Multi-Modal Foundation · mist-models
MIST 1.8B · FreeSolv
A molecular foundation model predicting hydration free energy for neutral molecules from SMILES representations.
What MIST 1.8B · FreeSolv is
MIST 1.8B FreeSolv is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It adapts the Molecular Insight SMILES Transformers (MIST) architecture to perform single-target regression of hydration free energy.
Hydration free energy measures the energetic change when transferring a molecule from gas into water, a critical parameter for evaluating solubility, solvent interactions, and pharmacokinetic behaviour in aqueous environments. The model addresses the computational expense of physical hydration free energy simulations by learning direct mappings from molecular representations to experimental thermodynamic values.
The model is initialised from a 1.8-billion parameter encoder pretrained on vast synthetic chemical space and fine-tuned on the FreeSolv benchmark from MoleculeNet.
How it works
The underlying architecture is a RoBERTa-PreLayerNorm encoder featuring 24 layers, 24 attention heads, a hidden dimension of 1536, and an intermediate dimension of 6144. It is pretrained 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 encoder is paired with a task network consisting of a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Unit (GELU) activations and dropout. Input molecules are supplied as Kekulized SMILES tokenised with SMIRK. Sequence representations are generated by taking the pooled hidden state of the first token, which the task network processes to predict hydration free energy in kcal/mol. Fine-tuning uses the Mean Squared Error (MSE) loss and the AdamW optimizer on the FreeSolv dataset.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with a 2-layer MLP task network
- Pretraining data
- Enamine REAL Space SMILES strings
- Fine-tuning data
- FreeSolv dataset from MoleculeNet (642 neutral molecules)
- Maximum sequence length
- 2048 tokens
- Evaluation metric
- Root Mean Squared Error (RMSE) in kcal/mol
- Required GPU memory
- 24 GB
What you can use it for
- Hydration Free Energy Prediction
- Estimating hydration free energy values in kcal/mol directly from Kekulized SMILES strings for neutral organic molecules.
- Aqueous Solubility Profiling
- Benchmarking molecule-solvent interaction energetics across candidate compounds during early-stage molecular screening.
- Molecular Representation Extraction
- Extracting sequence-level mean-pooled embedding vectors from SMILES strings for downstream computational chemistry 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 1.8B · FreeSolv
Strengths and limits
Good at
- Pretrained on the extensive Enamine REAL Space database covering synthetically accessible chemical regions.
- Utilises a 2048-token context window to accommodate long and complex Kekulized SMILES strings.
- Provides end-to-end hydration free energy predictions directly from string-based chemical inputs without requiring 3D conformer generation.
Watch out for
- Fine-tuned on FreeSolv, a small dataset restricted to only 642 neutral small molecules in water.
- Input tokenisation relies on SMIRK tokenizers, which require a local Rust installation.
- Only evaluates neutral solutes in water and does not predict non-aqueous solvation or charged ionic states.
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 1.8B · 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 1.8B · 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 1× A100 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What input formats does the model accept?
The model takes molecular SMILES strings, specifically formatted as Kekulized SMILES, tokenised using SMIRK.
What are the output units of the regression target?
The model outputs hydration free energy predictions expressed in kcal/mol.
What baseline model was this fine-tuned from?
It was fine-tuned from the pretrained encoder mist-1.8B-dh61satt.
What training infrastructure was used?
Pretraining was executed across 32 NVIDIA A100-SXM4-80GB GPUs in roughly 7 days, whereas fine-tuning was performed on a single NVIDIA A100 GPU.
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. Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 [physics], 2025. doi:10.48550/arXiv.2510.18900.
Written from MIST 1.8B · 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.