Predicting cancer targets modulated by Ayurvedic medicines

The recent availability of databases that provide both phenotypic descriptions and the chemical structures of the constituent compounds in traditional Chinese and Indian medicines, have enabled Bender et al  (J. Chem. Inf. Model. 2013 (53) 661 – 673, DOI: 10.1021/ci3005513 , to develop a cool algorithm to predict the mode of action (MOA) of these compounds and to predict novel protein targets for cancer therapies.

Traditional medicine has been utilised by human for thousands of years and normally viewed as complementary or alternative to mainstream therapies.  However, both Chinese and traditional Indian medicine (Ayervedic) have provided us with important drugs for instance Artemisinin an antimalarial drug and reserpine an antihypertensive drug.

From 1981 to 2007, 67% of the pharmaceuticals or new molecular entities (NMEs) introduced into the market were natural product based or a derivative there of.  These natural products often have desirable properties which make them good drugs; they are soluble despite breaking Lipinski’s Rule of Five, they embody privileged structures that are more frequently found to bind a variety of proteins in different organisms, and they are safe and well-tolerated, often having been commonly used for centuries.

However, there are major challenges that need to be resolved that enable the development of a new drug from a traditional medicine.  These include the isolation of the active constituents, the synthesis of the active constituents, the elucidation of the mode of action and finally the development of the compound as a “drug”.

The recent availability of databases that provide chemical structures and their corresponding phenotypes have enabled Bender et al to predict the MOA of compounds found in TCM and Ayurveda addressing one of these major challenges.  First they developed a classifier using bioactive compounds from the ChEMBL database, ChEMBL biological targets, ECFP_4 fingerprints for each compound and a Naïve Bayes classifier.


Figure 1: The compounds were represented using the Extended Connectivity Fingerprints, with a diameter of 4 bonds ECFP_4.  The ECPF is derived from the Morgan algorithm and was implemented in Sitegic’s Pipeline Pilot (Accelrys Inc).  Each atom identifier contains topological information on the atom that includes the number of immediate heavy atoms, the atom’s mass, its charge, the number of hydrogens attached, the valance minus the number of hydrogens and whether it is part of a ring.

This was used to predict which compound (fingerprint) would inhibit each protein target.  Then by creating fingerprints for each traditional medicine compound they could predict which protein targets they would hit.  For example they predicted the protein targets for some of the active ingredients of Panaz ginseng

FP5Figure 2:  Predicted targets for some of the active ingredients in Panax Ginseng

Next they correlated different proteins targets with different phenotypes. Predicting which molecular targets were modulated by the compounds in each different phenotype.  This enabled them to identify the protein targets most frequently modulated by Ayurvedic medicines, with possible anti-tumour effects.  The 10 most enriched protein targets are shown in the table below. The progesterone receptor currently has over 10 inhibitors with FDA approval.  Other proteins identified by this methods include regulators of other well-known cancer targets.


Figure 3:  Top 10 cancer targets in predicted to be inhibited by Ayurvedic medicines.

Practical recommendations – virtual screening hit selection and optimisation based on a literature analysis

Virtual screening is a part of modern drug discovery and has received much attention in the literature. The publications to date however largely focus on case studies or descriptions of methodologies and there are no reported views or guidelines on hit selection or optimisation critieria which exist for conventional screening. A recent publication (Zhu et al – Hit Identification and Optimization in Virtual Screening: Practical Recommendations Based on a Critical Literature Analysis ) highlights this issue and seeks to address the gap based upon a thorough review of the literature from 2007 to 2011.

The groups finding are interesting and their conclusions valuable to the practising medicinal chemist.

They found that whilst potency criteria varied the majority of groups employed potency cut offs in the low to mid micromolar range(1-100uM) similar to groups engaged in hit identification via conventional HTS.Another perhaps more disturbing trend observed was that groups were not using physicochemical filters or structural alerts as widely as might expected And they recommend the inclusion of ligand efficiency criteria and application of a structural alert filter.This observation is perhaps some what surprising given the number of recent publications highlighting the dangers of high lipophilicity etc and it would seem that too many chemists are still being seduced by potency and ignoring other important considerations. Another interesting observation is that typical hit rates for virtual screens are highly than those for high throughput screens.

The authors conclude by making some simple and logical recommendations for workers to use in association with future virtual screening campaigns amongst them they emphasise the need to include physico – chemical property filters and further qualify this by pointing out that certain newer target classes may require higher molecular weight ligands

Novel Rat model for Alzheimer’s disease

It is stating the obvious that having good animal models is critical to the success of any drug discovery program. In many more complex diseases however, good animal models are not available. The ‘gold standard’ animal models for Alzheimer’s disease, Aβ-overproducing transgenic AD mice; do not demonstrate robust tauopathy and subsequent neuronal loss without the addition of genes not linked to familial AD.

In a recent paper Cohen et al., (1) have generated transgenic rats bearing human mutant APP (amyloid precursor protein) and PS1 (presenilin 1). These animals appear to manifest the full spectrum of age-dependent Alzheimer’s disease pathologies alongside cognitive disturbances. They have age-dependent β-amyloid deposition as well as intraneuronal Aβ1-42 and soluble Aβ oligomers. Many mouse models do present with some tauopathy, however, they do not present with neurofibrillary tangles (NFT) as observed in human AD. In this rat model however, they identified striking tauopathy. As well as hyperphosphorylated Tau, structures reminiscent to NFTs were identified close to β-amyloid plaques in aged rats. In addition immunostaining revealed structures consistent with NFTs in 16 month old rats. These NFT-like structures were also frequently observed in areas without plaques, as is found in human AD.

In concert with the molecular pathology, these transgenic rats exhibited neuronal loss and neuronal degeneration that was progressive and age-dependent. There was also an inverse correlation between the neuronal numbers and Aβ1-42 abundance. TUNEL staining indicated the presence of nicked DNA and measurements of active caspase-3 suggested the neurons were apoptosing.  This neuronal loss paralleled changes in behavioural characteristics such as novel object recognition (which is a hippocampal-dependent measure of working memory) that was significantly impaired in older transgenic animals. This was repeated in the Barnes maze, where there were no difference between wild-type and transgenic animals at 6 months, but after 15 months the transgenic animals made significantly more errors than wild-type.

With recent late-stage failures of treatments for Alzheimers this new animal model opens up the possibility to test novel therapeutics in a more human disease-like model.