Publications and Presentations

Data Imputation through Deep Learning

Data Imputation through Deep Learning

Nov 20, 2020

This article was published in Innovations in Pharmaceutical Technology magazine, Autumn/Winter 2020

Data Imputation through Deep Learning: Expanding the data available for drug discovery by up to 100x
M. Segall*, B. Irwin*, T. Whitehead†, S. Mahmoud*, G. Shields*, G. Turner*, A. Elliott*, S.-B. Marcu*, R. Parini†, E. Champness*, G. Conduit†‡
*Optibrium Ltd., Cambridge, UK, [email protected], † Intellegens Ltd., Cambridge, UK, [email protected], ‡Cavendish Laboratory, University of Cambridge, Cambridge UK

Data imputation through deep learning IPT article

Machine learning (ML) methods are routinely used in drug discovery to build models that can predict the properties of compounds directly from their chemical structure. These quantitative structure-activity relationship (QSAR) models take ‘features’ of chemical structures (often referred to as ‘descriptors’) as input to predict one or more properties, including activities against biological targets or in phenotypic assays and a broad range of absorption, distribution, metabolism, excretion and toxicity (ADMET) properties. However, even the most sophisticated ML methods can struggle to produce high-quality predictions. This is due to the limitations of drug discovery data. The number of compounds with data for any given experimental endpoint is small when compared with machine learning data sets in many other fields; the overlap of compounds measured in different endpoints is even smaller; and the data generated by biological assays are noisy due to experimental variability.

Imputation methods take a different approach. They use the limited property data that are available as inputs, to ‘fill in the gaps’ where measured values are not yet available. An example of an imputation method is Alchemite™, which applies deep learning to both compound descriptors and sparse assay data, as illustrated in Figure 1. The resulting model ‘learns’ directly from correlations between experimental endpoints, in addition to relationships between structural features of compounds and the experimental data. This approach makes better use of the sparse and noisy data in drug discovery. It produces more accurate predictions than QSAR models, which enables better targeting of the most promising compounds.


Discover Cerella™

Cerella™ is a unique artificial intelligence platform which supports medicinal chemists and other discovery scientists. It escalates success rates and advances small molecule drug discovery, from working with early hits to nominating preclinical candidates.

Cerella’s AI platform is proven to overcome limitations in drug discovery data, confidently delivering results, seamlessly integrating with your med chem software platforms.