Estimation of reliability of predictions and model applicability domain evaluation in the analysis of acute toxicity (LD₅₀)

Abstract

This study presents a new type of acute toxicity (LD(50)) prediction that enables automated assessment of the reliability of predictions (which is synonymous with the assessment of the Model Applicability Domain as defined by the Organization for Economic Cooperation and Development). Analysis involved nearly 75,000 compounds from six animal systems (acute rat toxicity after oral and intraperitoneal administration; acute mouse toxicity after oral, intraperitoneal, intravenous, and subcutaneous administration). Fragmental Partial Least Squares (PLS) with 100 bootstraps yielded baseline predictions that were automatically corrected for non-linear effects in local chemical spaces--a combination called Global, Adjusted Locally According to Similarity (GALAS) modelling methodology. Each prediction obtained in this manner is provided with a reliability index value that depends on both compound's similarity to the training set (that accounts for similar trends in LD(50) variations within multiple bootstraps) and consistency of experimental results with regard to the baseline model in the local chemical environment. The actual performance of the Reliability Index (RI) was proven by its good (and uniform) correlations with Root Mean Square Error (RMSE) in all validation sets, thus providing quantitative assessment of the Model Applicability Domain. The obtained models can be used for compound screening in the early stages of drug development and prioritization for experimental in vitro testing or later in vivo animal acute toxicity studies.

Trainable structure–activity relationship model for virtual screening of CYP3A4 inhibition

“…Recently the successful application of this methodology in predicting various properties of continuous nature (e.g., LogP, LD 50 ) has been reported [24,25]. Two main parts constituting the basis of the method are a global QSAR model providing baseline predictions for the property of interest and local corrections calculated according to the experimental data for the most similar compounds from the training set.…”

“…The major part of this set is intended for the description of general chemical compound constitution and was comprised of conventional fragmental descriptors, such as atoms, functional groups, molecular shape fragments, and others. These descriptors have been already used and proved effective in previous projects, involving modeling various biological activities and chemical properties [24,25]. Additionally several typical fragments describing CYP3A4 specificity were added (e.g., nitrogen containing heterocycles, methylenedioxybenzene, fragments representing possible cytochrome P450 metabolism sites, etc.…”

“…A novel GALAS (Global, Adjusted Locally According to Similarity) modeling methodology was used, which allowed forecasting the reliability of each prediction. The successful applications of this method in predicting continuous properties, such as LogP and acute toxicity in terms of LD 50 , have been recently described in detail [24,25]. In case of CYP3A4 inhibition models it was adapted for binary data.…”

Target‐Driven Subspace Mapping Methods and Their Applicability Domain Estimation

“…Other methods make use of different distance metrics to measure distance from an unseen compound to cluster centroids or a number of nearest neighbors in the descriptor space of the training set and then, they establish a criterion of reliability based on the value of those distances. [29][30][31][32][33] The major problem of these metrics is that proximity in the descriptor space does not necessarily imply a close connection with the target vector.…”

“…[34] Such approaches, for example, analyze the variability of ensemble methods in the predictions, [35,36] while other approaches combine variability in the predictions with distances to the training set. [33,37,38,39] A recent work particularly addresses the problem of applicability domain for kernel-based machine learning models. [40] The method presented here is established in analogy with density-based methods.…”