Independent research by Lu Zhang · MIST transfer benchmark
Can a fine-tuned molecular foundation model match task-specific models?
I designed and ran this independent benchmark to first compare fine-tuned MIST with task-specific models on the same QM9 split and 12 properties, then test an ensemble across both model families.
Live prediction client
Predictions, not a benchmark score
Enter a public example to compare the four fixed models on one molecule. This page does not claim that one model wins; model quality needs a separately controlled evaluation.
How does the browser reach the models?
The browser sends only {"smiles":"CCO"}, without cookies, accounts, or API tokens. The backend must return all four models, all 12 targets and units, and schema_version="live-demo-prediction-v1" in one response; otherwise the page refuses to render it.
Native-unit model output
Comparison for —
| Property | Unit | Ridge | XGBoost | MLP | MIST-28M |
|---|
How to read these numbers
- They are model predictions, not measurements, confidence intervals, or official benchmark scores.
- For this QM9 benchmark, DFT-calculated values are treated as the target labels (ground truth) for training and evaluation; they are not experimental ground truth.
- QM9 mainly covers small organic molecules with DFT-computed labels; error can grow outside that chemical domain.
- Disagreement among the four models is not calibrated uncertainty.
- HOMO, LUMO, and gap can inform molecular screening, but they do not directly describe an electrolyte, electrode interface, or battery performance.
This page does not store the input, but the configured prediction service receives the submitted SMILES. Do not submit confidential molecules.