Assessing the Prediction Quality of the Anti-SARS-CoV-2 Activity Using the D3Targets-2019-nCoV Web Service

N.S. Ionov, P.V. Pogodin, V.V. Poroikov
2020 Biomedical Chemistry Research and Methods  
The D3Targets-2019-nCoV web service predicting the interaction of chemical compounds with SARS-CoV-2 virus proteins and human proteins involved in the pathogenesis of COVID-19 by structural similarity and molecular docking was evaluated. The quality of the prediction was assessed as a balanced accuracy, which was calculated based on the results of the prediction for the structures of chemical compounds from the test set we compiled. The test set consisted of 35 active and 59 inactive molecules,
more » ... inactive molecules, including compounds with the experimetnaly confirmed absence of activity against the selected targets and compounds active against SARS-CoV-2 targets, not presented in the CoViLigands database. The authors of the analyzed web service did not indicate the thresholds for the values of the similarity score and the docking scoring function, using which it would be possible to reliably divide the compounds into active and inactive with respect to target proteins. Therefore, we assessed the balanced accuracy of the predictive methods D3Targets-2019-nCoV at various thresholds for cutting off active substances from inactive ones. Using our test set it was found that the highest value of balanced accuracy (0.59) was achieved when choosing active molecules based on the results of 2D similarity assessment (cutoff threshold was 46%). Assessment of 3D similarity did not allow achieving balanced accuracy values exceeding 0.5. It is shown that using the 2Dх3D integral similarity assessment recommended by the authors, the maximum value of the balanced accuracy 0.57 was achieved at a threshold of 31%. The calculated balanced accuracy for molecular docking results does not exceed 0.51. On the case study for the tideglusib, it was shown that the values of the scoring function for two target proteins, the activity against which was confirmed in the experiment (3CLpro and GSK3B), do not differ significantly from the values of the scoring function for the remaining 44 targets were not confirmed.
doi:10.18097/bmcrm00140 fatcat:oefz3re7mvbe3hln72z7ioj42i