[Paper Review] Learning Multimodal Graph-to-Graph Translation for Molecular Optimization

Wengong Jin, Kevin Yang+2 more|arXiv (Cornell University)|Dec 3, 2018

Machine Learning in Materials Science51 references154 citations

TL;DR

The paper proposes a multimodal graph-to-graph translation model for molecular optimization using a junction tree encoder-decoder with latent codes and adversarial scaffold regularization, achieving state-of-the-art results and diverse outputs across several tasks.

ABSTRACT

We view molecular optimization as a graph-to-graph translation problem. The goal is to learn to map from one molecular graph to another with better properties based on an available corpus of paired molecules. Since molecules can be optimized in different ways, there are multiple viable translations for each input graph. A key challenge is therefore to model diverse translation outputs. Our primary contributions include a junction tree encoder-decoder for learning diverse graph translations along with a novel adversarial training method for aligning distributions of molecules. Diverse output distributions in our model are explicitly realized by low-dimensional latent vectors that modulate the translation process. We evaluate our model on multiple molecular optimization tasks and show that our model outperforms previous state-of-the-art baselines.

Motivation & Objective

• Motivate molecular optimization as a graph-to-graph translation problem with parallel data.
• Enable diverse (multimodal) outputs by injecting low-dimensional latent codes into decoding.
• Ensure chemical validity and target-domain alignment via adversarial scaffold regularization.
• Leverage a junction-tree based encoder-decoder to hierarchically generate valid molecular graphs.
• Demonstrate improvements over baselines across multiple property optimization tasks.

Proposed method

• Encode molecules with a shared junction-tree and graph encoder using graph message passing.
• Decode first a junction tree with a tree-GRU and attention to predict scaffolds, then assemble the graph from clusters.
• Introduce low-dimensional latent codes z to capture multimodal translations and train with a variational objective.
• Apply adversarial regularization on continuous scaffold representations to align translated outputs with the target domain.
• Train a graph decoder with a contextual loss (Eq. 10) and utilize an adversarial (WGAN-GP) objective over decoder embeddings to enforce scaffold realism.
• Optionally, use a VAE-like objective with a KL term to encourage latent codes to follow a prior.

Experimental results

Research questions

• RQ1 Can a multimodal graph-to-graph translation framework learn diverse, property-improving molecular mappings from paired data?
• RQ2 Does a junction-tree based encoder-decoder improve validity and quality of generated molecules compared to sequence-based or flat graph approaches?
• RQ3 Do latent codes enable meaningful diversity without sacrificing translation accuracy or chemical validity?
• RQ4 Does adversarial scaffold regularization help align translated molecules with the target domain distributions?
• RQ5 How does the proposed method compare to MMPA, JT-VAE, GCPN, and VSeq2Seq baselines on penalized logP, QED, and DRD2 optimization tasks?

Key findings

Method	δ=0.6 (Improvement)	δ=0.6 (Diversity)	δ=0.4 (Improvement)	δ=0.4 (Diversity)	Improvement	Diversity
MMPA	1.65±1.44	0.329	3.29±1.12	0.496	-	-
JT-VAE	0.28±0.79	-	1.03±1.39	-	-	-
GCPN	0.79±0.63	-	2.49±1.30	-	-	-
VSeq2Seq	2.33±1.17	0.331	3.37±1.75	0.471	-	-
VJTNN	2.33±1.24	0.333	3.55±1.67	0.480	-	-

• The model outperforms baselines on penalized logP, achieving higher improvements and diverse outputs across two similarity thresholds.
• On QED and DRD2 tasks, the method (VJTNN and VJTNN+GAN) shows higher success rates and competitive diversity and novelty metrics compared to baselines.
• VJTNN+GAN provides marginal diversity gains over VJTNN, with strong performance on QED and DRD2 under explicit target domains.
• The approach yields diverse and novel molecules while leveraging parallel data for sample efficiency beyond MMPA rules-based methods.
• A junction-tree based encoder-decoder with latent multimodal decoding attains superior translation accuracy and property improvements.