Modern QSAR approaches have wide practical applications in drug discovery for screening potentially bioactive molecules before their experimental testing. Most models predicting the bioactivity of compounds are based on molecular descriptors derived from 2D structure losing explicit information about the spatial structure of molecules which is important for protein-ligand recognition. The major problem in constructing models using 3D descriptors is the choice of a probable bioactive conformation that affects the predictive performance. Multi-instance (MI) learning approach considering multiple conformations upon the model training can be a reasonable solution to the above problem. Here, we compared MI-QSAR with the classical single-instance QSAR (SI-QSAR) approach, where each molecule was encoded by either 2D descriptors or 3D descriptors issued from the single lowest-energy conformation. The calculations were carried out on a sample of 175 datasets extracted from the ChEMBL23 database. It was demonstrated that (i) MI-QSAR outperforms SI-QSAR in numerous cases and (ii) MI algorithms can automatically identify plausible bioactive conformations. Instance-attention based network can be applied for most important conformer selection which was shown to correspond PDB conformer in 50-84% of molecules.