Mind your prevalence!

Multiple metrics are used when assessing and validating the performance of quantitative structure-activity relationship (QSAR) models. In the case of binary classification, balanced accuracy is a metric to assess the global performance of such models. In contrast to accuracy, balanced accuracy does not depend on the respective prevalence of the two categories in the test set that is used to validate a QSAR classifier. As such, balanced accuracy is used to overcome the effect of imbalanced test sets on the model's perceived accuracy. Matthews' correlation coefficient (MCC), an alternative global performance metric, is also known to mitigate the imbalance of the test set. However, in contrast to the balanced accuracy, MCC remains dependent on the respective prevalence of the predicted categories. For simplicity, the rest of this work is based on the positive prevalence. The MCC value may be underestimated at high or extremely low positive prevalence. It contributes to more challenging comparisons between experiments using test sets with different positive prevalences and may lead to incorrect interpretations. The concept of balanced metrics beyond balanced accuracy is, to the best of our knowledge, not yet described in the cheminformatic literature. Therefore, after describing the relevant literature, this manuscript will first formally define a confusion matrix, sensitivity and specificity and then present, with synthetic data, the danger of comparing performance metrics under nonconstant prevalence. Second, it will demonstrate that balanced accuracy is the performance metric accuracy calibrated to a test set with a positive prevalence of 50% (i.e., balanced test set). This concept of balanced accuracy will then be extended to the MCC after showing its dependency on the positive prevalence. Applying the same concept to any other performance metric and widening it to the concept of calibrated metrics will then be briefly discussed. We will show that, like balanced accuracy, any balanced performance metric may be expressed as a function of the well-known values of sensitivity and specificity. Finally, a tale of two MCCs will exemplify the use of this concept of balanced MCC versus MCC with four use cases using synthetic data. SCIENTIFIC CONTRIBUTION: This work provides a formal, unified framework for understanding prevalence dependence in model validation metrics, deriving balanced metric expressions beyond balanced accuracy, and demonstrating their practical utility for common use cases. In contrast to prior literature, it introduces the derived confusion matrix to express metrics as functions of sensitivity, specificity and prevalence without needing additional coefficients. The manuscript extends the concept of balanced metrics to Matthews' correlation coefficient and other widely used performance indicators, enabling robust comparisons under prevalence shifts.

JPlogP: an improved logP predictor trained using predicted data.

The partition coefficient between octanol and water (logP) has been an important descriptor in QSAR predictions for many years and therefore the prediction of logP has been examined countless times. One of the best performing models is to predict the logP using multiple methods and average the result. We have used those averaged predictions to develop a training-set which was able to distil the information present across the disparate logP methods into one single model. Our model was built using extendable atom-types, where each atom is distilled down into a 6 digit number, and each individual atom is assumed to have a small additive effect on the overall logP of the molecule. Beyond the simple coefficient model a consensus model is evaluated, which uses known compounds as a starting point in the calculation and modifies the experimental logP using the same coefficients as in the first model. We then test the performance of our models against two different datasets, one where many different models routinely perform well against, and another designed to more represent pharmaceutical space. The true strength of the model is represented in the pharmaceutical benchmark set, where both models perform better than any previously developed models.

Evaluation of a statistics-based Ames mutagenicity QSAR model and interpretation of the results obtained.

The relative wealth of bacterial mutagenicity data available in the public literature means that in silico quantitative/qualitative structure activity relationship (QSAR) systems can readily be built for this endpoint. A good means of evaluating the performance of such systems is to use private unpublished data sets, which generally represent a more distinct chemical space than publicly available test sets and, as a result, provide a greater challenge to the model. However, raw performance metrics should not be the only factor considered when judging this type of software since expert interpretation of the results obtained may allow for further improvements in predictivity. Enough information should be provided by a QSAR to allow the user to make general, scientifically-based arguments in order to assess and overrule predictions when necessary. With all this in mind, we sought to validate the performance of the statistics-based in vitro bacterial mutagenicity prediction system Sarah Nexus (version 1.1) against private test data sets supplied by nine different pharmaceutical companies. The results of these evaluations were then analysed in order to identify findings presented by the model which would be useful for the user to take into consideration when interpreting the results and making their final decision about the mutagenic potential of a given compound.

Self organising hypothesis networks: a new approach for representing and structuring SAR knowledge.

BACKGROUND: Combining different sources of knowledge to build improved structure activity relationship models is not easy owing to the variety of knowledge formats and the absence of a common framework to interoperate between learning techniques. Most of the current approaches address this problem by using consensus models that operate at the prediction level. We explore the possibility to directly combine these sources at the knowledge level, with the aim to harvest potentially increased synergy at an earlier stage. Our goal is to design a general methodology to facilitate knowledge discovery and produce accurate and interpretable models. RESULTS: To combine models at the knowledge level, we propose to decouple the learning phase from the knowledge application phase using a pivot representation (lingua franca) based on the concept of hypothesis. A hypothesis is a simple and interpretable knowledge unit. Regardless of its origin, knowledge is broken down into a collection of hypotheses. These hypotheses are subsequently organised into hierarchical network. This unification permits to combine different sources of knowledge into a common formalised framework. The approach allows us to create a synergistic system between different forms of knowledge and new algorithms can be applied to leverage this unified model. This first article focuses on the general principle of the Self Organising Hypothesis Network (SOHN) approach in the context of binary classification problems along with an illustrative application to the prediction of mutagenicity. CONCLUSION: It is possible to represent knowledge in the unified form of a hypothesis network allowing interpretable predictions with performances comparable to mainstream machine learning techniques. This new approach offers the potential to combine knowledge from different sources into a common framework in which high level reasoning and meta-learning can be applied; these latter perspectives will be explored in future work.