[論文レビュー] Multi-Objective Molecule Generation using Interpretable Substructures

Wengong Jin, Regina Barzilay+1 more|arXiv (Cornell University)|Feb 8, 2020

Computational Drug Discovery Methods参考文献 43被引用数 86

ひとこと要約

本論文はRationaleRLを提案する。解釈可能なサブ構造から分子を組み立て、強化学習を通じて複数の性質を最適化する、マルチ目的分子設計の根拠ベースのグラフ生成モデルで、いくつかのタスクで最先端の結果を達成している。

ABSTRACT

Drug discovery aims to find novel compounds with specified chemical property profiles. In terms of generative modeling, the goal is to learn to sample molecules in the intersection of multiple property constraints. This task becomes increasingly challenging when there are many property constraints. We propose to offset this complexity by composing molecules from a vocabulary of substructures that we call molecular rationales. These rationales are identified from molecules as substructures that are likely responsible for each property of interest. We then learn to expand rationales into a full molecule using graph generative models. Our final generative model composes molecules as mixtures of multiple rationale completions, and this mixture is fine-tuned to preserve the properties of interest. We evaluate our model on various drug design tasks and demonstrate significant improvements over state-of-the-art baselines in terms of accuracy, diversity, and novelty of generated compounds.

研究の動機と目的

• Address the challenge of designing molecules that satisfy multiple property constraints simultaneously.
• Identify small, property-driven substructures (rationales) that influence specific properties.
• Assemble full molecules by expanding rationales and fine-tuning mixtures to preserve target properties.
• Enable interpretable molecule generation by exposing the rationale vocabulary to users and domain experts.

提案手法

• Extract single-property rationales from positive molecules using Monte Carlo Tree Search to find connected subgraphs with high predicted property score and small size.
• Merge single-property rationales into multi-property rationales via maximum common substructure (MCS) and superposition to satisfy multiple constraints.
• Train a graph completion model P(G|S) as a variational autoencoder that expands a given rationale S into a full molecule G while ensuring S is contained in G.
• Learn the rationale distribution P(S) to prefer rationales that are more likely to yield positive molecules, with entropy regularization to encourage exploration.
• Pre-train the graph generator on ChEMBL-derived data to learn realistic expansion, then fine-tune with policy gradient using property predictors as reward.
• Use Frechet ChemNet Distance (FCD) and toxicity rationale evaluation to assess distributional similarity and fidelity of rationales.

実験結果

リサーチクエスチョン

• RQ1 How can multi-property molecular design be achieved by decomposing molecules into interpretable substructures (rationales)?
• RQ2 Can a rationale-conditioned graph generator expand rationales into realistic molecules that satisfy multiple property constraints?
• RQ3 Does learning a rationale distribution P(S) improve multi-property optimization compared to generation from scratch?
• RQ4 Do rationales correspond to chemically meaningful substructures, and can they aid toxicity-related explanations?
• RQ5 How does RationaleRL compare to state-of-the-art baselines under various multi-property constraint settings?

主な発見

手法	GSK3β_Success	GSK3β_Novelty	GSK3β_Diversity	JNK3_Success	JNK3_Novelty	JNK3_Diversity	GSK3β+JNK3_Success	GSK3β+JNK3_Novelty	GSK3β+JNK3_Diversity
JT-VAE	32.2%	11.8%	0.901	23.5%	2.9%	0.882	3.3%	7.9%	0.883
GCPN	42.4%	11.6%	0.904	32.3%	4.4%	0.884	3.5%	8.0%	0.874
GVAE-RL	33.2%	76.4%	0.874	57.7%	62.6%	0.832	40.7%	80.3%	0.783
REINVENT	99.3%	61.0%	0.733	98.5%	31.6%	0.729	97.4%	39.7%	0.595
RationaleRL	100%	53.4%	0.888	100%	46.2%	0.862	100%	97.3%	0.824

• RationaleRL achieves state-of-the-art performance across single-, two-, and four-property constraint tasks in terms of success, novelty, and diversity.
• On two-property constraints, RationaleRL attains 100% success with high novelty (100%) and strong diversity (0.824).
• On four-property constraints, RationaleRL substantially outperforms baselines (e.g., 74.8% vs 47.9% success, 0.701 vs 0.621 diversity).
• Ablation studies show rationales provide clear benefits over generation from scratch (GVAE-RL baseline).
• Rationales extracted via MCTS cover chemical space of known positives, and generated dual inhibitors lie distributionally closer to true positives (lower FCD than REINVENT).
• Rationale accuracy on toxicity-related evaluation indicates meaningful and faithful rationales, with partial and exact match metrics favoring the proposed approach.