Multi-Modal Foundation · mist-models
MIST 1.8B · BACE
A 1.8-billion-parameter model predicting small-molecule inhibition of Alzheimer's target BACE-1 from SMILES strings.
What MIST 1.8B · BACE is
MIST 1.8B · BACE is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It pairs a self-supervised transformer encoder with a classification head to predict whether small molecules inhibit human β-secretase 1 (BACE-1), a primary therapeutic target in Alzheimer's disease research.
The model builds on the MIST (Molecular Insight SMILES Transformers) foundation model family, which was pre-trained on the Enamine REAL Space database of commercially accessible chemical structures. It addresses the challenge of identifying viable BACE-1 inhibitors by evaluating chemical structures encoded as SMILES strings, assisting medicinal chemists and computational drug discovery researchers.
How it works
The model employs a RoBERTa-PreLayerNorm encoder comprising 24 layers, 24 attention heads, a hidden size of 1536, an intermediate size of 6144, and a maximum context length of 2048 tokens. For pre-training, the encoder was trained on Kekulized SMILES strings from Enamine REAL Space using a masked language modelling objective and cross-entropy loss with the FusedLAMB optimizer across 32 NVIDIA A100 GPUs.
For downstream prediction, hidden states are pooled by taking the representation of the first token, which is passed into a two-layer multi-layer perceptron (MLP) featuring Gaussian Error Linear Units (GELU) activations and dropout. The network was fine-tuned on 1,513 small molecules from the MoleculeNet BACE benchmark using binary cross-entropy loss and the AdamW optimizer on a single NVIDIA A100 GPU, evaluated via an 80/10/10 scaffold split.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with a 2-layer MLP classification head
- Parameter count
- 1.8 billion
- Pre-training data
- Enamine REAL Space
- Fine-tuning data
- MoleculeNet BACE benchmark (1,513 molecules)
- Context length
- 2048 tokens
- Input format
- Kekulized SMILES strings
- Primary task
- Binary classification of BACE-1 inhibition
What you can use it for
- BACE-1 inhibition screening
- Screening candidate compound libraries to predict whether small molecules inhibit β-secretase 1 for Alzheimer's disease drug discovery programmes.
- Virtual screening prioritisation
- Ranking synthesised or catalogue compounds by unnormalised BACE-1 binding logits to prioritise candidates for in vitro assays.
- Molecular representation extraction
- Extracting dense molecular embedding vectors from SMILES inputs for downstream computational chemistry 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 1.8B · BACE
Strengths and limits
Good at
- Pre-trained on the Enamine REAL Space dataset, covering synthetically accessible regions of chemical space.
- Supports sequence lengths up to 2048 tokens, allowing evaluation of long Kekulized SMILES strings.
- Evaluated using a scaffold split rather than random splitting to rigorously assess generalisation across molecular scaffolds.
Watch out for
- Fine-tuned on a relatively small dataset of 1,513 molecules, which may constrain generalisation to under-represented chemical classes.
- Restricted to binary classification of inhibition and does not predict continuous affinity measures such as IC50 or Ki.
- Requires a Rust environment installation to build and run the required 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 1.8B · BACE 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 · BACE 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 expects Kekulized SMILES strings tokenised using the SmirkTokenizerFast tokeniser.
What output does the model produce during inference?
Running prediction on SMILES strings outputs unnormalised logits for binary classification of BACE-1 binding.
What metric was used to evaluate this model?
The model was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUROC) on an 80/10/10 scaffold split of the BACE dataset.
Which checkpoint was used as the base for fine-tuning?
The model was fine-tuned from the pre-trained foundation model checkpoint mist-1.8B-dh61satt.
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., & Viswanathan, V. (2025). Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 [physics]. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 1.8B · BACE'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.