한국제약바이오협회

With the ongoing rapid growth of publicly available ligand–protein bioactivity data, there is a trove of valuable data that can be used to train a plethora of machine-learning algorithms. However, not all data is equal in terms of size and quality and a significant portion of researchers’ time is needed to adapt the data to their needs. On top of that, finding the right data for a research question can often be a challenge on its own. To meet these challenges, we have constructed the Papyrus dataset. Papyrus is comprised of around 60 million data points. This dataset contains multiple large publicly available datasets such as ChEMBL and ExCAPE-DB combined with several smaller datasets containing high-quality data. The aggregated data has been standardised and normalised in a manner that is suitable for machine learning. We show how data can be filtered in a variety of ways and also perform some examples of quantitative structure–activity relationship analyses and proteochemometric modelling. Our ambition is that this pruned data collection constitutes a benchmark set that can be used for constructing predictive models, while also providing an accessible data source for research.

Construction of Papyrus

The Papyrus dataset was obtained by collecting and processing ChEMBL’s (version 30) 19,286,751 activity data points measured on 2,157,379 compounds and 14,855 targets, ExCAPE-DB’s 70,850,163 activity data points of 998,131 compounds measured on 1667 targets, Sharma et al.’s dataset [23] of 258,060 activity data points of 76,017 compounds measured on 8 targets, Christmann-Franck et al.’s dataset [24] of 344,788 activity data points of 2065 compounds measured on 448 targets, Klaeger et al.’s dataset [13] of 5916 activity data points of 243 compounds measured on 520 targets and Merget et al.’s dataset[25] of 260,757 activity data points of 47,774 compounds measured on 341 targets. The data was standardized and filtered after which 59,775,087 activity values associated with 1,270,570 unique two-dimensional compound structures and 6926 proteins were obtained.

Complete preparation steps taken to create the Papyrus dataset are available in Additional file 4 and parameters in Additional file 2: Tables S1–S13. Briefly, only data associated with Ki, K_D, IC₅₀, EC₅₀ and their logarithm transforms were considered if expressed in molar concentrations, molecules structures were standardized using ChEMBL structure pipeline [27] as well as a combination of OpenBabel [28, 29], tautomer canonicalization and Dimorphite-DL [30]. Proteins were mapped to UniProt [31] identifiers, sequences, and ChEMBL’s tiered protein classification.

Throughout the filtering and standardization process, the data were prepared considering three levels of quality for machine learning: the data regression models can be developed from are labelled high-quality while those classifiers can model are labelled low-quality. Medium quality is available for regression models and is associated with bioactivity data points associated with lower quality of the associated bioassays.

Papyrus dataset statistics

The Papyrus dataset consists of 59,775,087 compound-protein pairs, each associated with at least either one activity value or activity class. Additionally, this represents the data of 1,270,570 unique two-dimensional compound structures and 6926 proteins across 499 different organisms. In terms of data quality, 1,238,835 data points are of high quality, i.e., representing exact bioactivity values measured and associated with a single protein or complex subunit. 335,661 data points are of medium quality, i.e., exact bioactivity values associated with either potentially multiple proteins or a homologous single protein. 58,200,591 data points are of low quality, i.e., exact bioactivity values associated with either multiple homologous proteins or homologous complex subunits, censored bioactivity values and binary activity classes. When considering data points across all quality types, 2,585,248 are associated with exact bioactivity values, 354,981 with censored data and 56,823,552 with binary activity classes. The repartition of data quality across the ten organisms with the most data (Table 1) indicates a clear bias towards humans, with 55,595,516 data points or more than 93% of the data related to it, but also emphasizes the interest towards rodent targets with 2,513,821 data points or more than 4% of the data associated with mouse and 1,244,385 data points or 2% with rats.