A quantum mechanics (QM)-based scoring function has been applied to complexes of cyclin-dependent kinase 2 (CDK2) and thirty-one pyrazolo[1,5-a]pyrimidine-based inhibitors and their bioisosteres. A hybrid three-layer QM/MM setup (DFT-D/PM6-D3H4X/AMBER in generalized Born solvent) was used here for the first time as an extension of our previous full QM and SQM/MM (SQM means semiempirical QM) approaches. Two approaches to obtain the structures of the CDK2/inhibitor complexes were examined: i) building the modifications from one X-ray structure available coupled with a conformational search and ii) docking the compounds into CDK2. The QM-based scoring entailed a QM/SQM/MM optimization followed by calculations of the binding scores which were subsequently correlated with the experimental binding free energies. The correlation for the building protocol was good (r(2) = 0.64, predictive index = 0.81), whereas the docking approach failed. A decomposition of the interaction energies to ligand fragments enabled us to rationalize the differences in the binding affinities. In conclusion, we have developed and refined a QM-based scoring protocol and successfully applied it to reproduce the binding affinities in congeneric series of CDK2 inhibitors and to rationalize their potency. We thus propose that such a tool can be used in computer-aided rational drug design.