SynCraft: Datasets for Synthesizability Optimization via LLM-Guided Molecular Editing

This dataset accompanies the paper "Guiding Large Language Models to Predict Edit Sequences for Molecular Synthesizability Optimization" (Nature Machine Intelligence, 2026). Contents: - Test sets: Unsynthesizable molecule pools from Pocket2Mol, ResGen, and RIPK1 case study - Synthesis-cliff training corpus: 3,332 pairs with reasoning traces - Multi-LLM inference outputs: Results from Gemini-2.5-Pro, GPT-5.2, DeepSeek-V4-Pro, DeepSeek-V4-Flash - PLIP retention tables: Binding-mode preservation data for Mpro and RIPK1 - Case-study structures: TargetDiff Mpro rescue demonstration All data are validated with SimpRetro + Enamine building blocks (30-minute timeout on Intel Xeon E5-2670 CPU).