[Paper Review] Data-Efficient Graph Grammar Learning for Molecular Generation

Minghao Guo, Veronika Thost+4 more|arXiv (Cornell University)|Mar 15, 2022

Machine Learning in Materials Science17 citations

TL;DR

This paper proposes a data-efficient graph grammar learning framework for molecular generation that automatically constructs interpretable, substructure-level production rules from minimal training data (as few as 11–32 samples). By jointly optimizing for molecular diversity, synthesizability, and distributional fidelity, the method achieves state-of-the-art performance on monomer and polymer generation tasks with orders-of-magnitude less data than standard deep learning baselines.

ABSTRACT

The problem of molecular generation has received significant attention recently. Existing methods are typically based on deep neural networks and require training on large datasets with tens of thousands of samples. In practice, however, the size of class-specific chemical datasets is usually limited (e.g., dozens of samples) due to labor-intensive experimentation and data collection. This presents a considerable challenge for the deep learning generative models to comprehensively describe the molecular design space. Another major challenge is to generate only physically synthesizable molecules. This is a non-trivial task for neural network-based generative models since the relevant chemical knowledge can only be extracted and generalized from the limited training data. In this work, we propose a data-efficient generative model that can be learned from datasets with orders of magnitude smaller sizes than common benchmarks. At the heart of this method is a learnable graph grammar that generates molecules from a sequence of production rules. Without any human assistance, these production rules are automatically constructed from training data. Furthermore, additional chemical knowledge can be incorporated in the model by further grammar optimization. Our learned graph grammar yields state-of-the-art results on generating high-quality molecules for three monomer datasets that contain only ${\sim}20$ samples each. Our approach also achieves remarkable performance in a challenging polymer generation task with only $117$ training samples and is competitive against existing methods using $81$k data points. Code is available at https://github.com/gmh14/data_efficient_grammar.

Motivation & Objective

• To address the challenge of molecular generation with extremely limited training data, common in specialized chemical domains like polymer monomers.
• To overcome the limitations of deep learning models that require large datasets and often generate invalid or non-synthesizable molecules.
• To develop a grammar-based generative model that is both data-efficient and capable of incorporating complex chemical knowledge beyond basic valency rules.
• To automatically construct graph grammars from data without human-designed rules, while optimizing for key chemical metrics such as diversity and synthesizability.
• To enable high-quality, physically plausible molecular generation even when training data is scarce, particularly for class-specific chemical spaces.

Proposed method

• The method learns a graph grammar from molecular training data by automatically discovering production rules that generate valid molecular structures through substructure-level transformations.
• The grammar construction process jointly optimizes for multiple chemical objectives—diversity, synthesizability (RS), and distributional similarity—using a weighted loss function with a balance factor λ.
• Production rules are learned via differentiable optimization over graph structures, allowing the model to capture complex substructures (e.g., rings, functional groups) rather than just atom-by-atom additions.
• The framework supports end-to-end training and can be fine-tuned with additional chemical constraints or metrics, enabling integration of domain-specific knowledge.
• The model uses a symbolic, rule-based representation that ensures interpretability and allows for functional group extraction and analysis.
• The approach is evaluated on monomer and polymer datasets, including a 117-sample polymer dataset and 3 monomer datasets with only ~20 samples each.

Experimental results

Research questions

• RQ1 Can a graph grammar be automatically constructed from minimal chemical data (e.g., 11–32 samples) without human-designed rules?
• RQ2 How well can a data-efficient grammar-based model perform on molecular generation tasks compared to deep learning baselines when training data is extremely limited?
• RQ3 Can the grammar construction process effectively optimize for complex chemical metrics such as diversity and synthesizability while maintaining structural validity?
• RQ4 To what extent can the learned grammar generalize to generate novel, physically plausible molecules not present in the training set?
• RQ5 How does the balance between different optimization objectives (e.g., diversity vs. synthesizability) affect the quality and diversity of generated molecules?

Key findings

• The proposed method achieved state-of-the-art performance on three monomer datasets with only ~20 training samples each, outperforming existing grammar-based and deep learning models in sample quality and distributional fidelity.
• On a challenging polymer generation task with 117 training samples, the model achieved performance competitive with existing methods trained on 81,000 data points.
• The model generated novel, complex molecules such as cyclooctane, which were not present in the training data, demonstrating strong generalization and structural creativity.
• The Chamfer distance of the generated molecules was twice as high as that of strong baselines like HierVAE, indicating superior distributional alignment despite minimal data.
• The balance factor λ successfully controlled the trade-off between diversity and synthesizability, with λ₁=1, λ₂=2 achieving a well-balanced performance.
• Functional groups characteristic of each monomer class (e.g., isocyanate) were automatically identified as shared production rules, confirming the model’s interpretability and alignment with chemical knowledge.