Multi-Modal Foundation · mist-models
MIST 28M · FreeSolv
A 28-million-parameter transformer predicting hydration free energies of small neutral molecules in water.
What MIST 28M · FreeSolv is
MIST 28M · FreeSolv is a specialised molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It adapts a 28-million-parameter transformer encoder to estimate the hydration free energy of small organic compounds directly from chemical structure representations.
Hydration free energy measures the free energy change when transferring a molecule from vacuum into water, serving as a fundamental quantity for understanding aqueous solubility, solvation thermodynamics, and molecule–solvent interactions. The model outputs quantitative regression predictions in kilocalories per mole (kcal/mol).
How it works
The underlying architecture is an 8-layer RoBERTa-PreLayerNorm encoder with 8 attention heads, a hidden dimension of 512, an intermediate dimension of 2048, and a maximum context length of 2048 tokens. For property regression, the encoder is paired with a task-specific network consisting of a two-layer multi-layer perceptron with Gaussian Error Linear Unit (GELU) activations and dropout. Molecular embeddings are pooled by extracting the hidden state vector of the first token.
The base model was pre-trained on Kekulized SMILES strings from the Enamine REAL Space dataset using a masked language modelling objective with cross-entropy loss and the FusedLAMB optimiser. It was subsequently fine-tuned on the FreeSolv benchmark from MoleculeNet using mean squared error loss and the AdamW optimiser to predict experimental and calculated hydration free energies.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder (8 layers) with a 2-layer MLP head
- Parameter count
- 28 million
- Pre-training data
- Enamine REAL Space (Kekulized SMILES)
- Fine-tuning data
- FreeSolv (642 neutral molecules)
- Input format
- Kekulized SMILES strings
- Context length
- 2048 tokens
- Prediction target
- Hydration free energy (kcal/mol)
What you can use it for
- Hydration Free Energy Prediction
- Estimate the aqueous hydration free energy of candidate small molecules directly from molecular SMILES notation.
- Aqueous Interaction Screening
- Evaluate molecule–solvent interaction energies when prioritising neutral compounds for aqueous chemical environments.
- Molecular Representation Extraction
- Extract mean-pooled or first-token embedding vectors from the encoder for downstream chemical analysis tasks.
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 · FreeSolv
Strengths and limits
Good at
- Pre-trained on the Enamine REAL Space chemical library, providing strong structural representations across synthetically accessible chemical space.
- Handles extended molecular sequences with support for inputs up to 2048 tokens.
- Outputs interpretable thermodynamic values directly in kcal/mol.
Watch out for
- Fine-tuned on the FreeSolv benchmark containing only 642 molecules, restricting empirical exposure to a small set of neutral compounds in water.
- Does not support charged or ionic species, as training targets are defined solely for neutral molecules.
- Requires Rust in the local environment to support SMIRK tokenisation.
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 · FreeSolv 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 · FreeSolv 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 require?
The model expects molecular structures in the form of Kekulized SMILES strings up to 2048 tokens long.
What are the units of the predicted outputs?
Regression predictions are returned in kilocalories per mole (kcal/mol).
How was the model evaluated?
The model was evaluated on an 80/10/10 random split of the FreeSolv dataset, tracking performance using Root Mean Squared Error (RMSE) in kcal/mol.
Can the model be used solely as an embedding extractor?
Yes, it can function as a feature-extraction model by taking the hidden representation of the first token to produce a single embedding vector per molecule.
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].
Written from MIST 28M · FreeSolv'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.