NO-Classifier: Prediction of Anti- Neuroinflammatory Agents Using Text Mining of 3D Molecular Fingerprints

Abstract

Abstract CNS Drug discovery has been challenging due to the lack of clarity on CNS diseases' basic biological and pathological mechanisms. Despite the difficulty, some CNS drugs have been developed based on phenotypic effects. Herein, we propose a phenotype-structure relationship model, which predicts an anti-neuroinflammatory potency based on 3D molecular structures of the phenotype-active or inactive compounds without specifying targets. For this chemo-centric study, a predictive model of the nitric oxide (NO) inhibitory potency in hyper-activated microglia is built from the 548 agents, which were collected from 95 research articles (28 substructures consisting of natural products and synthetic scaffolds) and doubly externally validated by the agents of 9 research articles as third set. 3D Structures (multi-conformer ensemble) of every agent were encoded into the E3FP molecular fingerprint of the Keiser group as a 3D molecular representation. The location information of the molecular fingerprints could be learned and validated to classify the inhibitory potency of compounds (IC50 cut-off between bi-classes: 37.1 μM): (1) multi-layer perceptron (MLP) (accuracy: 0.962, AUC: 0.994), (2) recurrent neural network (RNN) (accuracy: 0.966, AUC: 0.994), and (3) convolutional neural network (CNN) (accuracy: 0.969, AUC: 0.996). The high performance of these models was compared with that of four classical machine classification models (Logistic, Ridge, Lasso, and Naïve Bayes). We named the bi-class models NO-Classifier. Out-of-set validation and decision region analysis of the out-of-set doubly demonstrated NO-Classifier effectively discerned the anti-inflammatory potency of testing compounds in inflammatory cell phenotype with the webserver in https://no-classifier.onrender.com.