Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs

Dataset and additional libraries for "Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs" This repository contains the code implementation and experimental data used in our paper "Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs" (Priessner et al., 2025). Contents Experimental Dataset: Complete spectroscopic data (1H NMR, 13C NMR, HSQC, COSY) for 34 diverse organic molecules Simulated NMR data generated using the SGNN model Regioisomeric analogs (ROM) used for testing structure recovery from incorrect initial guesses Raw experimental spectra in industry-standard formats Code Implementation: Structure generation pipeline combining Chemformer, Mol2Mol, and MMST approaches HSQC peak matching and analysis algorithms LLM prompting templates for multimodal spectral analysis Scripts for DeepSeek-R1 and other LLM interaction Evaluation frameworks and metrics calculation Supplementary Materials: Detailed methodology documentation Performance comparison across different LLM models (Claude 3.5 Sonnet, Claude 3.7 Sonnet-Thinking, Gemini 2.0 Flash-Thinking, o3-mini, Kimi 1.5, and DeepSeek-R1) Confidence score analysis data This dataset supports the reproducibility of our integrated workflow that combines reasoning-capable language models with specialized chemical analysis tools to significantly improve molecular structure determination using nuclear magnetic resonance spectroscopy. Related Links Full code repository: GitHub Publication: [Journal reference when available] Citation If you use this dataset or code in your research, please cite: Priessner, M., Lewis, R.J., Johansson, M.J., Goodman, J.M., Janet, J.P., & Tomberg, A. (2025). Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs. ChemRxiv Contact For questions regarding the dataset or implementation, please contact: [email protected] [email protected] Note: This upload also contains additional work in two areas: (1) a multimodal spectrum transformer fine-tuned specifically on HSQC data, including the fine-tuned models, and (2) investigations into the latent spaces of encoded molecules using the multimodal spectrum transformer. These additional components can be found in the files labeled as mmt_explainability (main code) and mmt_identifier (latent space investigation).