Multi-Modal Foundation · mist-models
MIST 28M · BBBP
A molecular transformer predicting blood-brain barrier permeability from SMILES strings for CNS drug screening.
What MIST 28M · BBBP is
MIST 28M · BBBP is a molecular property prediction model designed to classify whether small molecules can cross the blood-brain barrier. Developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor, the model is built on the Molecular Insight SMILES Transformers (MIST) framework.
Assessing blood-brain barrier permeability is a critical filter in the development of therapeutics targeting the central nervous system. The model addresses this by training on the MoleculeNet BBBP benchmark, which comprises 2,039 compounds annotated as either penetrating or non-penetrating.
The underlying foundation model was pre-trained on vast chemical structures from the Enamine REAL Space database using masked language modelling, before being fine-tuned with a classification head on scaffold-split molecular data.
How it works
The architecture relies on a RoBERTa-PreLayerNorm encoder featuring 8 layers, 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a maximum sequence context of 2048 tokens. For downstream property prediction, sequence representations are pooled using the hidden state of the first token and fed into a task network consisting of a two-layer multi-layer perceptron (MLP) with GELU activations and dropout.
Pre-training was conducted on Kekulized SMILES strings using a masked language modelling objective with cross-entropy loss and the FusedLAMB optimizer. Supervised fine-tuning replaces the language modelling objective with binary cross-entropy loss optimised via AdamW, producing unnormalised logits for binary penetration classification.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with two-layer MLP task head
- Parameter count
- 28 million
- Encoder dimensions
- 8 layers, 8 attention heads, hidden size 512, intermediate size 2048
- Maximum sequence length
- 2048 tokens
- Pre-training data
- Enamine REAL Space SMILES strings
- Fine-tuning data
- MoleculeNet BBBP (2,039 compounds)
- Evaluation metric
- AUROC
What you can use it for
- Central nervous system drug screening
- Predicting whether candidate small-molecule therapeutics can permeate the blood-brain barrier to reach central nervous system targets.
- Virtual chemical library profiling
- Filtering and categorising large sets of Kekulized SMILES representations into penetrating and non-penetrating compounds.
- Molecular representation extraction
- Extracting pooled token embeddings of molecular structures for downstream statistical modelling and chemical analysis.
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 · BBBP
Strengths and limits
Good at
- Pre-trained on the Enamine REAL Space dataset, exposing the encoder to synthetically accessible chemical space prior to task adaptation.
- Benchmarked on a rigorous 80/10/10 scaffold split to test generalisation across structurally diverse chemical scaffolds.
- Lightweight 28-million parameter design allows for low-overhead inference compared to larger molecular language models.
Watch out for
- Provides only binary classification and does not quantify continuous rates or mechanisms of blood-brain barrier transport.
- Supervised fine-tuning is restricted to a small dataset of 2,039 compounds from MoleculeNet.
- Input sequences must be Kekulized SMILES strings rather than arbitrary chemical string formats.
- Requires an environment with Rust installed to support the associated 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 28M · BBBP 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 · BBBP 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 take?
The model expects Kekulized SMILES strings representing chemical structures, tokenized with the custom SMIRK fast tokenizer.
What are the outputs of the model?
The model outputs unnormalised logits for binary classification regarding blood-brain barrier penetration, or pooled embedding vectors when run as a feature extractor.
How was the model evaluated during development?
The model was evaluated on the MoleculeNet BBBP dataset using an 80/10/10 scaffold split, with performance assessed by the Area Under the Receiver Operating Characteristic curve (AUROC).
What computational resources were used to train this model?
Pre-training was conducted on two NVIDIA A100-SXM4-80GB GPUs over 12 hours and 15 minutes, while fine-tuning was performed using a single NVIDIA A100 GPU.
How to cite it
Wadell, A., Bhutani, A., Azumah, V., Ellis-Mohr, A. R., Kelly, C., Zhao, H., Nayak, A. K., Hegazy, K., Brace, A., Lin, H., Emani, M., Vishwanath, V., Gering, K., Alkan, M., Gibbs, T., Wells, J., Varshney, L. R., Ramsundar, B., Duraisamy, K., Mahoney, M. W., Ramanathan, A., and Viswanathan, V. (2025). Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 28M · BBBP'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.