Models trained on model-made ADME labels hold up on real data
- Researchers trained multitask graph neural networks on a large compound set whose ADME labels were themselves predictions from earlier models, covering roughly two dozen endpoints.
- The surrogate models were then tested on new in-house measurements, a prospective check, and on public experimental data.
- They stayed predictive across several properties, their uncertainty estimates flagged less reliable predictions, and fine-tuning on measured data improved most endpoints.
💡 Points Based on the abstract
Training on predictions of predictions risks inheriting the teacher models' biases, so I'd want to see a head-to-head against the teachers on the same held-out compounds, not just correlation with measurements. On my own series, I'd check whether the uncertainty flag holds on new chemotypes outside the training space, since that is where ADME models usually break.
📄 Paper · Journal of Chemical Information and Modeling (ACS) ↗ Surrogate Machine Learning Models for ADME Prediction: From Synthetic Labels to Experimental Data