Metformin directly targets the H3K27me3 demethylase KDM6A/UTX

Summary Metformin, the first drug chosen to be tested in a clinical trial aimed to target the biology of aging per se, has been clinically exploited for decades in the absence of a complete understanding of its therapeutic targets or chemical determinants. We here outline a systematic chemoinformatics approach to computationally predict biomolecular targets of metformin. Using several structure‐ and ligand‐based software tools and reference databases containing 1,300,000 chemical compounds and more than 9,000 binding sites protein cavities, we identified 41 putative metformin targets including several epigenetic modifiers such as the member of the H3K27me3‐specific demethylase subfamily, KDM6A/UTX. AlphaScreen and AlphaLISA assays confirmed the ability of metformin to inhibit the demethylation activity of purified KDM6A/UTX enzyme. Structural studies revealed that metformin might occupy the same set of residues involved in H3K27me3 binding and demethylation within the catalytic pocket of KDM6A/UTX. Millimolar metformin augmented global levels of H3K27me3 in cultured cells, including reversion of global loss of H3K27me3 occurring in premature aging syndromes, irrespective of mitochondrial complex I or AMPK. Pharmacological doses of metformin in drinking water or intraperitoneal injection significantly elevated the global levels of H3K27me3 in the hepatic tissue of low‐density lipoprotein receptor‐deficient mice and in the tumor tissues of highly aggressive breast cancer xenograft‐bearing mice. Moreover, nondiabetic breast cancer patients receiving oral metformin in addition to standard therapy presented an elevated level of circulating H3K27me3. Our biocomputational approach coupled to experimental validation reveals that metformin might directly regulate the biological machinery of aging by targeting core chromatin modifiers of the epigenome.

Investigaci on Sanitaria (FIS), Spain. This study was supported also by unrestricted research grants from Roche Pharma (Spain) and Astellas Pharma (Spain) to the Program Against Cancer Therapeutic Resistance (ProCURE, Catalan Institute of Oncology), and the Armangu e family (Girona, Catalonia) to the Metabolism and Cancer Group (Girona Biomedical Research Institute) in cultured cells, including reversion of global loss of H3K27me3 occurring in premature aging syndromes, irrespective of mitochondrial complex I or AMPK. Pharmacological doses of metformin in drinking water or intraperitoneal injection significantly elevated the global levels of H3K27me3 in the hepatic tissue of low-density lipoprotein receptor-deficient mice and in the tumor tissues of highly aggressive breast cancer xenograft-bearing mice. Moreover, nondiabetic breast cancer patients receiving oral metformin in addition to standard therapy presented an elevated level of circulating H3K27me3. Our biocomputational approach coupled to experimental validation reveals that metformin might directly regulate the biological machinery of aging by targeting core chromatin modifiers of the epigenome.

K E Y W O R D S
aging, cancer, chemoinformatics, computational screening, metformin

| INTRODUCTION
The TAME (Targeting Aging with Metformin) clinical trial has been designed to evaluate the capacity of the antidiabetic biguanide metformin to delay the manifestation of age-associated disorders (Barzilai, Crandall, Kritchevsky & Espeland, 2016). By enrolling patients aged 65-79 years diagnosed with one single age-associated condition and then assessing the global impact of metformin on a composite outcome including cardiovascular events, cancer, dementia, mortality, and other functional and geriatric endpoints, this paradigm-shifting study aimed to target the aging process per se (Newman et al., 2016; Figure S1A). If the positive consequences of metformin extend beyond an isolated impact on each separate agerelated disease, the TAME study might pave the way for the development of new healthspan-promoting treatments aimed to promote reduction in age-associated multimorbidity.
Metformin might exert multiple healthspan-promoting effects by correcting deregulated nutrient/energy-sensing metabolic axes such as insulin/IGF-1 and AMPK/mTOR, one of the metabolic hallmarks of aging (L opez-Ot ın, Blasco, Partridge, Serrano & Kroemer, 2013;L opez-Ot ın, Galluzzi, Freije, Madeo & Kroemer, 2016). The ability of metformin to promote such metabolic fitness and operate as an antiaging tool is commonly perceived as the sum of the pleiotropic effects due to its primary action on a single master mechanism. Since the first description of inhibition of mitochondrial energy transfer by guanidines in 1963, mitochondrial complex I (mCI) of the electron transport chain has been commonly viewed as the primary target of metformin (Bridges, Jones, Pollak & Hirst, 2014;Chance & Hollunger, 1963).
Although approved as an insulin-lowering agent for type 2 diabetes or other hyperinsulinemic conditions, the ongoing use of metformin has led to the discovery of unanticipated and multifaceted actions in several chronic conditions. Indeed, metformin is an archetypical example of an approved drug that has been clinically exploited without a complete understanding of its precise therapeutic targets or chemical determinants (Sweeney, Raymer & Lockwood, 2003). Here, using a systematic chemoinformatics approach (Figure S1B) coupled to laboratory-based confirmatory testing, we sought to computationally predict and experimentally validate new biomolecular targets through which metformin might operate as a polytherapeutic tool capable of targeting the biological machinery of aging.

| Structure-based virtual profiling of metformin targets
Thirteen structure-based virtual profiling (VP) predicted targets of metformin (Table 1) were selected based on interaction energies (≤6.1 kcal/mol). Among others, the targets included the metalloenzymes glutamate carboxypeptidase 2 and N-acetylated-alpha-linked acidic dipeptidase 2; purine nucleoside phosphorylase, an essential enzyme of the purine salvage pathway; CAD protein, a rate-limiting trifunctional protein involved in de novo synthesis of pyrimidine nucleotides; arginases 1 and 2, two enzymes involved in L-arginine/ nitric oxide metabolism that promote vascular endothelial inflammation and senescence, atherogenesis, and cardiovascular aging; dynamin-1, a member of the mitochondrial division machinery associated with neurodegenerative disorders; and kallikrein-7, a serine protease involved in skin homeostasis and inflammation.

| Ligand-based virtual profiling of metformin
Twenty-eight ligand-based VP targets were selected based on the similarity of structural and physicochemical properties to metformin (Table 1) The binding energy is obtained during the virtual profiling experiment as it is docking-based. The more negative the binding energy, the more plausible the interaction.
A notable number of DNA-interacting enzymes are listed in the ligand-based VP for metformin (Table 1)

| Docking and molecular dynamics validation of metformin targets
To validate the VP results, we performed in silico binding experiments using rigid docking + short molecular dynamics (MD) calculations, the latter aiming to simulate flexible docking conditions. In total, we performed docking calculations for 33 metformin targets.
Docking calculations of metformin against the selected targets, which were run twice to avoid false positives, revealed low binding energies, ranging from À3.2 kcal/mol to À6.5 kcal/mol (Table S1). To add protein flexibility to the analysis and to test the stability of the selected metformin-target complexes, allowing us to filter out poorly interacting compounds, we carried out short MD simulations of 1 ns. We then performed Molecular mechanics-generalized born surface area (MM/GBSA) calculations to estimate the free energy of the binding of small ligands such as metformin to biological macromolecules (Genheden & Ryde, 2015).
MM/GBSA-based estimation of ligand-binding affinities considers the dynamic nature of the protein and is therefore more reliable to provide a realistic view of metformin-binding affinity than rigid docking estimations. The ranked energies obtained following MM/ GBSA rescoring calculations over MD simulations are summarized in Tables S2 and S3.

| New possible indications and biomolecular targets for metformin
New possible indications of metformin were evaluated by crossing the ligand-and structure-based metformin targets obtained after VP with DisGeNET, a database containing 429,036 associations between 17,381 genes and 15,093 diseases, disorders, and clinical or abnormal human phenotypes (Piñero et al., 2017). A graphical summary of the diseases in which the predicted metformin targets are involved suggested an enrichment of the most prevalent aging-related diseases, including neoplasms/carcinoma, cardiovascular, and neurodegenerative disorders ( Figure S1B). Using the WEB-based Gene SeT AnaLysis Toolkit to statistically evaluate the DisGeNet findings, we confirmed the nonrandom and unbiased distribution of such metformin targetassociated morbidities (Table S4). Gene Ontology (GO) was extracted from UniProtKB database to provide a detailed description of the molecular functions and biological processes of the ligand-and structure-based metformin targets (Table S5).

We then performed a Protein ANalysis THrough Evolutionary
Relationship (PANTHER) overrepresentation test using as a reference list the comprehensive GO annotations directly imported from the GO database that includes 20,972 mapped identifiers. Fourteen different classes of molecular functions were found significantly enriched in the list of metformin targets (Table S6). Beyond the expected overrepresentation of molecular functions related to transmembrane transporter activity, most of the molecular functions were related to metal ion binding and peptidase activity.
Our computational de-orphanization of metformin further revealed that DNA/histone-interacting proteins might be viewed as unforeseen targets for anti-aging metformin. The best-aligned 3D and 2D metformin poses (after MD simulations) and the key interacting residues for metformin binding to such DNA/histone-related targets are illustrated in Figure S2 and Table S7. When considering the activity and phenotype of similar molecules against the DNA/histone-interacting targets of metformin, one could infer that, with the exception of the predicted activation of geminin, metformin mostly behaves as an inhibitor of chromatin structure modifiers (Tables S8 and S9; Figure S3A,B).

| Metformin specifically augments global levels of H3K27me3 irrespective of mitochondrial complex I and AMPK
To test the hypothesis that metformin might target the chromatinmodifying activities of specific KDMs that actively remove well-estab- value among all the biguanides tested (À54.4117 kcal/mol, Figure 4), which was notably higher than those binding energies found when using buformin, phenformin or cycloguanil (that ranged from À13.4961 to À23.3294 kcal/mol), all of them lower than those initially observed with metformin (À26.1494 kcal/mol, Table S3).
AlphaScreen assays confirmed that norMitoMet worked as the most potent inhibitor of the demethylase activity of purified KDM6A/UTX among all the biguanides tested (IC 50 = 0.15 mmol/L, ffi40-fold lower than metformin IC 50 ), whereas phenformin, cycloguanil, and buformin were slightly less potent than metformin at inhibiting the enzymatic activity of KDM6A/UTX (Figure 4b).
On the basis of the similar property principle (Nigsch & Mitchell, 2008), which states that structural similar molecules are more likely to have similar properties and biological activities, we used 2D and 3D molecular fingerprints to disentangle whether biguanides When we compared post-treatment and pretreatment serum, circulating H3K27me3 was found to be significantly augmented in those patients that, in addition to the standard neoadjuvant regimen including anthracycline/taxane-based chemotherapy and trastuzumab, simultaneously received 850 mg metformin twice-daily for 24 weeks before surgery (Figure 6c). Such differential effects of metformin were found in the serum of patients whose tumors responded favorably to treatment (achieving a pathological complete response) when compared to those who did not respond (Figure 6c).

| DISCUSSION
We provide biocomputational evidence to suggest that the capacity of metformin to operate as a polytherapeutic anti-aging tool likely involves diverse mechanisms of action, such as substrate competition phenomena with metabolites or nutrients, metal-interactive regulation of protein functioning (e.g., inflammation-related proteases), genome stability, and epigenome marking and functioning (Figure 6d).
Metformin is a highly hydrophilic drug and is understood to require transporters to cross membranes (Pernicova & Korbonits, 2014). Consistent with this view, our ligand-based VP, involving in silico screening, accurately predicted up to eight different membrane transporters interacting with metformin. These findings support the notion that since transporters involved in uptake and extrusion of metformin were designed by nature for endogenous substrates, the metabolic effects of metformin might involve substrate competition phenomena with metabolites or nutrients (Glossmann & Reider, 2013). Metformin and related biguanides are known to bind endogenous metals such as Zn 2+ , Cu 2+ , Fe 3+ that independently inhibit proinflammatory proteases (Glossmann & Reider, 2013;Lockwood, 2010;Logie et al., 2012;Sweeney et al., 2003;Thorne & Lockwood, 1991). Accordingly, our systematic chemoinformatics approach confirmed that metallopeptidases and proteins with transition metal ionbinding properties were significantly overrepresented among the predicted targets for metformin. These findings lend weight to the notion that metal-interactive regulation of protein functioning should be viewed as a central mechanism of action of metformin and, consequently, that zinc, copper, iron, and other metals might be viewed as primary targets of metformin. The metal-binding properties of metformin appear to be involved not only in its capacity to regulate a number of proteases (Lockwood, 2010;Sweeney et al., 2003), but also to interact with and regulate the activity of a significant number of aging-and cancer-related DNA/chromatin-interacting enzymes.
Interestingly, our systematic chemoinformatics approach coupled to confirmatory experimental testing revealed for the first time that metformin can directly target the chromatin-modifying activity of aging-related histone demethylases such as KMD6A/UTX.

The structural basis for metformin-induced inhibition of H3
Lys27 demethylation might rely on its capacity to interact with the same set of residues that are known to fix the orientation of one of the methyl groups such that it can be subjected to demethylation in the catalytic pocket of KDM6A/UTX (Sengoku & Yokoyama, 2011). Indeed, we verified that an entire family of five pharmacologically relevant biguanides sharing a common functional group interquartile ranges, whiskers, and ranges for H3K27me3 in hepatic tissues obtained in Lldr À/À female mice upon feeding with experimental diets and use for metformin for 14 weeks; CD, chow diet; HFD, high-fat diet. (b) Representative immunoblots for H3K27me3 histone modification in tumor tissues obtained from JIMT-1 xenograft-bearing mice treated with 250 mg/kg/day metformin during 7 weeks (Cufi et al., 2012). Also shown are total H3 controls. (c) Box plots indicating median, interquartile ranges, whiskers, and ranges for circulating H3K27me3 in nondiabetic breast cancer patients who were randomly assigned to receive daily metformin (850 mg twice-daily) for 24 weeks concurrently with 12 cycles of weekly paclitaxel plus trastuzumab followed by four cycles of 3-weekly fluorouracil, epirubicin, cyclophosphamide plus trastuzumab (arm A), or equivalent sequential chemotherapy plus trastuzumab without metformin, followed by surgery (arm B). Pathological complete response (pCR) was defined as absence of invasive tumor cells in breast and axilla following the completion of treatment at the time of surgery. (d) Biomolecular targets for metformin: A four-faceted approach to aging. The ability of metformin to operate as a polytherapeutic tool functioning as a bona fide anti-aging medicine might involve multifaceted mechanisms of action, including substrate competition phenomena with metabolites or nutrients, metal-interactive regulation of protein functioning (e.g., inflammation-related proteases), defense of the integrity/stability of the genome, and epigenome marking/functioning, including the H3K27me3 demethylase KDM6A/UTX small size, however, metformin was capable of simultaneously interacting, in the crystal structure of KDM6A/UTX (3AVS), with the residues required not only for iron binding in the demethylase reaction but also with the methyl groups of the H3K27me3 side chain as well as NOG-an analog of the cofactor AKG (H1146, E1148, Y1135; Sengoku & Yokoyama, 2011;Shpargel, Sengoku, Yokoyama & Magnuson, 2012). Moreover, metformin-induced inhibition of catalysis by isolated mCI, which is believed to be the primary target of metformin, requires even higher concentrations of metformin (i.e., 20-100 mmol/L; Bridges et al., 2014Bridges et al., , 2016 than those required to inhibit the demethylation activity of the purified KDM6A/UTX enzyme. We experimentally confirmed that neither mCI nor AMPK seem to be required to elicit specific augmentation of the global levels of the  (Holm et al., 2012;Wei et al., 2008) closely related to the maintenance of poorly differentiated CSC (Sakaki et al., 2015;Yan et al., 2017) (Jin et al., 2011;Maures, Greer, Hauswirth & Brunet, 2011;McCauley & Dang, 2014;Shah et al., 2013). A decrease in repression-associated H3K27me3 (and an increased activity of the H3K27me3 demethylase KDM6A/UTX) is a key feature of the global chromatin reconfiguration occurring not only in somatic cells during the normal aging process but also in prematurely aging cells in HGPS and WS (Scaffidi & Misteli, 2005;Shah et al., 2013;Shumaker et al., 2006). In addition, landmark observations in Caenorhabditis elegans have linked gain of H3K27me3 (and loss of the H3K27me3 demethylase UTX-1) to extended longevity, strongly suggesting that preserving high levels of H3K27me3 by inhibiting KDM6/UTX may be critical for maintaining youthfulness (Jin et al., 2011;Maures et al., 2011;McCord et al., 2013;Shah et al., 2013). Despite conflicting data from different model systems, there is a trend for increases in activating histone marks (e.g., H3K4m2/3, H3K36me3) and decreases in repressive histone marks (e.g., H3K9m2/3, H3K27me3) indicative of a more actively transcribed genome, which is consistent with a well-recognized open chromatin conformation in aging cells and organisms that culminates in the so-called heterochromatin loss model of aging (Pal & Tyler, 2016

| Structure modeling
The metformin structure was modeled from the CHEMBL database using the CHEMBL 1431 entry. All predicted targets were modeled from their crystal structures with the exception of those for which structures were not available, in which case homology modeling was carried out using SWISS-MODEL.

| Molecular dynamics calculations
Short (1 ns) MD simulations were performed using NAMD version 2.10 over the best-docked complexes, which were selected based on the interaction energy. The Ambers99SB-ILDN and the GAFF forcefield set of parameters were employed for receptors and ligands, respectively. The ligand parameters were obtained using Acpype software, whereas the receptor structures were modeled using the leap module of Amber Tools. Simulations were carried out in explicit solvent using the TIP3P water model with the imposition of periodic boundary conditions via a cubic box. Electrostatic interactions were calculated by the particle-mesh Ewald method using constant pressure and temperature conditions. Each complex was solvated with a minimum distance of 10 A from the surface of the complex to the edge of the simulation box. Na + or Cl À ions were also added to the simulation to neutralize the overall charge of the systems. The temperature was maintained at 300 K using a Langevin thermostat, and the pressure was maintained at 1 atm using a Langevin Piston barostat. The time step employed was 2 fs. Bond lengths to hydrogens were constrained with the SHAKE algorithm. Before production runs, the structure was energy minimized followed by a slow heating-up phase, using harmonic position restraints on the heavy atoms of the protein. Subsequently, the system was energy minimized until volume equilibration followed by the production run without any position restraints.

Surface Area
Molecular mechanics-generalized born surface area rescoring was

CONF LICT OF I NTEREST
The authors declare that they have no conflict of interest.