The euroSAMPL1 pK_a blind prediction and reproducible research data management challenge

The development and testing of methods in computational chemistry for the prediction of physicochemical properties is by now a mature form of scientific research, with a number of different methods ranging from molecular mechanics simulations, over quantum calculations, to empirical and machine learning models. Blind prediction challenges for these properties are regularly organized to allow researchers from academia and industry to test their methods in a fair and unbiased manner. At the same time, research data management (RDM) is still not utilized as extensively as it could be in the development and application of such models, especially in academia. In particular, the FAIR standards (Findable, Accessible, Interoperable, Reusable) can serve as guidelines for good RDM, but many models, the data used to train them, and the data they generate fall short of one, or multiple, of these standards. The goal of the first euroSAMPL pK_a blind prediction challenge was to promote and help develop good RDM standards for computational chemistry. To achieve this, the challenge was designed to rank not just the predictive performance of the models but also evaluate the adherence to the FAIR principles by cross-evaluation of the participants themselves. We here present the analysis of the blind prediction quality by their statistical metrics as well as of the cross-evaluation by a newly defined "FAIRscore". The results suggest that multiple methods can predict the pK_a to within chemical accuracy, but also that "consensus" predictions constructed from multiple, independent methods may outperform each individual prediction. Furthermore, the state of research data management in the field of computational chemistry is discussed, and suggestions for future improvements developed.