Learning Multimodal Graph-to-Graph Translation for Molecular Optimization

TL;DR

This work reframes molecular optimization as multimodal graph-to-graph translation, proposing a junction tree encoder-decoder that generates valid molecular graphs and accommodates diverse outputs through latent codes. It introduces a variational extension (VJTNN) that derives latent representations from input–output differences and an adversarial scaffold regularization to align generated scaffolds with the target domain. Empirical results across three optimization tasks show superior translation quality and diversity compared to state-of-the-art baselines, highlighting improved data efficiency from supervised, parallel-data learning. The approach offers a principled, scalable route to generate multiple plausible, property-improving molecules for drug design.

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.

Figures (4)

• Figure 1: Illustration of our encoder-decoder model. Molecules are represented by their graph structures and junction trees encoding the scaffold of molecules. Nodes in the junction tree (which we call clusters) are valid chemical substructures such as rings and bonds. During decoding, the model first generates its junction tree and then combines clusters in the predicted tree into a molecule.
• Figure 2: Multiple ways to assemble neighboring clusters in the junction tree.
• Figure 3: Multimodal graph-to-graph learning. Our model combines the strength of both variational JTNN and adversarial scaffold regularization.
• Figure 4: Examples of diverse translations learned by VJTNN+GAN on QED and DRD2 dataset.