Multi-Modal Foundation · mist-models
MIST 28M · SIDER
A molecular transformer predicting 27 adverse drug reaction classes from SMILES strings.
What MIST 28M · SIDER is
MIST 28M · SIDER is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It belongs to the Molecular Insight SMILES Transformers (MIST) family of foundation models, which are pre-trained on molecular representations and adapted to downstream chemical and biological tasks.
The model addresses the challenge of anticipating adverse drug reactions from chemical structure alone. By fine-tuning a pre-trained 28-million-parameter encoder on the Side Effect Resource (SIDER) benchmark, it predicts drug side effects mapped across 27 Medical Dictionary for Regulatory Activities (MedDRA) system-organ classes.
This system aids researchers screening small-molecule drug candidates by providing automated, multi-label side-effect profiles directly from chemical SMILES notations before advancing compounds to experimental or clinical testing.
How it works
The model uses a RoBERTa-PreLayerNorm encoder architecture consisting of 8 layers, 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a maximum sequence length of 2048 tokens. It was pre-trained using a masked language modelling objective on Kekulised SMILES strings from the Enamine REAL Space dataset, optimised via DeepSpeed FusedLAMB with cross-entropy loss.
For downstream side-effect prediction, the pre-trained encoder (mist-28M-ti624ev1) is coupled with a task-specific network comprising a two-layer multi-layer perceptron with Gaussian Error Linear Units (GELU) activations and dropout. Sequence pooling is performed using the hidden state of the first token. The full model was fine-tuned on the SIDER dataset using PyTorch Lightning, AdamW optimisation, and binary cross-entropy loss to output unnormalised logits for 27 adverse reaction categories.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with a 2-layer MLP classification head
- Parameter count
- 28M (encoder)
- Pre-training data
- Enamine REAL Space SMILES strings
- Fine-tuning data
- SIDER dataset from MoleculeNet (1,427 marketed drugs)
- Context length
- 2048 tokens
What you can use it for
- Adverse drug reaction profiling
- Predict unnormalised logits across 27 MedDRA system-organ side-effect terms for novel small-molecule structures.
- Preclinical safety screening
- Triage drug candidates and chemical libraries computationally based on predicted side-effect risks prior to synthesis or biological assays.
- Molecular representation extraction
- Generate 512-dimensional molecular embedding vectors from Kekulised SMILES strings via the pooled encoder states for downstream 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 · SIDER
Strengths and limits
Good at
- Trained on diverse, synthetically accessible chemical space from the Enamine REAL Space compound collection.
- Outputs comprehensive multi-label predictions across 27 distinct MedDRA side-effect categories simultaneously.
- Accommodates long sequence inputs with a maximum sequence length of 2048 tokens.
Watch out for
- Fine-tuning is restricted to the relatively small SIDER dataset of 1,427 marketed drugs.
- Requires Kekulised SMILES strings as input rather than raw or arbitrary molecular formats.
- Utilises SMIRK tokenisation which requires a Rust installation to run.
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 · SIDER 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 · SIDER 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 MIST 28M · SIDER accept?
The model takes molecular sequences formatted as Kekulised SMILES strings.
What does the model output during inference?
Calling predict outputs unnormalised logits for 27 MedDRA adverse drug reaction categories. It can also produce a pooled embedding vector representing the input molecule.
Which software packages are required to use the tokenizer?
The model uses the SMIRK tokenizer, which requires Rust to be installed in the operating environment.
How were the evaluation splits organised for SIDER?
The model was evaluated using a random 80/10/10 split on the SIDER dataset, measured using the Area Under the Receiver Operating Characteristic curve (AUROC) averaged across all 27 tasks.
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, 2025, arXiv:2510.18900 [physics], doi:10.48550/arXiv.2510.18900
Written from MIST 28M · SIDER'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.