Structural Alerts and Random Forest Models in a Consensus Approach for Receptor Binding Molecular Initiating Events

Abstract

A molecular initiating event (MIE) is the gateway to an adverse outcome pathway (AOP), a sequence of events ending in an adverse effect. In silico predictions of MIEs are a vital tool in a modern, mechanism-focused approach to chemical risk assessment. For 90 biological targets representing important human MIEs, structural alert-based models have been constructed with an automated procedure that uses Bayesian statistics to iteratively select substructures. These models give impressive average performance statistics (an average of 92% correct predictions across targets), significantly improving on previous models. Random Forest models have been constructed from physicochemical features for the same targets, giving similarly impressive performance statistics (93% correct predictions). A key difference between the models is interpretation of predictionsthe structural alert models are transparent and easy to interpret, while Random Forest models can only identify the most important physicochemical features for making predictions. The two complementary models have been combined in a consensus model, improving performance compared to each individual model (94% correct predictions) and increasing confidence in predictions. Variation in model performance has been explained by calculating a modelability index (MODI), using Tanimoto coefficient between Morgan fingerprints to identify nearest neighbor chemicals. This work is an important step toward building confidence in the use of in silico tools for assessment of toxicity.

Neural network activation similarity: a new measure to assist decision making in chemical toxicology

“…These are the targets KCNH2 and MAPK1, which were identied in our previous publication as challenging classications in this dataset. 38 Positive probability values were also calculated for AR binders using the Somax function on an optimal trained DNN (Fig. 2).…”

“…Clustered statistics are taken from Table 3 and random statistics generated from Table 5 The best performing DNN for each biological target was compared to models previously generated using SAs and RFs. 38 In our previous study, the SAs were constructed using structures obtained from a maximal common substructure algorithm and selected using Bayesian statistics to iteratively select the best alerts. The RF models were based on 200 physicochemical descriptors calculated in RDKit and modelled in sklearn.…”

“…Data for 79 pharmacologically important biological targets were extracted from the publicly available databases ChEMBL 40 and ToxCast 41 (Table 1). These targets are a subset of those used in our previous work 38 including targets published by Bowes et al 37 and Sipes et al 39 These targets were chosen as they provide valuable toxicological information for risk assessment. ChEMBL and ToxCast were combined to provide a relatively balanced dataset with more than 1000 chemicals per target for model construction and evaluation, an amount that was found to be required for DNN training.…”

“…As an additional comparison point, RF similarity (RFS), was calculated using the Euclidean distance between normalized vectors consisting of the 50 most important physicochemical descriptors identied in RF model construction. 38 This gives an appropriate comparison point based on trained machine learning models, and Tanimoto similarity gives a comparison based on chemical similarity and the DNN model inputs.…”

“…16,34 While machine learning algorithms such as DNNs, 13 shallow NNs and decision trees, 35 and convolutional neural networks 36 have been used in ligand-receptor binding binary classication tasks in the past, this is the rst time DNNs have been applied to predicting MIEs. A wide variety of diverse and important human targets were chosen for DNN classier construction, including the well-known Bowes targets 37 and an extended list identied in our previous work 38 including targets published by Sipes et al 39 Many of the previously published papers on machine learning only provide models for a single biological target or endpoint. [1][2][3][4][5]11,12,15 This limits their usefulness and coverage of human toxicology, which we aim to overcome by constructing models for a large number of MIEs.…”