GraphNVP: An Invertible Flow Model for Generating Molecular Graphs

TL;DR

GraphNVP introduces an invertible normalizing-flow model for molecular graphs, decomposing generation into two latent spaces: adjacency structure and node attributes. By applying two dedicated reversible coupling flows and a dequantization step, it enables exact likelihood training and one-shot generation of valid, highly unique molecular graphs. The approach yields near-perfect reconstruction, high uniqueness, and a latent space that supports property-directed molecule optimization (e.g., QED). Empirically, GraphNVP outperforms several VAE/GAN baselines on QM9 and ZINC-250k and offers a scalable framework for graph generation with precise probabilistic grounding. Limitations include permutation-invariance handling, suggesting directions for future work to strengthen symmetry properties while preserving performance.

Abstract

We propose GraphNVP, the first invertible, normalizing flow-based molecular graph generation model. We decompose the generation of a graph into two steps: generation of (i) an adjacency tensor and (ii) node attributes. This decomposition yields the exact likelihood maximization on graph-structured data, combined with two novel reversible flows. We empirically demonstrate that our model efficiently generates valid molecular graphs with almost no duplicated molecules. In addition, we observe that the learned latent space can be used to generate molecules with desired chemical properties.

Figures (7)

• Figure 1: Forward transformation of the proposed GraphNVP.
• Figure 2: Masking schemes used in proposed affine coupling layers.
• Figure 3: Generative process of the proposed GraphNVP. We apply the inverse of the coupling layers in the reverse order, so that the original input can be reconstructed.
• Figure 4: Visualization of the learned latent spaces along two randomly selected orthogonal axes. The red circled molecules are centers of the visualizations (not the origin of the latent spaces). An empty space in the grid indicates that an invalid molecule is generated.
• Figure 5: Chemical property optimization. Given the left-most molecule, we interpolate its latent vector along the direction which maximizes its QED property.