MIST: Molecular Insight SMILES Transformers

MIST is a family of molecular foundation models for molecular property prediction. The models were pre-trained on SMILES strings from the Enamine REAL Space dataset using the Masked Language Modeling (MLM) objective, then fine-tuned for downstream prediction tasks.

Model Details

• Architecture: Encoder-only transformer RoBERTa-PreLayerNorm
• Pre-training: Masked Language Modeling on molecular SMILES
• Tokenization: Smirk tokenizer

Model Inputs and Outputs

Inputs

• SMILES strings: Standard SMILES notation for molecular structures
• Batch size: Variable, automatically padded during inference

Outputs

• Predictions: Task-specific numerical or categorical predictions
• Format: Dictionary with channel names and predicted values (if channels are configured), or raw tensor output

Quick Start

Tutorials are available in Google Colab:

• Inference
• Finetuning

Running Locally

To run the model locally, create a virtual environment and install dependencies:

python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
pip install -r requirements.txt

Note: SMIRK tokenizers require Rust to be installed. See the Rust installation guide for details.

Use the model! For a full list of model IDs and properties see the list of provided models below. For details on the specific inputs and outputs formats for each model variant see the model card.

from transformers import AutoModel
from smirk import SmirkTokenizerFast

# Load the model
model = AutoModel.from_pretrained(
    "mist-models/mist-{size}-{model_id}-{property}",
    trust_remote_code=True
)

# Make predictions
smiles_batch = [
    "CCO",           # Ethanol
    "CC(=O)O",       # Acetic acid
    "C1=CC=CC=C1"       # Benzene
]
results = model.predict(smiles_batch)

Provided Models

Pre-trained

• mist-1.8B-dh61satt: Flagship MIST model (MIST-1.8B)
• mist-28M-ti624ev1 **: Smaller MIST model (MIST-28M).

** Indicates publically released models. Below is a full list of finetuned variants hosted on HuggingFace:

MoleculeNet Benchmark Models

** Indicates publically released models. * Indicates models currently not available on hugging-face due to storage limits.

QM9 Benchmark Models

The single target (MIST-1.8B encoder) models for properties in QM9 are available.

* Indicates models currently not available on hugging-face due to storage limits

Finetuned Single Task Models

These models consist of a MIST-encoder and task network finetuned on a single dataset used in the applications demonstrated in the manuscript.

Finetuned Multi-Task Models

These are additional multi-target finetuned models consisting of a MIST encoder and task network.

** Indicates publically released models.

Finetuned Mixture Models

These models consist of a MIST-encoder and physics informed task network for mixture property prediction.

Citation

If you use this model in your research, 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},  
}

License and Notice

Model weights are provided as-is for research purposes only, without guarantees of correctness, fitness for purpose, or warranties of any kind.

Restrictions:

• Research use only
• No redistribution without permission
• No commercial use without licensing agreement

For questions, issues, or licensing inquiries, please contact venkvis@umich.edu.