Background estimates and data normalization
To reveal systematic biases, the DIPS scores of drug-induced gene expression profiles using biological controls were classified into four drug and batch categories (). The DIPS scores between different drugs in the same batch are significantly higher than between different drugs from different batches, implying a considerable batch effect as has already been hinted at in the original CMap publication 
(, Label 3). Still, characteristic drug features are reflected in the gene expression profiles, i.e. the DIPS scores between the same drugs (, Blue and Red) are significantly higher than between different drugs from different batches (, Grey) (t-test p-values<2.2×10−16
Analysis of systematic biases and benchmarking with independent features of chemicals.
We utilized the large number of treatments to infer the background gene expression (by mean-centering) instead of the few biological controls provided by CMap, in order to eliminate the batch effect. In this way, also common (e.g. stress) responses will be down-weighted to reveal the characteristic expression response of each chemical perturbation.
After this normalization the batch effect was largely eliminated and the data reflect the characteristic features of drug perturbations better. The DIPS scores between different drugs from the same batch are no longer higher than between different drugs from different batches (, Label 3). Additionally, the DIPS scores between the same drugs from different batches (, Red) are higher than the between different drugs from the same batch (, Yellow), revealing the concordance of drug-induced gene expression profiles across batches (t-test p-values<2.2×10−16
). Same conclusions were also derived from the distributions of Pearson correlations for drug-induced gene expression profiles across four drug/batch categories (Figure S1
Assessment of the drug-induced gene expression profile similarity score
We prove the integrity and reliability of the homogeneous gene expression profiles constructed with mean centering, by employing benchmark sets representing different features of drugs such as chemical structure similarity and shared Anatomical Therapeutic Chemical (ATC) classification of the World Health Organization (WHO) 
. Chemical structural similarity is an indicator of shared drug targets and mechanism of action 
. It is reported that high chemical similarity (i.e. with Tanimoto 2D coefficients >0.85) tends to result in similar biological responses 
. The ATC classification is based on both the therapeutic and chemical properties of the drug also referred to as the drug mode of action. Thus, we expect that pairs of drugs with high structural similarity or shared ATC classification result in similar gene expression profiles.
Benchmarking shows that the DIPS scores calculated using the mean-centered procedure are clearly superior to the method proposed by Iorio et al.
(). The area under the Receiver operator characteristic (ROC) curve (AUC) for the combined DIPS scores for 989 drugs (average over three cell lines), are higher both when using chemical similarity (Tanimoto 2D coefficient >0.8) and the 4th
level of the available ATC code shared between drug pairs (Figure S2
). This confirms that the mean-centered data reflect the specific response after drug perturbation better than the treatment-control comparisons that were used previously 
Finally, the drug induced gene expression profiles were found to be concordant across cell lines (Figure S3
). Although only cancer cell lines were used in CMap, the procedure proposed here should be applicable to drug perturbation profiles across multiple tissues and even organ systems.
Differential expression of drug-induced drug targets
Integrating 4849 CMap arrays with 40,656 drug target relations from STITCH resulted in a set of 1,290 drug-target relations for which a genome-wide cellular response is available. We found that thirteen out of 167 distinct drug targets in this set (8%; 86 drug target relations) are subjected to significant differential expression upon drug treatment () by comparing the drug-induced expression changes of the drug target against all other treatments present in CMap (see methods
). We found supporting evidence in the literature for seven out of thirteen (q-value <0.05) significant differential regulations of drug targets shown in , confirming the rationale and predictive power of our systematic approach. For the remaining six targets we can predict a hitherto unknown drug-induced differential regulation.
Drug-induced differentially regulated drug targets.
The identified, differentially regulated drug targets are enriched in G-protein coupled receptors (GPCRs) (Figure S4
), in agreement with previous reports that members of the GPCR family are generally regulated by several mechanisms including receptor desensitization, endocytosis at the protein level and regulation of the cellular receptor content 
. In the three cancer cell lines used, we observe agonist-induced down-regulation of GPCR mRNAs for beta-2 adrenergic receptor (ADRB2), prostaglandin E2 receptor subtype EP2 and prostaglandin E4 receptor subtype EP4 (, Genes 3,4,12), which were previously reported in DDT1 MF-2 smooth muscle cells (ADRB2) and 293-EBNA human embryonic kidney cells (prostaglandin E2/E4 receptor subtypes EP2/EP4) 
. This indicates that drugs can induce similar feedback loops in a wide variety of cell types. However, we cannot rule out that cross-regulation among signaling pathways may be responsible for the regulation of GPCR mRNAs as it has been described before 
. For example, it has been shown that a beta adrenergic mRNA-binding protein, ELAV-like protein 1 (ELAVL1) can be induced by ADRB2 agonist or elevated levels of cyclic adenosine monophosphate (cAMP) 
and destabilizes ADRB2 mRNA. The ELAVL1 protein binds to GPCR mRNAs and recognizes a cognate sequence located at the 3′-UTR of ADRB2, proteinase-activated receptor and M2, M3 muscarinic acetylcholine receptor mRNAs 
. Therefore cAMP provides cross-talk among GPCR regulatory networks. However, it is shown that intracellular cAMP accumulation is not the only factor contributing to the reduction of ADRB2 mRNA levels 
. Moreover, we find that GPCR-targeting drugs regulate the transcription of their specific targets (Figure S5
). We conclude that in addition to the cross-regulation of G-protein signaling pathways drug target-specific feedback loops are also responsible for the regulation of drug target mRNAs.
In addition to cross-regulation of multiple drugs through the same signaling pathways, promiscuous drugs targeting multiple proteins may cause complex regulatory networks. In order to explore the cross-regulation of drug targets induced by a promiscuous drug, we searched and found that 259 out of 466 total drugs are multi-target drugs and 4 of these drugs act on multiple differentially regulated drug targets upon drug treatment. For example, Podophyllotoxin used in various chemotherapies is known to target both tubulin beta 2C and DNA topoisomerase 2-alpha. The tubulin beta 2C and DNA topoisomerase 2-alpha mRNAs are both down-regulated upon drug treatment in three cell lines (, Genes 8,13). Tubulin beta 2C inhibitors induce microtubule depolymerization that leads to the specific down-regulation of tubulin beta 2C mRNAs preventing the translational synthesis and thus the further accumulation of abundant tubulin monomers 
. Moreover, we found that DNA topoisomerase 2-alpha mRNAs are not down-regulated upon treatment of other tubulin inhibitors (Figure S5
). Therefore, we conclude that feedback loops of tubulin beta 2C and DNA topoisomerase 2-alpha are not cross-regulated. Two other examples of multi-target drugs are vorinostat used for the treatment of cutaneous T cell lymphoma and trichostatin A that serves as an antifungal antibiotic. Vorinostat and trichostatin A are considered to be nonspecific histone deacetylase inhibitors. These drugs lead to the up-regulation of histone deacetylase 3 (HDAC3) and down-regulation of histone deacetylase 7 (HDAC7) (, Genes 1,5). In this case it is unclear whether there is cross-regulation, although HDAC7 siRNA experiments failed to induce the up-regulation of HDAC3 mRNAs 
, a result that is disfavoring the cross-regulation.
While the above cases only confirm literature reports in other cell lines or tissues, we also identified new cases of drug-induced expression regulation of drug targets. The significant novel findings are the inhibitor-induced down-regulation of calmodulin 1, DNA topoisomerase 2-alpha and up-regulation of endoplasmin, lanosterol 14-alpha demethylase and cAMP-specific phosphodiesterase 4D (, Genes 9,8,2,7,10). Lanosterol 14-alpha demethylase is actually an off-target of antifungal drugs that bind the mammalian version with lower affinity than the fungal lanosterol 14-alpha demethylase. Probably, the up-regulation of the mammalian lanosterol 14-alpha demethylase compensates for the undesired inhibition and modulates the adverse effects. On the contrary, we observed a feedback loop that accelerates the down regulation of calmodulin 1 mRNA induced by calmodulin inhibitors. Calmodulin targeting drugs can provide a rapid and effective therapeutic effect, while at the same time small variations of drug concentrations can increase adverse effects. Therefore, it would be interesting to study further the functional effects upon target inhibition of lanosterol 14-alpha demethylase, endoplasmin and calmodulin 1 to elucidate the roles of feedback loops in drug mode of action and adverse effects.
Drug-induced target regulation might be implicated in tolerance development and thus identifying potential target regulation should be an integral part of drug discovery to prevent failures in later stages of clinical trials. For example, we have observed the inhibitor-induced up-regulation of ADRB2 and thymidylate synthetase (TYMS)(, Genes 3,11) 
. TYMS is an essential enzyme for DNA replication/repair and an important drug target in cancerous cells. Indeed, it has been shown that inhibitor-induced TYMS over-expression obstructs the clinical efficiency by inducing tumor drug resistance 
. In addition to over-expression, down-regulation of drug targets upon agonist treatment may also cause treatment tolerance as observed for ADRB2 long-acting agonist treatment. ADRB2 is a therapeutic target activated to treat the symptoms of asthma. We observe the agonist-induced up-regulation of ADRB2 and already in 2005, the FDA warned patients that ADRB2 might be down-regulated (desensitization) and be unresponsive for asthma treatment due to long-acting agonist exposure 
. Thus, robustness in biological systems could prevent the applicability of the long-term treatments via positive/negative feedback loops of the drug target affecting the clinical efficiency of drugs in trial and on the market. Drug-induced regulation of drug targets can thus be linked to tolerance development, which restricts the efficiency of clinical treatments where the drug concentration is limited to avoid an excess of adverse drug reactions.
Taken together, we have identified drug-induced differential regulation of drug targets. Due to the limited signal to noise ratio in the data at hand, the identified 8% of all drug targets that show feedback loops has to be seen as a lower limit, i.e. target-regulation appears as a wide-spread biological phenomenon that has to be taken into account during drug development.