Multi-Modal Foundation · mist-models
MIST 26.9M · dn
A molecular language model fine-tuned to predict Lewis basicity and boron trifluoride affinity from SMILES strings.
What MIST 26.9M · dn is
MIST 26.9M · dn is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. Belonging to the Molecular Insight SMILES Transformers (MIST) family, it is designed to predict molecular chemical properties directly from simplified molecular-input line-entry system (SMILES) representations.
The model addresses downstream property prediction tasks, specifically targeting Lewis basicity. Fine-tuned on an experimental boron trifluoride (BF₃) affinity benchmark, the network provides numerical property regressions useful for understanding molecular interactions and solvation dynamics.
This capability is particularly relevant to researchers designing advanced electrolytes, where Lewis basicity influences lithium-ion salt solubility and the coordination environment in battery systems.
How it works
The architecture uses a RoBERTa-PreLayerNorm encoder with 8 layers, a hidden dimension of 512, an intermediate size of 2048, and 8 attention heads, capped at a maximum sequence length of 2048 tokens. For downstream property prediction, the encoder connects to a task network consisting of a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Unit (GELU) activations and dropout. The representation is formed by pooling the hidden state of the first sequence token.
The encoder was originally pre-trained on Kekulized SMILES strings from the Enamine REAL Space dataset using a masked language modelling (MLM) cross-entropy objective optimised by FusedLAMB. Fine-tuning uses the AdamW optimiser and a mean squared error (MSE) loss function over an experimental dataset of 344 BF₃ affinity measurements of Lewis bases taken at 298 K and 1 atm.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder (8 layers, 8 heads, hidden size 512) with a 2-layer MLP head
- Base model
- mist-28M-ti624ev1
- Pre-training data
- Enamine REAL Space SMILES strings
- Fine-tuning data
- 344 Lewis base BF₃ affinity measurements at 298 K and 1 atm
- Input format
- Kekulized SMILES strings
- Maximum context length
- 2048 tokens
What you can use it for
- Lewis basicity quantification
- Predict numerical BF₃ affinity values directly from candidate SMILES strings to evaluate molecular Lewis basicity.
- Battery electrolyte screening
- Assess candidate solvent and additive molecules for lithium-ion salt solubility and solvation shell behaviours.
- Molecular feature extraction
- Extract mean-pooled numerical embedding vectors from molecular sequences for downstream clustering or property analysis.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
MIST 26.9M · dn
Strengths and limits
Good at
- Pre-trained on Enamine REAL Space, capturing synthetically accessible chemical space.
- Supports long sequence contexts up to 2048 tokens for complex molecular SMILES representations.
- Enables rapid scalar prediction directly from Kekulized SMILES without requiring 3D conformer generation.
Watch out for
- Fine-tuning relied on a small dataset of only 344 experimental measurements.
- Predictions are calibrated specifically to standard conditions of 298 K and 1 atm.
- Tokenizer dependencies require a Rust environment when using SMIRK tokenizers.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use MIST 26.9M · dn commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run MIST 26.9M · dn without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on CPU only, typically in 12s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
What does the model take as input?
It accepts molecular SMILES strings, specifically pre-processed into Kekulized form.
What physical property does the model output?
It outputs a regression prediction for BF₃ affinity, an experimental measure of Lewis basicity.
Which base model was this variant built upon?
It was fine-tuned from the mist-28M-ti624ev1 pre-trained RoBERTa-style checkpoint.
What metrics are used to assess the model's accuracy?
Performance is measured using Root Mean Squared Error (RMSE) across a random 80/10/10 data split.
How to cite it
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 (2025). Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 [physics]. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 26.9M · dn's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.