Multi-Modal Foundation · mist-models
MIST 28M · MUV
Predicts 17 bioassay activity endpoints from SMILES strings for unbiased virtual screening.
What MIST 28M · MUV is
MIST 28M · MUV is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It belongs to the Molecular Insight SMILES Transformers (MIST) family of chemical foundation models, which are designed to learn chemical representations directly from simplified molecular-input line-entry system (SMILES) text.
The model is fine-tuned on the Maximum Unbiased Validation (MUV) benchmark from MoleculeNet. Comprising 93,087 chemical compounds evaluated across 17 bioassay targets, MUV was designed to eliminate analogue bias and provide a realistic evaluation framework for virtual screening pipelines.
How it works
The underlying architecture is a RoBERTa-PreLayerNorm encoder with 8 layers, 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a context window of 2048 tokens. For property prediction, a two-layer multi-layer perceptron with Gaussian Error Linear Unit (GELU) activations and dropout is appended to the encoder. Token embeddings are pooled into a single molecular representation by selecting the hidden state of the initial sequence token.
The base encoder was pre-trained on Kekulized SMILES strings from Enamine REAL Space using masked language modelling with cross-entropy loss and the FusedLAMB optimizer. The model was subsequently fine-tuned on Kekulized SMILES from the MUV dataset using binary cross-entropy loss and the AdamW optimizer to perform multi-label binary classification.
Key facts
- Model architecture
- RoBERTa-PreLayerNorm with a 2-layer MLP classification head
- Parameter count
- 28 million
- Context length
- 2,048 tokens
- Pre-training dataset
- Enamine REAL Space
- Fine-tuning dataset
- MoleculeNet MUV (93,087 compounds)
- Input format
- Kekulized SMILES strings
- Primary outputs
- Unnormalised logits for 17 bioassay activity endpoints
What you can use it for
- Unbiased virtual screening
- Screens candidate compound libraries against 17 challenging bioassay targets designed to counteract analogue bias.
- Multi-target bioactivity prediction
- Outputs simultaneous classification logits indicating potential activity across 17 distinct biological targets.
- Molecular feature extraction
- Produces pooled continuous vector embeddings from input chemical sequences for downstream 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 28M · MUV
Strengths and limits
Good at
- Trained on the MUV dataset, mitigating chemical analogue bias during virtual screening evaluation.
- Pre-trained on Enamine REAL Space, anchoring representations within synthetically accessible regions of chemical space.
- Compact 28-million parameter footprint enables efficient inference and lower computational requirements.
Watch out for
- Predictions are restricted to the 17 predefined bioassay activity targets of the MUV benchmark.
- Requires Kekulized SMILES as input and relies on the SMIRK tokenizer, which requires Rust to run.
- Performance is evaluated on an 80/10/10 random split, which may not assess cross-scaffold generalisation.
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 28M · MUV 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 28M · MUV 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 inputs does the model expect?
The model expects molecular structures formatted as Kekulized SMILES strings, processed using the SMIRK tokenisation library.
What predictions does the fine-tuned model generate?
The model generates unnormalised logits for multi-label binary classification across 17 bioassay endpoints from the MUV benchmark.
Which checkpoint was used for fine-tuning?
The model was fine-tuned from the pre-trained checkpoint mist-28M-ti624ev1.
What pooling strategy is applied to the sequence representations?
For property prediction, the encoder representations are pooled by extracting the hidden state vector of the first token in the sequence.
How to cite it
Alexius Wadell, Anoushka Bhutani, Victor Azumah, Austin R. Ellis-Mohr, Celia Kelly, Hancheng Zhao, Anuj K. Nayak, Kareem Hegazy, Alexander Brace, Hongyi Lin, Murali Emani, Venkatram Vishwanath, Kevin Gering, Melisa Alkan, Tom Gibbs, Jack Wells, Lav R. Varshney, Bharath Ramsundar, Karthik Duraisamy, Michael W. Mahoney, Arvind Ramanathan, and Venkatasubramanian Viswanathan. Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 [physics], 2025. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 28M · MUV'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.