The machine learning-based virtual screening of molecular databases is a commonly used approach to identify hits. However, many aspects associated with training predictive models can influence the final performance and, consequently, the number of hits found. Thus, we performed a systematic study of the simultaneous influence of the proportion of negatives to positives in the testing set, the size of screening databases and the type of molecular representations on the effectiveness of classification. The results obtained for eight protein targets, five machine learning algorithms (SMO, Naïve Bayes, Ibk, J48 and Random Forest), two types of molecular fingerprints (MACCS and CDK FP) and eight screening databases with different numbers of molecules confirmed our previous findings that increases in the ratio of negative to positive training instances greatly influenced most of the investigated parameters of the ML methods in simulated virtual screening experiments. However, the performance of screening was shown to also be highly dependent on the molecular library dimension. Generally, with the increasing size of the screened database, the optimal training ratio also increased, and this ratio can be rationalized using the proposed cost-effectiveness threshold approach. To increase the performance of machine learning-based virtual screening, the training set should be constructed in a way that considers the size of the screening database.