High-resolution view of compound promiscuity

Introduction

Polypharmacology is an emerging theme in drug discovery^1,2. It is generally accepted that drugs often elicit their therapeutic effects through interactions with different targets and the ensuing modulation of multiple signaling pathways. In some therapeutic areas such as oncology, polypharmacology is heavily exploited, for example, through the use of promiscuous ATP site-directed protein kinase inhibitors³. In other areas, such as the treatment of infectious or chronic inflammatory diseases, achieving a high degree of target selectivity of drug candidates plays a major role.

The study of drug polypharmacology has become an important topic in pharmaceutical research^4,5, especially focusing on combined computational and experimental analysis⁵. On the basis of drug-target networks, it was estimated early on that a drug interacts on average with approximately two targets⁴. More recent estimates from computational data analysis suggest that drugs might bind on average to two to seven targets, depending on the primary target families, and that more than 50% of current drugs might interact with more than five targets⁶.

Compound promiscuity as defined herein is the origin of polypharmacology. Promiscuity analysis can be extended from drugs to bioactive compounds through computational mining of currently available activity data. The results of activity data analysis are generally affected by data incompleteness⁷. This potential influence can only be eliminated by reaching the ultimate (and probably elusive) goal of chemogenomics⁸, i.e., testing all compounds against all targets. In the presence of data incompleteness, compound promiscuity rates are likely underestimated. However, it is not certain that further increasing amounts of assay data will indeed significantly alter the currently emerging view of compound promiscuity (vide infra).

Recent studies have generated a differentiated picture of compound promiscuity. The interested reader is also referred to comprehensive reviews of compound promiscuity analysis⁹ and polypharmacology⁶. In this commentary, we summarize key messages from recent promiscuity analysis in a compact format. It is hoped that this summary might be helpful as a reference for further studies.

Key results of compound promiscuity analysis

Public data sources for compound promiscuity analysis discussed herein have been ChEMBL¹⁰, the major repository of compound activity data from medicinal chemistry (currently in May 2013 containing 1,295,510 compounds with a total of 11,420,351 activity annotations), the PubChem BioAssay database¹¹, the major repository of screening data (with more than 3300 confirmatory assays), and DrugBank¹², which currently contains 1518 approved and 5080 experimental drugs.

It is important to note that collecting all activity annotations for a compound reported in the literature including, for example, reporter gene or other cell-based assays is at best providing a measure of assay promiscuity, but not of specific interactions with different targets⁹. Therefore, it is generally required to apply data confidence criteria such as the presence of well-defined activity measurements or evidence for direct ligand-target interactions⁹ (as provided in ChEMBL as activity data filters).

Up-to-date promiscuity rates

In Table 1, current average promiscuity rates are summarized for compounds from ChEMBL, PubChem, and DrugBank. For promiscuity assessment of drugs, all targets reported in DrugBank were considered.

If all compounds with single or multiple target annotations are analyzed, ChEMBL compounds interact on average with one to two targets and PubChem compounds with two to three. However, approved drugs have on average close to six targets. In contrast, the degree of promiscuity of experimental drugs is considerably lower, with less than two targets per drug candidate. If only promiscuous compounds or drugs are taken into account (i.e., if compounds with single target annotations are excluded), promiscuity rates only slightly increase by about one target per compound, the exception being experimental drugs whose average number of targets increases from 1.8 to 4.7. Furthermore, median promiscuity rates were also calculated for promiscuous compounds from different sources, i.e., ChEMBL compounds with activity against at least two targets (K_i and IC₅₀), approved and experimental drugs annotated with more than four or at least two targets, respectively, and PubChem compounds active against at least three targets. Compared to the average promiscuity rates reported in Table 1, the median rates were consistently lower. However, the differences between the average and median rates were small, i.e., less than one for ChEMBL and PubChem compounds. By contrast, differences were larger than one for approved and experimental drugs, i.e., on the basis of median rates, drug target numbers were reduced by 1.9 and 2.7, respectively. Hence, average promiscuity rates for drugs were likely biased by highly promiscuous drugs.

In Table 2, the probability of promiscuity is reported for compounds from different sources (calculated from target distributions of compounds). For a ChEMBL compound with available IC₅₀ and K_i measurements, the current probability of activity against two or more targets is ~25% and ~38%, respectively (if both IC₅₀ and K_i measurements were available for a compound, they were separately considered). However, for activity against more than five targets, the probabilities are reduced to only ~1%. Similar observations are made for confirmed PubChem screening hits (providing an upper-limit promiscuity assessment for bioactive compounds, vide supra). In this case, the probability of activity against two or more, or against more than five targets is ~51% and ~8%, respectively. Furthermore, the probability of promiscuity of approved drugs from DrugBank is ~84% and the probability to interact with more than five targets still ~37%. For experimental drugs, the corresponding probabilities are much lower, with only ~24% and ~3%, respectively.

Compound promiscuity for different target families

All available compounds active against targets belonging to the five target families, including G protein-coupled receptor (GPCR) class A, protein kinases, ion channels, proteases, and nuclear hormone receptors, were assembled from ChEMBL release 14 and separated into K_i and IC₅₀ value-based subsets, as described above. Average promiscuity rates were calculated for all compounds active against a given family as well as compounds active against multiple targets within the family, as reported in Table 3. With the exception of the K_i subset of the ion channel family, promiscuity degrees for compounds active against these target families were similar to those reported in Table 1. In Table 4, the probability of promiscuity (i.e., activity against at least two or more than five targets) is reported for compounds active against these families (according to Table 2). Similar observations were made. A significant relative increase (~10%) in probability of promiscuity was only observed for compounds active against two or more targets from the nuclear receptor family on the basis of the IC₅₀ subset. Thus, for prominent target families, no above-average compound promiscuity rates were detected.

Promiscuity of compounds with increasing molecular weight

The degree of promiscuity was also determined for compounds with different sizes, i.e., molecular weight (MW). Seven subsets of compounds with increasing MW were collected from ChEMBL release 14 and organized into K_i and IC₅₀ value-based subsets, as reported in Table 5. Average promiscuity rates of compounds with increasing MW were found to be comparable to the global rates. However, a significant relative increase in promiscuity was observed for the smallest compounds with MW ≤ 200 in K_i subset. Furthermore, the probability of activity against two or more targets also increased by more than 10% for the smallest compounds in both subsets, as reported in Table 6. For larger compounds across all MW ranges, no significant increases in promiscuity were observed compared to the global degree and probability of compound promiscuity reported in Table 1 and Table 2, respectively.

Conclusions

Herein, we have provided a detailed and up-to-date view of compound promiscuity, the molecular basis of polypharmacology. For active compounds from medicinal chemistry and biological screening sources, the degree of promiscuity is lower than for drugs. There is a notable increase in promiscuity from bioactive compounds over drug candidates to approved drugs. The exploration of possible reasons for this apparent "promiscuity enrichment" along the drug discovery pathway should provide interesting opportunities for future research. On the basis of currently available high-confidence activity data, promiscuity of bioactive compounds is limited (and very low across different target families). However, if compounds are promiscuous, they typically bind to their targets with relatively high potency. Given the overall low degree of promiscuity of bioactive compounds including screening hits in the presence of nearly exponential data growth in recent years, it remains an open question if future chemogenomics efforts might substantially change the current picture of compound promiscuity (vide supra). The majority of available bioactive compounds have single target annotations and we believe it is unlikely that most of them will display a high degree of currently undiscovered promiscuity. Hence, we would also conclude that the target specificity paradigm that has long dominated small molecule discovery efforts should continue to play a major role, despite emerging "anti-reductionism" and the increasing focus on phenotypic readouts.