Reproduce “The Synthesizability of Molecules Proposed by Generative Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wenhao Gao, Connor W. Coley

This research quantifies the practical synthesizability of molecules proposed by state-of-the-art deep generative models, revealing that a substantial portion are experimentally intractable. The study demonstrates that biasing molecular generation through careful training set selection or integration of heuristic scores, particularly the SA Score, can significantly improve the synthetic accessibility of proposed compounds.

Key insights

  • Generative models, especially those optimized for specific objectives, frequently propose molecules that are experimentally challenging or impossible to synthesize, despite achieving high in silico performance.
  • The synthesizability of generated molecules is highly influenced by the synthetic accessibility of the compounds in their respective training datasets.
  • Integrating computationally inexpensive heuristic scores, particularly the SA Score, can effectively bias goal-directed generation towards synthetically tractable compounds, providing a practical method to enhance model utility.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.