Multi-Modal Foundation · mist-models

MIST 1.8B · HOMO

A molecular foundation model predicting highest occupied molecular orbital energy directly from chemical SMILES.

What MIST 1.8B · HOMO is

MIST 1.8B · HOMO is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan. It belongs to the Molecular Insight SMILES Transformers (MIST) family of foundation models, which are designed to support molecular discovery across chemically synthesizable space.

The model addresses the computational cost of quantum chemical calculations by estimating highest occupied molecular orbital (HOMO) energies directly from molecular structure. HOMO energy is a fundamental electronic property that correlates with a molecule's reactivity and ionization potential.

Built by fine-tuning the 1.8-billion-parameter MIST encoder on the QM9 benchmark dataset, the system maps chemical SMILES strings to continuous energy estimates expressed in Hartree units.

How it works

The core architecture is a RoBERTa-PreLayerNorm encoder comprising 24 layers, 24 attention heads, a hidden dimension of 1536, and an intermediate size of 6144, supporting context lengths up to 2048 tokens. The model was initially pre-trained using a masked language modelling objective on SMILES strings from the Enamine REAL Space dataset, optimised with cross-entropy loss and FusedLAMB.

For downstream regression, a task-specific head is attached to the encoder, consisting of a two-layer multi-layer perceptron with Gaussian Error Linear Unit (GELU) activations and dropout. Input representations are pooled by extracting the hidden state of the first token. The model was fine-tuned on QM9 using mean squared error loss and the AdamW optimiser to predict HOMO energy.

Key facts

Encoder architecture: RoBERTa-PreLayerNorm (24 layers, 24 heads, 1536 hidden size)
Task head: Two-layer MLP with GELU activations and dropout
Pre-training objective: Masked Language Modelling (MLM)
Pre-training dataset: Enamine REAL Space SMILES
Fine-tuning dataset: QM9 HOMO Energy (MoleculeNet)
Maximum sequence length: 2048 tokens
Input format: Kekulized SMILES strings
Output format: HOMO energy regression in Hartree
Pre-training hardware: 32 NVIDIA A100-SXM4-80GB GPUs (~7 days)
Fine-tuning hardware: 1 NVIDIA A100 GPU

What you can use it for

HOMO Energy Screening: Predict the highest occupied molecular orbital energy in Hartree for candidate small molecules supplied as SMILES strings.
Reactivity and Ionisation Estimation: Assess relative chemical reactivity and ionization tendencies across molecular libraries without performing density functional theory simulations.
Molecular Representation Extraction: Extract pooled embedding vectors from the 1.8-billion-parameter encoder for downstream computational chemistry workflows.

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 · HOMO

Strengths and limits

Good at

Pre-trained on Enamine REAL Space, anchoring representations in synthetically accessible chemical space.
Directly outputs physically meaningful regression predictions in Hartree units.
Accommodates sequences up to 2048 tokens, enabling evaluation of extended chemical representations.

Watch out for

Requires Kekulized SMILES strings as input, limiting direct applicability to arbitrary non-Kekulized notations.
Fine-tuning is restricted to the QM9 domain, which covers small organic molecules.
Tokenizer dependency requires a Rust installation to use the SMIRK tokenizer.

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 · HOMO 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 · HOMO 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 inputs does the model require?

The model takes molecular SMILES strings, specifically Kekulized SMILES, as input.

What units are the HOMO energy predictions in?

The model outputs regression values for HOMO energy in Hartree.

What base model was fine-tuned for this task?

It was fine-tuned from the pre-trained mist-1.8B-dh61satt checkpoint.

How are token embeddings pooled before the regression head?

Pooling is performed by taking the final hidden state of the first sequence token.

What dataset was used for fine-tuning?

The model was fine-tuned on the QM9 HOMO energy dataset from the MoleculeNet benchmark using an 80/10/10 random train/validation/test split.

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, 2025, arXiv:2510.18900 [physics], doi:10.48550/arXiv.2510.18900

Written from MIST 1.8B · HOMO'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.