Multi-Modal Foundation · mist-models
MIST 28M · ESOL
A 28M-parameter transformer predicting aqueous solubility (log S) of small molecules from SMILES strings.
What MIST 28M · ESOL is
MIST 28M · ESOL is a molecular foundation model specialised for aqueous solubility regression. Developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor, it adapts the MIST-28M architecture to predict measured log solubility (log S in mols per litre) for small molecules.
Solubility in water is a critical physical property across drug discovery, electrolyte formulation, and chemical design, determining whether a compound will dissolve sufficiently to function in biological or chemical systems. This model replaces experimental solubility screening by predicting log S values directly from chemical structure representations.
The model builds on the MIST (Molecular Insight SMILES Transformers) family, which uses self-supervised learning over extensive collections of commercially available, synthetically accessible molecules to capture underlying chemical patterns before supervised fine-tuning.
How it works
The underlying architecture is an 8-layer RoBERTa-PreLayerNorm encoder with 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a maximum sequence length of 2048 tokens. It was pre-trained using a masked language modelling (MLM) objective with cross-entropy loss and the FusedLAMB optimizer on SMILES strings from the Enamine REAL Space dataset.
For property prediction, the pre-trained encoder (mist-28M-ti624ev1) is paired with a task network consisting of a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Units (GELU) activations and dropout. Sequence embeddings are derived by taking the hidden state of the first token. The full model was fine-tuned on the Delaney ESOL dataset using mean squared error (MSE) loss and the AdamW optimizer.
Key facts
- Base architecture
- RoBERTa-PreLayerNorm (8 layers, 8 attention heads, hidden dimension 512)
- Parameter count
- 28M
- Context length
- 2048 tokens
- Input format
- Kekulized SMILES strings
- Output format
- Aqueous solubility (log S in mols per litre) or mean-pooled embedding vectors
- Pre-training data
- Enamine REAL Space
- Fine-tuning data
- MoleculeNet ESOL dataset (1,128 small molecules)
What you can use it for
- Aqueous Solubility Prediction
- Predicting log S values (in mols per litre) for drug candidates, electrolyte components, and organic molecules from their SMILES strings.
- High-Throughput Virtual Screening
- Prioritising chemical libraries by filtering out poorly soluble molecules early in discovery pipelines.
- Molecular Representation Extraction
- Generating embedding vectors from the encoder's pooled token states for downstream chemoinformatics workflows and property regression models.
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 · ESOL
Strengths and limits
Good at
- Pre-trained on Enamine REAL Space, providing foundational familiarity with synthetically accessible regions of chemical space.
- Compact 28-million parameter encoder architecture enables efficient inference on moderate hardware.
- Supports long sequence contexts up to 2048 tokens, handling complex molecular SMILES representations.
Watch out for
- Requires input molecules to be provided as Kekulized SMILES strings for expected behaviour.
- Fine-tuning was conducted on the Delaney ESOL dataset of only 1,128 compounds, which may limit generalisation to distant chemical spaces.
- Environment setup requires a Rust installation to support 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 28M · ESOL 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 · ESOL 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 units are used for the solubility predictions?
The model outputs log solubility (log S) measured in mols per litre.
What input format does the model expect?
The model expects molecular SMILES strings, specifically Kekulized SMILES strings.
How are molecular representations pooled?
In line with standard encoder pooling, the embedding vector is produced by extracting the hidden state of the first token.
What hardware is needed to run or fine-tune this model?
The pre-trained model was originally pre-trained on two NVIDIA A100-SXM4-80GB GPUs and fine-tuned on a single NVIDIA A100 GPU, with registry environments allocating 16 GB of GPU memory for inference.
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 · ESOL'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.