[Paper Review] De Novo Molecular Generation via Connection-aware Motif Mining

Zijie Geng, Shufang Xie+7 more|arXiv (Cornell University)|Feb 2, 2023

Machine Learning in Materials Science8 citations

TL;DR

MiCaM mines connection-aware motifs from large molecular libraries and uses them in a generator to create de novo molecules, achieving state-of-the-art results on distribution and goal-directed benchmarks.

ABSTRACT

De novo molecular generation is an essential task for science discovery. Recently, fragment-based deep generative models have attracted much research attention due to their flexibility in generating novel molecules based on existing molecule fragments. However, the motif vocabulary, i.e., the collection of frequent fragments, is usually built upon heuristic rules, which brings difficulties to capturing common substructures from large amounts of molecules. In this work, we propose a new method, MiCaM, to generate molecules based on mined connection-aware motifs. Specifically, it leverages a data-driven algorithm to automatically discover motifs from a molecule library by iteratively merging subgraphs based on their frequency. The obtained motif vocabulary consists of not only molecular motifs (i.e., the frequent fragments), but also their connection information, indicating how the motifs are connected with each other. Based on the mined connection-aware motifs, MiCaM builds a connection-aware generator, which simultaneously picks up motifs and determines how they are connected. We test our method on distribution-learning benchmarks (i.e., generating novel molecules to resemble the distribution of a given training set) and goal-directed benchmarks (i.e., generating molecules with target properties), and achieve significant improvements over previous fragment-based baselines. Furthermore, we demonstrate that our method can effectively mine domain-specific motifs for different tasks.

Motivation & Objective

• Motivate improved fragment-based de novo molecular generation by moving beyond heuristic motif vocabularies.
• Develop a data-driven method to mine frequent, connection-aware motifs from large molecule libraries.
• Create a generator that simultaneously selects motifs and determines how they connect to build valid molecules.
• Demonstrate superior distribution-learning and goal-directed generation performance on standard benchmarks.
• Show that domain-specific motifs can be effectively mined for task-specific generation.

Proposed method

• Mine a motif vocabulary by iteratively merging frequent adjacent subgraphs to form connection-aware motifs.
• Preserve connection information by marking broken bonds with * and using motif graphs represented with GNNs.
• Use a VAE framework to map molecules to latent vectors and condition generation on z and motif representations.
• In generation, query either motif connection sites or current molecule sites to decide the next connection or cyclization.
• Generate by either attaching a new motif or merging sites to form rings, guided by a start and query network.
• Train with reconstruction loss, KL-divergence regularization, and a property-prediction loss to align latent space with molecular properties.

Experimental results

Research questions

• RQ1 Can a data-driven merging strategy discover meaningful, connection-aware motifs that improve generation quality over heuristic vocabularies?
• RQ2 Does a connection-aware motif vocabulary plus a motif-aware generator improve distribution-fitting (KL Div, FCD) and uniqueness/novelty on standard benchmarks?
• RQ3 Can the motif vocabulary and network parameters be jointly adapted to domain-specific tasks to achieve state-of-the-art goal-directed generation?
• RQ4 How does controlling merging-operation counts affect similarity to training data versus novelty?
• RQ5 Do greedy versus distributional generation modes trade off KL Div/FCD against novelty?

Key findings

• MiCaM achieves best KL Divergence and Fréchet ChemNet Distance (FCD) across QM9, ZINC, and GuacaMol datasets among compared baselines.
• MiCaM maintains high validity, uniqueness, and novelty while improving distributional similarity to training sets.
• A medium number of merging operations (~500) yields high similarity; more operations increase motif size and similarity but can reduce novelty.
• Distributional-mode generation yields higher novelty than greedy mode, while greedy mode slightly boosts similarity metrics.
• In goal-directed benchmarks, MiCaM attains strong scores and, when combined with iterative target augmentation, achieves state-of-the-art results on multiple tasks.
• Case studies show domain-specific motifs drive target-property improvements in complex molecules.