Comparative efficiency of structure activity relationship and proteochemometric modelling

Virtual screening of biologically active compounds is widely applied for the search of drug leads. The well-known methods of structure-activity relationship (SAR) are based on the chemical structure comparison. In the last years, an approach known as proteochemometrics (PCM) has also gained popularity. PCM extends the capabilities of SAR by incorporating the protein target descriptors into the model. Unlike SAR, PCM can be used to predict new targets with unknown spectra of ligands. As both approaches can be used to predict ligands for the known proteins, several researchers apply PCM to solve this task, without providing compelling reasons to support the superiority of the PCM approach over SAR. To correctly compare the performance of SAR and PCM in the given situation, we have developed a special validation scheme. As a result, we did not find any advantages of PCM over SAR in the prediction of ligands for the protein with an established ligand spectrum. At the same time, the validation procedure commonly used for PCM models considerably inflates the evaluation scores compared to our technique. Widespread use of such validation scheme leads to conclusions that PCM has great advantage over SAR in contrast to our findings. Thus, our study emphasizes that a transparent and correct validation scheme is essential for comparison of different methods.