Meet with us at ACS Fall 2024

AI-guided discovery

Denver, CO, USA

August 18 -22 , 2024

Join Daniel Barr, PhD at the ACS Fall elevating chemistry meeting for his talk on how “Explainable deep learning imputation prioritises the most relevant data, accounts for uncertainty, and guides experiment selection to bring additional value to small molecule discovery”

Where: Room 709

When: Sunday 18th at 9am

Session: Talk – 4104164. COMP ‘Machine Learning in Chemistry: Chemical Representations & Property Prediction’ session.

Daniel’s presentation is all about our Cerella™ AI platform – let us know if you’d like a chat about how Cerella could transform your discovery workflows.

Get in touch

Abstract:

The last decade has seen a remarkable rise in the number and scope of artificial intelligence and machine learning (especially deep learning) algorithms for small molecule discovery. Great interest is arising around questions of explainability with the expanding scope of these algorithms; ‘black box’ models are increasingly insufficient when dealing with sensitive or confidential data. Because many deep learning models use thousands or even millions of parameters, explaining how they arrive at any particular result is difficult. On the other hand, explainable models typically use only dozens of parameters and are often unable to achieve the same accuracy as the ‘black box’ models. Using case studies, we will demonstrate how the Alchemite™ method for deep learning imputation, as implemented in Cerella, outperforms traditional QSAR methods for the prediction of molecular activities and properties without sacrificing interpretability. We will show that these models capture more than twice the information about the relationships between endpoints than correlation analysis. We will discuss how this approach improves the accuracy of prediction by excluding extraneous data and utilising indirect correlations that other methods miss. These results provide strong evidence for incorporating explainable deep learning methods in platforms such as Cerella for drug discovery and related research.