Upgrading Reliability in Molecular Property Prediction by Robust Quantification of Uncertainty from Machine Learning Models