Does your model know its limits? Leveraging uncertainty estimates in compound prioritisation
Authors
- Matthew Segall
- Samar Mahmoud
- Charlotte Wharrick
- Bailey Montefiore
Accurate uncertainty quantification is critical for the effective, reliable application of in silico models in drug discovery. If a model’s estimates of prediction uncertainty correlate with the observed accuracy of predictions, this means that low-uncertainty predictions can be used with confidence in decision-making.
The applications of uncertainty estimates can go far beyond this: identifying unlikely experimental results that can reveal false negatives and missed opportunities, revealing novel high-potential chemical space where the model is extrapolating, and prioritising new experimental measurements that will add the greatest additional information to improve a model and extend its domain of applicability, i.e. active learning.
