MolPuzzle: Spectrum Analysis using LLMs

Overview

MolPuzzle is a comprehensive framework for elucidating molecular structures from spectral data (IR, Mass Spectrometry, H-NMR, C-NMR) using Large Language Models (LLMs). It automates the generation of questions based on spectral features, samples data for evaluation, and benchmarks various models including GPT-4, Claude, and open-source VLM models.

Features

• Multi-Modal Analysis: Supports IR, Mass Spec, H-NMR, and C-NMR spectrum analysis.
• Automated Question Generation: Converts spectral data into natural language QA pairs.
• Model Evaluation: Benchmarks LLM performance on chemical structure elucidation tasks.
• Extensible Architecture: Easy to add new models and datasets.

Structure

• src/molpuzzle/: Core package containing the analysis logic.
• notebooks/: Jupyter notebooks for data conversion and experimental workflows.
• Data/: Dataset containing JSON and image files for spectrum analysis.
• sample_data/: Directory for generated sample data.

Installation

Prerequisites

• Python 3.8+
• PyTorch
• CUDA (optional, for local model inference)

Steps

1. Clone the repository:

   git clone https://github.com/yourusername/MolPuzzle.git
   cd MolPuzzle
   

2. Install dependencies:

   pip install -r requirements.txt
   

   Or install the package in editable mode:

   pip install -e .
   

Quick Start

We provide a quick start script to demonstrate how to load data and perform sampling.

python quick_start.py

This script will:

1. Load sample data from Data/Stage1.json.
2. Convert it to a CSV format suitable for processing.
3. perform a data sampling session to generate a test set.

Usage

1. Data Sampling

To sample questions from a dataset:

python src/molpuzzle/spectrum_analysis.py \
    --task H-NMR \
    --action sample_data \
    --input_csv sample_data/stage1_sample.csv \
    --output_csv sample_data/sampled_questions.csv

2. Generate Responses

To generate responses using a model (e.g., GPT-4):

export OPENAI_API_KEY='your_key'
python src/molpuzzle/spectrum_analysis.py \
    --task H-NMR \
    --action generate_responses \
    --models gpt-4 \
    --input_csv sample_data/sampled_questions_0.csv

3. Evaluation

To evaluate model performance:

python src/molpuzzle/spectrum_analysis.py \
    --task H-NMR \
    --action evaluate \
    --models gpt-4 \
    --input_csv sample_data/sampled_questions_0.csv

Data Sample

The Data/ directory contains the core datasets.

• Stage1.json: Basic property questions (Saturation, etc.)
• Stage2.json: Functional group identification.
• Stage3.json: Full structure elucidation.

Example entry from Stage 1:

{
    "Molecule Index": "99",
    "SMILES": "CCCCC1=CC=CC=C1",
    "cls": "Saturation",
    "Formula": "C10H14",
    "Question": "Could the molecule with the formula C10H14 potentially be Saturated?",
    "Answer": "No"
}

License

This project is licensed under the MIT License - see the LICENSE file for details.