Multi-Modal Foundation · mist-models
MIST 1.8B · BBBP
A molecular foundation model predicting blood-brain barrier penetration from SMILES strings.
What MIST 1.8B · BBBP is
MIST 1.8B · BBBP is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It is designed to assess whether small molecules can cross the blood-brain barrier, a key consideration for central nervous system drug discovery and neurological pharmacology.
The model builds upon the MIST (Molecular Insight SMILES Transformers) family of foundation models. After large-scale pre-training across commercially accessible chemical space, the 1.8-billion-parameter architecture was fine-tuned on the MoleculeNet Blood-Brain Barrier Penetration (BBBP) benchmark for binary classification.
How it works
The model architecture uses a RoBERTa-PreLayerNorm encoder consisting of 24 layers, 18 attention heads, a hidden size of 2304, and an intermediate size of 9216, supporting sequences up to 2048 tokens. Molecules are represented as Kekulized SMILES strings. During pre-training on the Enamine REAL Space dataset, the encoder learned representations via a masked language modelling objective using Cross-Entropy Loss and the FusedLAMB optimizer.
For downstream prediction, the pre-trained encoder (mist-1.8B-dh61satt) is paired with a task network comprising a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Units (GELU) activations and dropout. Hidden states are pooled by taking the representation of the first token. The task network was fine-tuned with binary cross-entropy loss and the AdamW optimizer on the BBBP dataset using an 80/10/10 scaffold split, outputting unnormalized logits for binary penetration classification.
Key facts
- Model architecture
- RoBERTa-PreLayerNorm encoder with a two-layer MLP task head
- Base model
- mist-1.8B-dh61satt
- Pre-training objective
- Masked Language Modelling (MLM)
- Pre-training data
- Enamine REAL Space SMILES
- Fine-tuning data
- MoleculeNet BBBP dataset (2,039 compounds)
- Maximum sequence length
- 2048 tokens
- Evaluation metric
- AUROC
What you can use it for
- Blood-Brain Barrier Penetration Screening
- Predicting whether small-molecule drug candidates can penetrate the central nervous system based on their Kekulized SMILES strings.
- Molecular Representation Extraction
- Extracting pooled sequence embeddings from the fine-tuned encoder for use in downstream cheminformatics analyses.
- Lead Prioritisation for Neurotherapeutics
- Ranking candidate compounds during early-stage discovery to filter for or against blood-brain barrier permeability.
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 · BBBP
Strengths and limits
Good at
- Pre-trained on Enamine REAL Space, capturing synthetically accessible and drug-relevant regions of chemical space.
- Evaluated using a scaffold split to test generalisation across distinct chemical classes rather than random train-test partitions.
- Accommodates molecular inputs up to 2048 tokens using an 18-head RoBERTa-PreLayerNorm backbone.
Watch out for
- Fine-tuned on a relatively small single-task dataset of 2,039 compounds, which limits exposure to broader bioactivity profiles.
- Expects input molecules strictly as Kekulized SMILES strings tokenised using the SMIRK tokenizer, which requires a Rust environment.
- Outputs unnormalized logits rather than calibrated probabilities for blood-brain barrier crossing.
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 · 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 1.8B · 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 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 takes SMILES strings of molecules, which should be Kekulized before processing by the SmirkTokenizerFast tokenizer.
What is the model's output?
When using its prediction pipeline, the model outputs unnormalized logits for binary classification indicating whether a compound penetrates the blood-brain barrier. It can also produce mean-pooled embedding vectors when run as a feature-extraction model.
What dataset was used to fine-tune this model?
The model was fine-tuned on the MoleculeNet BBBP (Blood-Brain Barrier Penetration) dataset, which contains 2,039 annotated compounds split 80/10/10 using scaffold partitioning.
What hardware was used to train the model?
Pre-training took approximately 7 days across 32 NVIDIA A100-SXM4-80GB GPUs, while fine-tuning on the BBBP dataset was completed on a single NVIDIA A100 GPU.
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 1.8B · 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.