Virtual Screening of Chinese Herbs with Random Forest

Abstract

Random Forest, a form of multiple decision trees, has been used to screen a database of Chinese herbal constituents for potential inhibitors against several therapeutically important molecular targets. These comprise cyclic adenosine 3′-5′-monophosphate phosphodiesterases, protein kinase A, cyclooxygenases, lipoxygenases, aldose reductase, and three HIV targets-integrase, protease, and reverse transcriptase. In addition, compounds were identified which may inhibit the expression of inducible nitric oxide synthase and/or nitric oxide production in vivo. A total of 240 Chinese herbs containing 8264 compounds were screened in silico, including many used on a regular basis in traditional Chinese medicine. Active compounds were selected from another database of 2597 phytochemicals and related natural products with known target affinities and covered a wide range of structural classes. Random Forest was found to perform well, even on highly unbalanced data characteristic of ligand-based screening where the compounds to be screened are far more numerous than the number of active compounds used in training. Despite a conservative screening protocol, a wide variety of compounds from Chinese herbs were hit. Of particular interest were the relatively large number of herbs predicted to inhibit multiple targets, as well as a number which appeared to contain inhibitors of the same target from different phytochemical classes. The latter point to the possibility that individual species may make use of alternative phytochemical strategies in target inhibition. A literature search provided evidence to support 83 herb-target predictions.

Phytochemical Informatics of Traditional Chinese Medicine and Therapeutic Relevance

“…The sesquiterpenes, which are slightly larger than both monoterpenes and the smallest phenolics, are closely associated with Regulate Qi, InVigorate Blood, and Aromatic (Damp) categories (39), though a smaller number also appear to have associations with Wind Damp and Phlegm Cold categories (15). It is intriguing to find parallels again in this respect among small phenolics, in that a fairly large group of these are also associated with Aromatic (Damp), Regulate Qi, and InVigorate Blood categories (38), and these are on average slightly larger compounds than those phenolics which share TCM characteristics with the monoterpenes.…”

“…The reader is referred to these papers for further details. 5,15 In this instance, we interpolated these, and additional scores for 17 other targets, onto the map and identified the major clusters into which they fell. Targets were chosen for which a reasonable number of active compounds are known (over 40).…”

“…In addition, relationships between TCM profile and predicted activities against a number of therapeutically important targets are analyzed. The latter are taken from results which have, in part, been previously published 15 and, despite the limited number of targets, illustrate how such an approach may be used to discover links between ethnopharmacological and molecular data.…”

Screening of herbal constituents for aromatase inhibitory activity

“…Random Forest (RF) models for prediction of aromatase inhibitory activity were derived using activity data taken from the Bio-active Plant Compounds Database (BPCD) [7]. The RF modelling, performed using Random Forests, version 1 (Salford Systems, San Diego, California) was carried out as described by Ehrman et al [11][12]. The MOE, Kier-Hall and Labute descriptor sets were used in generating three separate RF models, each comprising an ensemble of 500 decision trees.…”

“…For each tree in the ensemble, a bootstrap sample was randomly chosen from the minority class of 44 BPCD compounds with known anti-aromatase activity, and the same number of cases then randomly selected to add to this group from the majority class of 8,264 CHCD compounds. Decision trees were then constructed, following the CART algorithm, and employing the Gini splitting criterion (see [11][12]). The error rates associated with the resulting RF predictions were quantified using the technique of "out-of-bag" (OOB) cross-validation, with ~33% of the bootstrap compounds withheld for each tree trained in order to provide an independent test sample.…”