Synthesis planning relies on retrosynthesis models, yet this relationship is under-analyzed. We investigate the effect of contemporary single-step models trained on public and proprietary reaction data to analyze the synthesis routes produced.
Introducing novel evaluation metrics for retrosynthesis that address dataset imperfections using custom scoring metrics and presenting SynFormer, a transformer-based model that achieves state-of-the-art accuracy efficiently without pre-training.
Expanding the chemical space by inferring new chemical reactions through link prediction (SEAL) in a Chemical Reaction Knowledge Graph (CRKG). From high probability links, de novo products can be generated using a molecular transformer (Chemformer).
Pretraining of NERF models on chemically related mechanisms significantly improves the performance compared to pretraining by larger, mechanistically dissimilar reaction datasets.
NNAA-Synth is a synthesis assistance tool for SPPS-compatible non-natural amino acids that integrates protection strategies and evaluates reaction feasibility.