Multi-Modal Foundation · mist-models
MIST 28M · tmQM
A molecular transformer predicting quantum-mechanical properties of transition-metal complexes from SMILES strings.
What MIST 28M · tmQM is
MIST 28M · tmQM is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It is fine-tuned from MIST-28M, an encoder in the Molecular Insight SMILES Transformers family designed for downstream chemical tasks.
The model addresses the computational expense of quantum-chemical calculations for organometallic systems. By processing SMILES representations of transition-metal complexes, it provides fast regression predictions for quantum-mechanical targets originally derived via density functional theory (DFT).
How it works
The architecture consists of a RoBERTa-PreLayerNorm encoder with 8 layers, 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a maximum sequence context of 2048 tokens. The base encoder was pre-trained on SMILES strings from the Enamine REAL Space dataset using masked language modelling (MLM) and cross-entropy loss with the FusedLAMB optimizer.
For downstream prediction, the encoder is paired with a task network consisting of a two-layer multi-layer perceptron (MLP) featuring Gaussian Error Linear Unit (GELU) activations and dropout. Hidden states are pooled from the sequence's first token. The network was fine-tuned on the tmQM dataset across 8 quantum mechanical properties using mean squared error (MSE) loss and the AdamW optimizer.
Key facts
- Base architecture
- RoBERTa-PreLayerNorm (8 layers, 8 attention heads, hidden dimension 512)
- Task architecture
- Two-layer MLP with GELU activations and dropout
- Pre-training objective
- Masked Language Modelling (MLM)
- Fine-tuning objective
- Multi-task regression (MSE loss)
- Pre-training data
- Enamine REAL Space SMILES
- Fine-tuning data
- tmQM (108,000 transition organometallic compounds)
- Context length
- 2048 tokens
- Primary input format
- Kekulized SMILES strings
What you can use it for
- Quantum property prediction for organometallics
- Predict 8 quantum-mechanical targets, including HOMO, LUMO, HOMO/LUMO gap, dipole moment, and natural charge, directly from transition-metal complex SMILES.
- Molecular representation extraction
- Generate pooled sequence embedding vectors from transition-metal compound SMILES for downstream analytical workflows or feature extraction.
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 · tmQM
Strengths and limits
Good at
- Trained on tmQM, granting coverage over broad elemental and stereochemical diversity in transition organometallic chemistry.
- Pre-trained on Enamine REAL Space, anchoring chemical representations in synthetically accessible chemical space.
- Evaluated across 8 quantum mechanical properties using an 80/10/10 random train-validation-test split.
Watch out for
- Requires Kekulized SMILES strings as inputs unless otherwise specified.
- Tokenization relies on SMIRK tokenizers, which require a Rust installation to run.
- Served pipelines configured strictly for feature extraction return mean-pooled vectors rather than multi-property regression outputs.
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 · tmQM 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 · tmQM 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 MIST 28M · tmQM take?
The model accepts molecular SMILES strings, specifically expecting Kekulized SMILES format.
What properties does the fine-tuned model predict?
The task network predicts 8 quantum-mechanical properties, including HOMO, LUMO, HOMO/LUMO gap, dipole moment, and natural charge.
What quantum level of theory was used for the fine-tuning data?
The tmQM dataset contains quantum properties computed using DFT at the TPSSh-D3BJ/def2-SVP level with geometries pre-optimized at the GFN2-xTB level.
How does the model pool sequence representations?
The encoder hidden states are pooled using the hidden state of the first token before being passed to the task regression network, though generic extraction pipelines may provide mean-pooled embeddings.
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. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 28M · tmQM'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.