English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Advances in deep neural network (DNN) based molecular property predictionhave recently led to the development of models of remarkable accuracy andgeneralization ability, with graph convolution neural networks (GCNNs)reporting state-of-the-art performance for this task. However, some challengesremain and one of the most important that needs to be fully addressed concernsuncertainty quantification. DNN performance is affected by the volume and thequality of the training samples. Therefore, establishing when and to whatextent a prediction can be considered reliable is just as important asoutputting accurate predictions, especially when out-of-domain molecules aretargeted. Recently, several methods to account for uncertainty in DNNs havebeen proposed, most of which are based on approximate Bayesian inference. Amongthese, only a few scale to the large datasets required in applications.Evaluating and comparing these methods has recently attracted great interest,but results are generally fragmented and absent for molecular propertyprediction. In this paper, we aim to quantitatively compare scalable techniquesfor uncertainty estimation in GCNNs. We introduce a set of quantitativecriteria to capture different uncertainty aspects, and then use these criteriato compare MC-Dropout, deep ensembles, and bootstrapping, both theoretically ina unified framework that separates aleatoric/epistemic uncertainty andexperimentally on the QM9 dataset. Our experiments quantify the performance ofthe different uncertainty estimation methods and their impact onuncertainty-related error reduction. Our findings indicate that ensembling andbootstrapping consistently outperform MC-Dropout, with differentcontext-specific pros and cons. Our analysis also leads to a betterunderstanding of the role of aleatoric/epistemic uncertainty and highlights thechallenge posed by out-of-domain uncertainty.

Evaluating Scalable Uncertainty Estimation Methods for DNN-Based Molecular Property Prediction