Multi-Modal Foundation · mist-models
MIST 28M · pKa
Predicts acid dissociation constants (pKa) in DMSO for small molecules from Kekulized SMILES strings.
What MIST 28M · pKa is
MIST 28M · pKa is a specialised molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It adapts the 28-million-parameter MIST (Molecular Insight SMILES Transformers) foundation model to predict acid dissociation constants (pKa) in dimethyl sulfoxide (DMSO).
Understanding acid-base equilibria and ionisation behaviour in non-aqueous environments is vital in areas such as electrolyte development and physical organic chemistry. The model maps chemical structures directly to numerical dissociation values for dipolar non-hydrogen-bond-donor solvent systems.
How it works
The architecture consists of a RoBERTa-PreLayerNorm encoder with 8 transformer layers, 8 attention heads, a hidden size of 512, and an intermediate dimension of 2048, supporting sequences up to 2048 tokens. The base encoder was pre-trained using a masked language modelling objective on Kekulized SMILES strings from the Enamine REAL Space database.
For pKa prediction, the encoder is paired with a task network comprising a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Unit (GELU) activations and dropout. The hidden state of the first sequence token is pooled and fed into the task head. Fine-tuning was carried out using mean squared error loss on experimental pKa measurements extracted from an IUPAC technical report on dipolar non-hydrogen-bond-donor solvents, with duplicate entries averaged by InChI key.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with a 2-layer MLP regression head
- Parameters
- 28M
- Maximum sequence length
- 2048 tokens
- Input format
- Kekulized SMILES strings
- Output format
- Scalar pKa prediction (or pooled sequence embedding vectors)
- Pre-training dataset
- Enamine REAL Space
- Fine-tuning dataset
- IUPAC Technical Report on Acid Dissociation Constants in Selected Dipolar Non-Hydrogen-Bond-Donor Solvents (DMSO subset)
What you can use it for
- DMSO pKa Prediction
- Estimating acid dissociation constants in dimethyl sulfoxide for single organic molecules from SMILES representations.
- Non-Aqueous Electrolyte Screening
- Evaluating the acidity and ionisation tendencies of candidate electrolyte additives and organic salts in dipolar non-hydrogen-bond-donor solvents.
- Molecular Feature Extraction
- Generating first-token pooled molecular embeddings from the trained encoder to feed into alternative chemical learning 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 · pKa
Strengths and limits
Good at
- Fine-tuned on curated IUPAC experimental data where duplicate measurements were systematically resolved by InChI key averaging.
- Initialised from pre-training on the synthetically accessible Enamine REAL Space chemical collection.
- Accommodates molecular sequence lengths up to 2048 tokens.
Watch out for
- Predictions are restricted to dissociation behaviour in DMSO and cannot be applied directly to aqueous systems or protic solvents.
- Inputs must be provided as Kekulized SMILES strings.
- Cannot evaluate chemical mixtures, as multi-component records were excluded during training dataset preparation.
- Requires Rust toolchains to be installed to support the underlying SMIRK tokeniser.
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 · pKa 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 · pKa 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 input format does the model expect?
The model takes molecular SMILES strings, specifically requiring Kekulized SMILES, processed with the SMIRK tokeniser.
Which solvent does this pKa predictor model?
The model specifically predicts pKa in dimethyl sulfoxide (DMSO), trained on records filtered from IUPAC tables of dipolar non-hydrogen-bond-donor solvents.
How does the model aggregate token representations across a SMILES string?
Consistent with standard sequence encoder classification approaches, the task head pools representations by extracting the hidden state of the first sequence token.
Can this model predict the pKa of multicomponent mixtures?
No, mixture measurements were specifically removed during training data curation, limiting predictions to pure single compounds.
How to cite it
Wadell, Alexius, Anoushka Bhutani, Victor Azumah, Austin R. Ellis-Mohr, Celia Kelly, Hancheng Zhao, Anuj K. Nayak, et al. 'Foundation Models for Discovery and Exploration in Chemical Space.' arXiv preprint arXiv:2510.18900 (2025). https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 28M · pKa'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.