Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs

Zenodo Dataset Description: ChemStructLLM Title Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs - Dataset and Pre-trained Models Creators Priessner, Martin (Contact person) - ORCID: 0009-0005-7349-283X Lewis, R.J. Johansson, M.J. Goodman, J.M. Janet, J.P. Tomberg, A. Description This Zenodo repository contains the essential data files and pre-trained models required to reproduce the experimental results from our paper "Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs" (Priessner et al., 2025). Contents Overview This dataset provides three main components necessary for full system reproduction: 1. Pre-trained Models (models.zip) Contains all trained machine learning models used in the ChemStructLLM pipeline: Chemformer Models: Forward and retrosynthesis prediction models Location: models/chemformer/ Pre-trained transformer models for molecular transformation prediction MMST (Multimodal Spectrum Transformer) Models: Location: models/mmst/base_models/ Fine-tuned models for spectral data analysis Includes HSQC-specific fine-tuned variants Mol2Mol Models: Location: models/mol2mol/Alessandro_big/ Molecular transformation and structure generation models 2. Molecular Data and Test Cases (data_molecular_structure.zip) Experimental and reference data for system validation: Molecular Database: data/molecular_data/molecular_data.json Complete spectroscopic data (1H NMR, 13C NMR, HSQC, COSY) for 34 diverse organic molecules Experimental NMR measurements in structured JSON format SMILES representations and molecular identifiers Archive Data: data/molecular_data/archive/ Historical data versions and backup files Additional molecular datasets used during development 3. Pre-computed Workflow Results (intermediate_results_selected.zip) Example workflow outputs for analysis and validation: Simulation Results: _run_1_sim_finished/ & _run_1_sim_finished_clean/: Basic simulation workflow _run_2_sim_aug_finished/ & _run_2_sim_aug_finished_clean/: Augmented simulation data _run_3_sim+noise_finished/ & _run_3_sim+noise_finished_clean/: Noise-augmented simulations Experimental Results: _run_4_exp_d1_finished/ & _run_4_exp_d1_finished_clean/: Experimental dataset 1 results _run_5_exp_d1_aug_finished/ & _run_5_exp_d1_aug_finished_clean/: Augmented experimental results _run_6_exp_d4_finished/ & _run_6_exp_d4_finished_clean/: Experimental dataset 4 results Each result folder contains: Structure elucidation outputs LLM reasoning traces Performance metrics and evaluation data Intermediate analysis files System Requirements Operating System: Linux/macOS (tested on Ubuntu 20.04+, macOS 12+) Python: 3.7+ (multiple environments required) GPU: CUDA-compatible GPU recommended for model inference Storage: ~10GB for complete dataset Memory: 16GB+ RAM recommended Installation and Usage Download all three zip files from this Zenodo repository Extract to correct locations in your ChemStructLLM installation: # Extract models to root directory cd ChemStructLLM_/ unzip models.zip # Extract data to LLM_Structure_Elucidator directory cd LLM_Structure_Elucidator/ unzip data_molecular_structure.zip unzip intermediate_results_selected.zip Follow setup instructions in the main repository README.md Research Applications This dataset enables researchers to: Reproduce Paper Results: All experimental data and pre-trained models included Benchmark New Methods: Use our molecular dataset for comparative studies Extend the Framework: Build upon our pre-trained models and workflows Analyze LLM Performance: Study reasoning traces and confidence scores across different models Technical Specifications Data Format: JSON for molecular data, PyTorch models for ML components Spectral Data: Industry-standard NMR formats with chemical shift and coupling information Model Architecture: Transformer-based models with multimodal capabilities Evaluation Metrics: Structure recovery accuracy, confidence calibration, reasoning quality Related Links Full Code Repository: https://github.com/mpriessner/ChemStructLLM_ Publication: Priessner, M., et al. (2025). Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs. ChemRxiv Documentation: Complete setup and usage guide available in repository Citation If you use this dataset or models in your research, please cite: @article{priessner2025chemstructllm, title={Enhancing Molecular Structure Elucidation with Reasoning-Capable LLMs}, author={Priessner, Martin and Lewis, R.J. and Johansson, M.J. and Goodman, J.M. and Janet, J.P. and Tomberg, Anna}, journal={ChemRxiv}, year={2025}, doi={[DOI when available]} } Contact Information For questions regarding the dataset, models, or implementation: Primary Contact: [email protected] Corresponding Author: [email protected] License This dataset is released under [specify license - e.g., Creative Commons Attribution 4.0 International License]. Acknowledgments We thank the contributors to the open-source chemical informatics community and the developers of the underlying machine learning frameworks that made this work possible.