Unraveling Retrosynthetic Pathways through Template-Free Generative Models

Retrosynthesis, the process of identifying reaction pathways to synthesize target molecules, is a critical task in computational chemistry. This work discusses various approaches to framing and solving the retrosynthetic problem using deep neural networks, distinguishing between single-step and multi-step methods, template-based and template-free models, and graph-based versus sequence-based representations. This project uses a generative model to generate the reaction pathways where each molecule is encoded in an embedding space. Three encoder architectures—Normalizing Flows (NF), Variational Autoencoders (VAE), and Autoencoders (AE)—are evaluated for their ability to encode molecular structures effectively. The generative model employed, Insert-Fill-Halt (IFH), generates retrosynthetic routes. The study also discusses appropriate evaluation metrics: reconstruction accuracy for encoders and tree edit distance, clustering diversity, and the proportion of solved routes for generative performance. The results reveal the strengths and limitations of the VAE and AE in reconstructive accuracy and concluding remarks emphasize directions for future works.