[論文レビュー] A Two-Step Graph Convolutional Decoder for Molecule Generation

Xavier Bresson, Thomas Laurent|arXiv (Cornell University)|Jun 8, 2019

Computational Drug Discovery Methods参考文献 22被引用数 55

ひとこと要約

本論文は、分子生成のための2段階オートエンコーダを提案する。最初に bag-of-atoms を生成し、次に グラフ畳み込みデコーダを用いて結合を組み立てる。ZINC分子で90.5%の再構成率と100%の妥当性を達成。さらにビームサーチとVAEフレームワークを用いて化学的性質を最適化する。

ABSTRACT

We propose a simple auto-encoder framework for molecule generation. The molecular graph is first encoded into a continuous latent representation $z$, which is then decoded back to a molecule. The encoding process is easy, but the decoding process remains challenging. In this work, we introduce a simple two-step decoding process. In a first step, a fully connected neural network uses the latent vector $z$ to produce a molecular formula, for example CO$_2$ (one carbon and two oxygen atoms). In a second step, a graph convolutional neural network uses the same latent vector $z$ to place bonds between the atoms that were produced in the first step (for example a double bond will be placed between the carbon and each of the oxygens). This two-step process, in which a bag of atoms is first generated, and then assembled, provides a simple framework that allows us to develop an efficient molecule auto-encoder. Numerical experiments on basic tasks such as novelty, uniqueness, validity and optimized chemical property for the 250k ZINC molecules demonstrate the performances of the proposed system. Particularly, we achieve the highest reconstruction rate of 90.5\%, improving the previous rate of 76.7\%. We also report the best property improvement results when optimization is constrained by the molecular distance between the original and generated molecules.

研究の動機と目的

• Motivate the design of a simple, efficient auto-encoder for generating valid molecules.
• Decouple atom generation from bond construction to simplify molecule decoding.
• Leverage a graph neural network in the decoder to place bonds given a latent representation.
• Integrate a variational auto-encoder framework to improve latent space structure.
• Demonstrate reconstruction, novelty, uniqueness, and property-optimization capabilities on ZINC data.

提案手法

• Encode molecular graphs into a fixed-size latent vector z using a graph convolutional network with node and edge features.
• Decode by first generating a molecular formula (bag of atoms) from z via a one-hidden-layer MLP.
• Assemble bonds by applying a graph convolutional network to the bag of atoms to predict bond types between atoms.
• Use a beam search to enforce chemical validity and select high-probability, valency-respecting bond configurations.
• Optionally adopt a variational auto-encoder formulation to model z as z=μ+σ⊙ε and optimize via a KL-divergence loss.
• Train and evaluate on the ZINC dataset for reconstruction, validity, novelty, uniqueness, and property-optimization metrics.

実験結果

リサーチクエスチョン

• RQ1 Can a two-step, non-autoregressive decoder reliably reconstruct and generate valid molecular graphs?
• RQ2 Does separating atom generation from bond placement improve reconstruction rates and validity on large molecule datasets?
• RQ3 How does a VAE formulation affect the latent space and reconstruction quality for molecules?
• RQ4 What is the capability of the model to generate novel molecules and optimize chemical properties under constraints?

主な発見

方法	再構成	妥当性
CVAE (Gómez-Bombarelli et al. 2018)	44.6%	0.7%
GVAE (Kusner et al. 2017)	53.7%	7.2%
SD-VAE (Dai et al. 2018)	76.2%	43.5%
GraphVAE (Simonovsky & Komodakis 2018)	-	13.5%
JT-VAE (Jin et al. 2018)	76.7%	100.0%
GCPN (You et al. 2018)	-	-
OURS	90.5%	100.0%

• Achieved 90.5% reconstruction rate on 250k ZINC molecules, improving prior state-of-the-art 76.7% (with 100% validity).
• Obtained 100% validity for reconstructed molecules, including those not perfectly reconstructed.
• Generated 100% novel and unique molecules when sampling from the prior distribution (n=5000).
• Best property-improvement results when optimization is constrained by molecular distance, compared to prior VAE methods.
• Beam search contributes to producing chemically valid molecules and can be parallelized for efficiency.