MIST Demo

This repository contains tutorials for fine-tuning and applying MIST (Molecular Insight SMILES Transformer) foundation models to chemical problems. Model checkpoints for MIST models are available on HuggingFace and on Zenodo. The full code, including pre-training, model development and full scale application demos can be found in the mist repository.

Tutorials

Molecular Property Prediction

Inference demonstrations using fine-tuned MIST models:

• Loading pretrained MIST checkpoints from HuggingFace
• Predicting boiling point, flash point, and melting point
• Analyzing property trends for alkenes and alcohols

Run Finetuning

Complete fine-tuning workflow for MIST encoder models:

• Finetuning with LoRA (Low-Rank Adaptation) for parameter-efficient training
• Hyperparameter optimization for task network
• Training on the QM9 dataset for molecular property prediction
• Model evaluation

Olfaction Prediction

Exploring how molecules smell with the MIST olfaction model:

• Generating the scent profile for a molecule
• Modeling discontinuities in the structure-perception mapping

Installation

Local Installation

1. Clone the repository:

git clone <repository-url>
cd mist-demo

2. Create a virtual environment and install dependencies using uv

uv sync
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

Running the Notebooks

Launch Jupyter and open any notebook in mist-demo/tutorials:

jupyter notebook

Cite

If you use the MIST models in your work please cite:

@online{MIST,
  title = {Foundation Models for Discovery and Exploration in Chemical Space},
  author = {Wadell, Alexius and Bhutani, Anoushka and Azumah, Victor and Ellis-Mohr, Austin R. and Kelly, Celia and Zhao, Hancheng and Nayak, Anuj K. and Hegazy, Kareem and Brace, Alexander and Lin, Hongyi and Emani, Murali and Vishwanath, Venkatram and Gering, Kevin and Alkan, Melisa and Gibbs, Tom and Wells, Jack and Varshney, Lav R. and Ramsundar, Bharath and Duraisamy, Karthik and Mahoney, Michael W. and Ramanathan, Arvind and Viswanathan, Venkatasubramanian},
  date = {2025-10-20},
  eprint = {2510.18900},
  eprinttype = {arXiv},
  eprintclass = {physics},
  doi = {10.48550/arXiv.2510.18900},
  url = {http://arxiv.org/abs/2510.18900},