jointly consider all three aspects, and show that we achieve a score of 52.7% compared with 29.5% by the best baseline. We also ask experienced chemists to evaluate top-50 generated molecules and find that 30 of them are as good as existing ones. The main contributions of this paper are summarized below: • We propose a novel EM-like evolution-by-explanation algorithm for molecule optimization; • We present a novel, principled, explainable graph model based on an information-theoretic approach to extract subgraphs essential for maintaining certain desired properties; • Our approach outperforms existing state-of-the-arts by a large margin in terms of success rate (50% better), novelty (14% better), and an overall metric (79% better) on a real-world multiproperty optimization task.

2. RELATED WORK

There has been a surge of interest in using machine learning to discover novel molecules with certain properties in recent years. Most of the existing work defines a generative model for either the SMILES strings (Weininger, 1988) or molecular graphs, and uses Reinforcement Learning algorithms to optimize the properties of the generated molecules (Segler et al., 2018; Olivecrona et al., 2017; Guimaraes et al., 2017; You et al., 2018a; Popova et al., 2018; 2019; Samanta et al., 2019; Zhou et al., 2019; De Cao & Kipf, 2018; Kearnes et al., 2019; Shi et al., 2020; Jin et al., 2020) . Others optimize the continuous representation of molecules in a latent space learned by variants of variational autoencoders (Kusner et al., 2017; Dai et al., 2018b; Jin et al., 2018; Gómez-Bombarelli et al., 2018; Kang & Cho, 2018; Liu et al., 2018; Kajino, 2019) . More recent work attempts Evolutionary algorithms (Nigam et al., 2020; Leguy et al., 2020; Winter et al., 2019) , or focuses on finding high-quality molecules with synthesis paths (Bradshaw et al., 2019; Korovina et al., 2020; Gottipati et al., 2020) . Most similar to our approach is RationaleRL (Jin et al., 2020) , which extracts subgraphs from seed molecules using Monte Carlo Tree Search (MCTS) and generates full molecules by completing the subgraphs. Compared with previous work, our approach is the first to incorporate an explainable model in the iterative search process. Existing work on explainable models approaches the problems from three directions. The first line of work uses gradients of the outputs with respect to inputs to identify the salient features in the inputs (Simonyan et al., 2013; Springenberg et al., 2014; Baehrens et al., 2010) ; the second line of work approximates the model with simple interpretable models, such as locally additive mod-



Figure1: Overview of our EM-like evolution-by-explanation algorithm. Left: climbing up the energy landscape J(θ, p(s)) by alternatively taking an E-step and M-step. Right: illustrations for the E-step and M-step. In the E-step of time t, we draw samples from Q(p(s)|θ t-1 ) to approximate Q(θ|p t (s)) using rationales extracted from the seed molecules via an explainable model. Then in the M-step, we optimize Q(θ|p t (s)) w.r.t. θ, i.e. pushing the graph completion model p θ (•|s) towards generating higher scoring molecules conditioned on the rationale samples.

