Multi-Modal Foundation · mist-models
MIST 1.8B · G298
Predicts molecular Gibbs free energy at 298.15 K from SMILES strings using a 1.8-billion-parameter transformer.
What MIST 1.8B · G298 is
MIST 1.8B · G298 is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It predicts molecular Gibbs free energy at 298.15 K (G298) directly from chemical structure representations. Gibbs free energy is a fundamental thermodynamic quantity that determines the spontaneity and equilibrium of chemical reactions, making it critical for thermodynamic profiling in chemical synthesis, electrolyte evaluation, and drug discovery.
The model is a specialised variant of the Molecular Insight SMILES Transformers (MIST) family. It builds upon a 1.8-billion-parameter foundation model pre-trained on synthetically accessible compounds from the Enamine REAL Space database, and is subsequently fine-tuned on the QM9 dataset from the MoleculeNet benchmark. This enables fast property estimation without requiring costly quantum mechanical electronic structure calculations.
How it works
The model architecture uses a RoBERTa-PreLayerNorm transformer encoder with 24 layers, a hidden dimension of 2304, an intermediate size of 9216, 18 attention heads, and a maximum sequence context of 2048 tokens. Pre-training was conducted on Kekulized SMILES strings from Enamine REAL Space using a masked language modelling objective with cross-entropy loss, trained across 32 NVIDIA A100 GPUs using the FusedLAMB optimizer.
For property prediction, the pre-trained encoder is coupled with a task network consisting of a two-layer multi-layer perceptron with Gaussian Error Linear Unit (GELU) activations and dropout. Chemical sequences are tokenised with the SMIRK tokeniser. Sequence embeddings are pooled by taking the hidden state of the first token, which is passed into the regression head. Fine-tuning on QM9 was conducted on a single NVIDIA A100 GPU using the AdamW optimizer and mean squared error loss to predict G298 in Hartree.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with a 2-layer MLP head
- Parameter count
- 1.8 billion
- Pre-training data
- Enamine REAL Space
- Fine-tuning data
- QM9 (MoleculeNet)
- Context length
- 2048 tokens
- Target property
- Gibbs free energy at 298.15 K (Hartree)
What you can use it for
- Gibbs Free Energy Estimation
- Predicting the standard Gibbs free energy at 298.15 K in Hartree directly from a molecule's Kekulized SMILES string.
- Thermodynamic Screening
- Evaluating reaction feasibility and thermodynamic stability across candidate molecules in early-stage chemical discovery pipelines.
- Molecular Feature Extraction
- Extracting pooled sequence representations from the 1.8-billion-parameter encoder for downstream chemical modelling 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 · G298
Strengths and limits
Good at
- Pre-trained on the Enamine REAL Space collection, reflecting commercially and synthetically accessible chemical space.
- Accommodates long sequence representations with a maximum context window of 2048 tokens.
- Outputs physical thermodynamic energies directly from 1D SMILES strings without requiring 3D quantum chemical conformer generation.
Watch out for
- Expects Kekulized SMILES representations tokenised with the SMIRK tokeniser, which requires a Rust installation.
- Fine-tuned on the QM9 benchmark, limiting empirical regression calibration to the small organic molecule domain of that dataset.
- Demands significant computational resources, requiring approximately 24 GB of GPU memory for deployment.
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 · G298 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 · G298 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 format does the model require?
The model expects molecular sequences as Kekulized SMILES strings, processed using the SMIRK tokeniser.
What are the units of the predicted Gibbs free energy?
The regression head outputs Gibbs free energy at 298.15 K in Hartree.
How are sequence embeddings pooled before prediction?
The model extracts the final hidden state of the first token in the sequence to represent the entire molecule before passing it to the MLP head.
What hardware is necessary to run this model?
The model requires roughly 24 GB of GPU memory for execution.
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. doi:10.48550/arXiv.2510.18900.
Written from MIST 1.8B · G298'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.