GraViti: Graph-Level Variational Autoencoder for Molecular Generation

Researchers have unveiled GraViti, a transformer-based variational autoencoder designed for graph-level applications, which translates entire graphs into compact latent vectors, facilitating seamless interpolation and property-guided searches. This innovative approach creates a genuine graph-level latent space, in contrast to node-level embeddings. In molecular benchmarks, GraViti successfully decodes valid samples that comply with chemical constraints, directly retrieving domain rules from graph-level representations. The findings indicate that enforcing permutation invariance may hinder consistent reconstruction in domains with dependable canonical node orderings, like molecules or Bayesian networks. GraViti excels in reconstruction accuracy on extensive datasets and demonstrates strong generative capabilities with single-step decoding, serving as an efficient alternative to more intricate generation methods.