Anita Rácz

BACKGROUND: Cheminformaticians are equipped with a very rich toolbox when carrying out molecular similarity calculations. A large number of molecular representations exist, and there are several methods (similarity and distance metrics) to quantify the similarity of molecular representations. In this work, eight well-known similarity/distance metrics are compared on a large dataset of molecular fingerprints with sum of ranking differences (SRD) and ANOVA analysis. The effects of molecular size, selection methods and data pretreatment methods on the outcome of the comparison are also assessed. RESULTS: A supplier database (https://mcule.com/) was used as the source of compounds for the similarity calculations in this study. A large number of datasets, each consisting of one hundred compounds, were compiled, molecular fingerprints were generated and similarity values between a randomly chosen reference compound and the rest were calculated for each dataset. Similarity metrics were compared based on their ranking of the compounds within one experiment (one dataset) using sum of ranking differences (SRD), while the results of the entire set of experiments were summarized on box and whisker plots. Finally, the effects of various factors (data pretreatment, molecule size, selection method) were evaluated with analysis of variance (ANOVA). CONCLUSIONS: This study complements previous efforts to examine and rank various metrics for molecular similarity calculations. Here, however, an entirely general approach was taken to neglect any a priori knowledge on the compounds involved, as well as any bias introduced by examining only one or a few specific scenarios. The Tanimoto index, Dice index, Cosine coefficient and Soergel distance were identified to be the best (and in some sense equivalent) metrics for similarity calculations, i.e. these metrics could produce the rankings closest to the composite (average) ranking of the eight metrics. The similarity metrics derived from Euclidean and Manhattan distances are not recommended on their own, although their variability and diversity from other similarity metrics might be advantageous in certain cases (e.g. for data fusion). Conclusions are also drawn regarding the effects of molecule size, selection method and data pretreatment on the ranking behavior of the studied metrics. Graphical AbstractA visual summary of the comparison of similarity metrics with sum of ranking differences (SRD).

Applied datasets can vary from a few hundred to thousands of samples in typical quantitative structure-activity/property (QSAR/QSPR) relationships and classification. However, the size of the datasets and the train/test split ratios can greatly affect the outcome of the models, and thus the classification performance itself. We compared several combinations of dataset sizes and split ratios with five different machine learning algorithms to find the differences or similarities and to select the best parameter settings in nonbinary (multiclass) classification. It is also known that the models are ranked differently according to the performance merit(s) used. Here, 25 performance parameters were calculated for each model, then factorial ANOVA was applied to compare the results. The results clearly show the differences not just between the applied machine learning algorithms but also between the dataset sizes and to a lesser extent the train/test split ratios. The XGBoost algorithm could outperform the others, even in multiclass modeling. The performance parameters reacted differently to the change of the sample set size; some of them were much more sensitive to this factor than the others. Moreover, significant differences could be detected between train/test split ratios as well, exerting a great effect on the test validation of our models.

BACKGROUND: Interaction fingerprints (IFP) have been repeatedly shown to be valuable tools in virtual screening to identify novel hit compounds that can subsequently be optimized to drug candidates. As a complementary method to ligand docking, IFPs can be applied to quantify the similarity of predicted binding poses to a reference binding pose. For this purpose, a large number of similarity metrics can be applied, and various parameters of the IFPs themselves can be customized. In a large-scale comparison, we have assessed the effect of similarity metrics and IFP configurations to a number of virtual screening scenarios with ten different protein targets and thousands of molecules. Particularly, the effect of considering general interaction definitions (such as Any Contact, Backbone Interaction and Sidechain Interaction), the effect of filtering methods and the different groups of similarity metrics were studied. RESULTS: The performances were primarily compared based on AUC values, but we have also used the original similarity data for the comparison of similarity metrics with several statistical tests and the novel, robust sum of ranking differences (SRD) algorithm. With SRD, we can evaluate the consistency (or concordance) of the various similarity metrics to an ideal reference metric, which is provided by data fusion from the existing metrics. Different aspects of IFP configurations and similarity metrics were examined based on SRD values with analysis of variance (ANOVA) tests. CONCLUSION: A general approach is provided that can be applied for the reliable interpretation and usage of similarity measures with interaction fingerprints. Metrics that are viable alternatives to the commonly used Tanimoto coefficient were identified based on a comparison with an ideal reference metric (consensus). A careful selection of the applied bits (interaction definitions) and IFP filtering rules can improve the results of virtual screening (in terms of their agreement with the consensus metric). The open-source Python package FPKit was introduced for the similarity calculations and IFP filtering; it is available at: https://github.com/davidbajusz/fpkit .

Machine learning classification algorithms are widely used for the prediction and classification of the different properties of molecules such as toxicity or biological activity. the prediction of toxic vs. non-toxic molecules is important due to testing on living animals, which has ethical and cost drawbacks as well. The quality of classification models can be determined with several performance parameters. which often give conflicting results. In this study, we performed a multi-level comparison with the use of different performance metrics and machine learning classification methods. Well-established and standardized protocols for the machine learning tasks were used in each case. The comparison was applied to three datasets (acute and aquatic toxicities) and the robust, yet sensitive, sum of ranking differences (SRD) and analysis of variance (ANOVA) were applied for evaluation. The effect of dataset composition (balanced vs. imbalanced) and 2-class vs. multiclass classification scenarios was also studied. Most of the performance metrics are sensitive to dataset composition, especially in 2-class classification problems. The optimal machine learning algorithm also depends significantly on the composition of the dataset.

Recent implementations of QSAR modelling software provide the user with numerous models and a wealth of information. In this work, we provide some guidance on how one should interpret the results of QSAR modelling, compare and assess the resulting models, and select the best and most consistent ones. Two QSAR datasets are applied as case studies for the comparison of model performance parameters and model selection methods. We demonstrate the capabilities of sum of ranking differences (SRD) in model selection and ranking, and identify the best performance indicators and models. While the exchange of the original training and (external) test sets does not affect the ranking of performance parameters, it provides improved models in certain cases (despite the lower number of molecules in the training set). Performance parameters for external validation are substantially separated from the other merits in SRD analyses, highlighting their value in data fusion.