Ensemble modeling with machine learning and deep learning to provide interpretable generalized rules for classifying CNS drugs with high prediction power

Abstract

The trade-off between a machine learning (ML) and deep learning (DL) model’s predictability and its interpretability has been a rising concern in central nervous system-related quantitative structure–activity relationship (CNS-QSAR) analysis. Many state-of-the-art predictive modeling failed to provide structural insights due to their black box-like nature. Lack of interpretability and further to provide easy simple rules would be challenging for CNS-QSAR models. To address these issues, we develop a protocol to combine the power of ML and DL to generate a set of simple rules that are easy to interpret with high prediction power. A data set of 940 market drugs (315 CNS-active, 625 CNS-inactive) with support vector machine and graph convolutional network algorithms were used. Individual ML/DL modeling methods were also constructed for comparison. The performance of these models was evaluated using an additional external dataset of 117 market drugs (42 CNS-active, 75 CNS-inactive). Fingerprint-split validation was adopted to ensure model stringency and generalizability. The resulting novel hybrid ensemble model outperformed other constituent traditional QSAR models with an accuracy of 0.96 and an F1 score of 0.95. With the power of the interpretability provided with this protocol, our model laid down a set of simple physicochemical rules to determine whether a compound can be a CNS drug using six sub-structural features. These rules displayed higher classification ability than classical guidelines, with higher specificity and more mechanistic insights than just for blood–brain barrier permeability. This hybrid protocol can potentially be used for other drug property predictions.

When Machine Learning and Deep Learning Come to the Big Data in Food Chemistry

“…DT method is an open-box model that constructs a binary tree containing decision nodes and branches. It can be used for both classification and regression . While nodes represent the attributes in a group to be classified, branches represent the values taken .…”

“…Multiple decision trees form an ensemble in the RF algorithm and predict more accurate results, mainly when the individual trees are uncorrelated. As an ensemble learning algorithm, the RF method overcomes the main drawback of the DT method, including the overfitting of training data sets . In addition, RF makes it easy to evaluate the resultant model’s variable importance or contribution levels.…”

Re-Routing Drugs to Blood Brain Barrier: A Comprehensive Analysis of Machine Learning Approaches With Fingerprint Amalgamation and Data Balancing

“…As mentioned previously, MACCS is a substructure fingerprint that include predefined atom symbols, bonds, atom properties and environment. Specifically, MACCS provides information like the presence of nitrogen heterocycle which is a major contributing factor to BBB permeability [19]. The presence of this nitrogen heterocycle can further influence physicochemical properties like lipophilicity, polarity and hydrogen bonding capacity.…”

“…Recent years have seen unprecedented applications of AI/ML methods in addressing diverse problems ranging from medical image analysis [14], [15], [16] to drug discovery [17]. Several AI/ML-based models have been proposed to facilitate expeditious CNS drug discovery/repurposing by minimizing the number of laborious and time-consuming BBB permeability studies [4], [18], [19], [20], [21]. Several approaches for the identification and optimal generation of key molecular properties that are involved in BBB permeability have been reported.…”

“…In 2018, Wang and team developed BBB permeability prediction algorithms using SVM, RF, kNN and multilayer perceptron (MLP) neural network for a dataset of 2358 compounds [28]. RF is an ensemble learning model that circumvents the drawback of overfitting seen in the decision tree (DT) model that works by constructing a binary tree of decision nodes for either regression or classification [19]. A recent study employed DL algorithms with 5-fold crossvalidation to minimize redundancies involving undetected overfitting and overestimation.…”