Multi-Modal Foundation · mist-models
MIST 1.8B · SIDER
Predicts 27 MedDRA adverse drug reactions from molecular SMILES strings using a 1.8-billion-parameter encoder.
What MIST 1.8B · SIDER is
MIST 1.8B · SIDER is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It adapts the Molecular Insight SMILES Transformers (MIST) foundation architecture to predict potential adverse drug reactions directly from chemical structures encoded as SMILES strings.
The model assists chemical biologists and medicinal chemists in safety profiling by predicting small-molecule liabilities across 27 Medical Dictionary for Regulatory Activities (MedDRA) side-effect classes. By evaluating adverse drug reaction risks from chemical notation alone, researchers can flag potential toxicological issues before synthesis or experimental assays.
How it works
The architecture uses a RoBERTa-PreLayerNorm encoder with 24 layers, 24 attention heads, a hidden dimension of 1536, and an intermediate dimension of 6144, supporting sequences up to 2048 tokens. The base encoder was pre-trained on molecular SMILES from the Enamine REAL Space database using a masked language modelling objective with cross-entropy loss and the FusedLAMB optimiser on 32 NVIDIA A100 GPUs.
For downstream side-effect classification, the encoder is paired with a task network consisting of a two-layer multi-layer perceptron with Gaussian Error Linear Units (GELU) activations and dropout. The hidden state of the first token is pooled and the network is fine-tuned on Kekulized SMILES from the SIDER dataset using binary cross-entropy loss and the AdamW optimiser on a single NVIDIA A100 GPU.
Key facts
- Encoder architecture
- RoBERTa-PreLayerNorm (24 layers, hidden size 1536, 24 attention heads)
- Task network
- Two-layer multi-layer perceptron (MLP) with GELU activations and dropout
- Pre-training dataset
- Enamine REAL Space
- Fine-tuning dataset
- SIDER from MoleculeNet (1,427 marketed drugs across 27 MedDRA terms)
- Input format
- Kekulized SMILES strings
- Maximum sequence length
- 2048 tokens
- Hardware memory requirement
- 24 GB GPU memory
What you can use it for
- Adverse drug reaction profiling
- Generate unnormalised logits for 27 MedDRA adverse reaction categories directly from candidate SMILES strings.
- Early safety screening
- Prioritise small-molecule drug candidates by evaluating predicted multi-label toxicity risk factors prior to synthesis.
- Molecular feature extraction
- Extract high-dimensional vector representations of chemical structures for downstream analysis and clustering.
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 · SIDER
Strengths and limits
Good at
- Pre-trained on the Enamine REAL Space compound collection, covering synthetically accessible regions of chemical space.
- Supports long context inputs up to 2048 tokens, accommodating large and complex molecular SMILES strings.
- Outputs simultaneous multi-label binary predictions across 27 distinct MedDRA clinical side-effect categories.
Watch out for
- Supervised fine-tuning is limited to the 1,427 marketed drugs present in the SIDER dataset.
- Deployment requires Rust to be installed in the environment to build the required SMIRK tokeniser.
- Inference requires significant compute resources, demanding up to 24 GB of GPU memory.
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 · 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 1.8B · 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 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 notation does the model require?
The model expects molecular structures provided as Kekulized SMILES strings, processed using the SMIRK tokeniser.
What does the model output during inference?
It outputs unnormalised logits across 27 MedDRA side-effect categories for multi-label classification, or single pooled feature vectors when used as a feature extractor.
Which base model was fine-tuned for this task?
The model was fine-tuned from the pre-trained foundation encoder mist-1.8B-dh61satt.
How was the model evaluated?
It was evaluated on the SIDER dataset using an 80/10/10 random split, measuring the Area Under the Receiver Operating Characteristic curve (AUROC) averaged across all 27 classification tasks.
How to cite it
Alexius Wadell, Anoushka Bhutani, Victor Azumah, Austin R. Ellis-Mohr, Celia Kelly, Hancheng Zhao, Anuj K. Nayak, Kareem Hegazy, Alexander Brace, Hongyi Lin, Murali Emani, Venkatram Vishwanath, Kevin Gering, Melisa Alkan, Tom Gibbs, Jack Wells, Lav R. Varshney, Bharath Ramsundar, Karthik Duraisamy, Michael W. Mahoney, Arvind Ramanathan, and Venkatasubramanian Viswanathan. 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 · 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.