JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 ª 2019 THE AUTHORS. PUBLISHED BY ELSEVIER ON BEHALF OF THE AMERICAN COLLEGE OF CARDIOLOGY FOUNDATION. THIS IS AN OPEN ACCESS ARTICLE UNDER THE CC BY-NC-ND LICENSE (http://creativecommons.org/licenses/by-nc-nd/4.0/). PRECLINICAL RESEARCH Unraveling the Molecular Mechanism of Action of Empagliflozin in Heart Failure With Reduced Ejection Fraction With or Without Diabetes Oriol Iborra-Egea, MSC,a,b,c Evelyn Santiago-Vacas, MD,a,b,c Salva R. Yurista, MD,d Josep Lupón, MD, PHD,a,b,c Milton Packer, MD,e Stephane Heymans, MD,f Faiez Zannad, MD,g Javed Butler, MD,h Domingo Pascual-Figal, MD,i Antonio Lax, PHD,i Julio Núñez, MD, PHD,j Rudolf A. de Boer, MD, PHD,d Antoni Bayés-Genís, MD, PHDa,b,c VISUAL ABSTRACT Iborra-Egea, O. et al. J Am Coll Cardiol Basic Trans Science. 2019;4(7):831–40. HIGHLIGHTS  Using artificial intelligence, followed by in vivo validation, this study identified the key cardiac mechanism of action of empagliflozin in heart failure in patients with or without diabetes mellitus.  The most robust mechanism of action involved the NHE-1 co-transporter with 94.7% accuracy.  NHE-1 blockade by empagliflozin administration in rats restored the antiapoptotic activity of XIAP and BIRC5.  The beneficial reduction in cardiomyocyte cell death after empagliflozin treatment is independent of the presence of diabetes mellitus.  Empagliflozin could emerge as a new treatment for heart failure patients regardless of their glycemic status. ISSN 2452-302X https://doi.org/10.1016/j.jacbts.2019.07.010 832 Iborra-Egea et al. JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Empagliflozin in Heart Failure NOVEMBER 2019:831–40 ABBREVIATIONS SUMMARY AND ACRONYMS ANN = artificial neural network The mechanism of action of empagliflozin in heart failure with reduced ejection fraction (HFrEF) was deciphered DM = diabetes mellitus using deep learning in silico analyses together with in vivo validation. The most robust mechanism of action HF = heart failure involved the sodium-hydrogen exchanger (NHE)-1 co-transporter with 94.7% accuracy, which was similar for HFrEF = HF with reduced diabetics and nondiabetics. Notably, direct NHE1 blockade by empagliflozin ameliorated cardiomyocyte cell ejection fraction death by restoring expression of X-linked inhibitor of apoptosis (XIAP) and baculoviral IAP repeat-containing MI-HF = post-infarct heart protein 5 (BIRC5). These results were independent of diabetes mellitus comorbidity, suggesting that empa- failure gliflozin may emerge as a new treatment in HFrEF. (J Am Coll Cardiol Basic Trans Science 2019;4:831–40) NHE = sodium-hydrogen © 2019 The Authors. Published by Elsevier on behalf of the American College of Cardiology Foundation. This is an exchanger open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). RNAseq = RNA sequencing B SGLT2i = sodium-glucose co- transporter 2 inhibitor oth type 1 and 2 diabetes mellitus stroke); and b) HF-related hospitalizations; death (DM) are major risk factors for the from cardiovascular causes; and death from any cause development of cardiovascular dis- (3). Similar results with other SGLT2is have been re- eases, increasing morbidity and mortality (1). For de- ported subsequently by both randomized clinical tri- cades, long-term clinical studies of DM have shown als and observational studies (4,5). value in reducing glycemic levels without a substan- SEE PAGE 841 tial effect on cardiovascular outcomes and certainly no benefit in heart failure (HF) (2). The advent of a SGLT2 receptors are located in the proximal renal new class of agents, sodium-glucose co-transporter 2 tubule and are responsible for 90% of glucose reab- inhibitors (SGLT2i), heralds a new era and may repre- sorption into the bloodstream. Inhibition of SGLT2 sent a turning point. Indeed, in the EMPA-REG receptors causes glucosuria and reduces glycemic OUTCOME (Empagliflozin Cardiovascular Outcome levels (6). However, control of glycemic levels in Event Trial in Type 2 Diabetes Mellitus Patients) response to SGLTis alone seems insufficient to study, patients with type 2 DM who were at high explain the reported cardiovascular benefits. In fact, risk for cardiovascular events and who received the EMPA-TROPISM (Are the “Cardiac Benefits” of empagliflozin in addition to standard care showed Empagliflozin Independent of Its Hypoglycemic lower rates of the following compared to those Activity? [ATRU-4]) trial (NCT03485222) is an who received placebo: a) primary composite cardio- ongoing trial that aims to elucidate whether the vascular outcome (death from cardiovascular causes, benefits obtained in the EMPA-REG OUTCOME trial nonfatal myocardial infarction (MI) or nonfatal were mediated, at least in part, by a glucose- From the aHeart Institute, Hospital Universitari Germans Trias i Pujol, Badalona, Spain; bDepartment of Medicine, Universitat c Autònoma de Barcelona, Barcelona, Spain; Centro de Investigación Biomédica en Red Enfermedades Cardiovascualres (CIBERCV), Madrid, Spain; dDepartment of Cardiology, University Medical Center Groningen, University of Groningen, Groningen, the Netherlands; eBaylor Heart and Vascular Institute, Baylor University Medical Center, Dallas, Texas; fMaastricht University Medical Centre, Cardiovascular Research Institute Maastricht, Maastricht, the Netherlands; gCentre d’Investigation Clinique Plurithématique 1433, INSERM U1116, Université de Lorraine, Centre Hospitalier Régional et Universitaire de Nancy, French Clinical Research Infrastructure Network (F-CRIN), Investigation Network Initiative-Cardiovascular and Renal Clinical Trialists h (INI-CRCT), Nancy, France; Department of Medicine, University of Mississippi, Jackson, Mississippi; iDomingo Hospital Universitario Virgen de la Arrixaca, University of Murcia, Centro Nacional de Investigaciones Cardiovasculares, CIBERCV, Murcia, Spain; and the jCardiology Department, Hospital Clínico Universitario, CIBERCV, INCLIVA, Universitat de València, València, Spain. Dr. Yurista is supported by Indonesia Endowment Fund for Education grant LPDP 20150722083422. Dr. Packer is a consultant with Abbvie, Amgen, Akcea, AstraZeneca, Bayer, Boehringer Ingelheim, Cardiorentis, Gilead, Johnson and Johnson, NovoNordisk, Pfizer, Sanofi, Synthetic Biologics, Relypsa, and Theravance. Dr. Zannad is a member of the Boehringer Ingelheim steering committee; and is a consultant for AstraZeneca, Mundipharma, NovoNordisk, and Merck. Dr. Butler is a consultant with Abbott, Adrenomed, Amgen, Array, AstraZeneca, Bayer, BerlinCures, Boehringer Ingelheim, Bristol-Myers Squibb, Cardiocell, Corvidia, CVRx, G3 Phar- maceutical, Innolife, Janssen, Lantheus, LinaNova, Luitpold, Medscape, Medtronic, Merck, Novartis, NovoNordisk, Relypsa, Roche, Sanofi, StealthPeptide, SC Pharma, V-Wave Limited, Vifor, and ZS Pharma. Dr. Núñez has received speaker fees from Boehringer Ingelheim. Dr. De Boer holds equity in scPharmaceuticals; has received honoraria from Mandalmed, Novartis, and Servier; and has received research support through his institution from AstraZeneca, Abbott, Bristol-Myers Squibb, Novartis, Roche, Trevena, and ThermoFisher GmbH. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose. The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the JACC: Basic to Translational Science author instructions page. Manuscript received May 20, 2019; revised manuscript received July 9, 2019, accepted July 9, 2019. JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Iborra-Egea et al. 833 NOVEMBER 2019:831–40 Empagliflozin in Heart Failure F I G U R E 1 Flow Chart Depicting the Machine Learning Approach Used in This Study See Supplemental Video 1. DM ¼ diabetes mellitus; GEO ¼ Gene Expression Omnibus database; HF ¼ heart failure; TPMS ¼ therapeutic performance mapping system. independent mechanism (7). A recent viewpoint full-length articles in PubMed was conducted to postulated the hypothesis that the benefit of SGLTis obtain information about HF and DM. When specific in HF may be mediated by the sodium-hydrogen molecular information describing the condition was exchanger (NHE) rather than by the effect on found, the articles were thoroughly reviewed to glucose reabsorption (8); however, a comprehensive identify protein/gene candidates to feed the mathe- mechanistic explanation of the extra-renal car- matical models. When the evidence of the implication dioprotective SGLTi effects remains elusive. of a candidate in the condition was judged inconsis- Machine learning is an interdisciplinary field based tent enough to be assigned as an effector, an addi- on computational and mathematical models that aims tional search was performed specifically for the to unravel key interactions within complex biological candidate, including all the protein names according networks (9). It is particularly well suited for inves- to UniProtKB (UniProt Knowledge Base, Geneva, tigating the mechanisms underlying the effects of Switzerland). No text-mining tools were used to avoid drugs, including SGLTis. Accordingly, this study used intrinsic technical biases, and only English-language deep learning analysis to investigate what the key car- articles were included, which accounted for more diac mechanisms may be by which empagliflozin exerts than 6,509 items reviewed. its effects in HF and reduced ejection fraction (HFrEF) in Moreover, empagliflozin pharmacological data and patients with or without DM (Figure 1). Machine molecular relationships described for HF and DM learning-driven pathways were validated in an in vivo were extracted by integrating massive, publicly available empagliflozin-treated post-infarct HF (MI-HF) rat model. databases such as Drugbank, Reactome, MINT, and Bio- METHODS Grid. All this information accounted for more than 350,000 interactions (Figure 1). Briefly, by integrating MACHINE LEARNING IN SILICO ANALYSES. Molecular published data, a set of molecular profiles characterizing characterization of HF and DM pathology and drug HF and DM were defined, which were used to build the treatments. First, extensive and careful review of protein network and the mathematical model. 834 Iborra-Egea et al. JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Empagliflozin in Heart Failure NOVEMBER 2019:831–40 C o m p i l a t i o n o f t r a n s c r i p t o m i c d a t a . Next, an mathematical models to form complex relationships RNA sequencing (RNAseq) dataset (public data avail- between datasets or to find patterns in the data able in Gene Expression Omnibus [GEO], accession (14,15). The system built an ANN, which was trained GSE57345 and subseries GSE57338 and GSE26887) of with 1,000 iterations. During this process, the models myocardial tissue data derived from HFrEF patients generated different MoAs, which progressively with (10) or without DM (11) was thoroughly analyzed increased in accuracy by correcting itself with to model the molecular effects of empagliflozin in already-known clinical and biomedical information, each scenario. Gene information was mapped one-to- and cross-validation with independent sets of data one to proteins for introduction into the protein (Figure 1, Supplemental Video 1). Eventually, the al- network, which included 280 proteins of interest gorithms reached a plateau which yielded the final (Supplemental Table 1). This collection of experi- accuracy performance (maximum ¼ 100%) to explain mental pathophysiological signals was used as a list of the MoA of empagliflozin in HFrEF (Figure 1). principles, termed a “truth table,” which apply to a spe- POST-INFARCT HEART FAILURE IN VIVO RAT cific condition (i.e., some proteins are overexpressed MODEL. A total of 74 rats were used for the experi- only in HF patients with DM) (Figure 1) and allow delin- mental validations. The nondiabetic, post-infarct eation of the molecular behavior of each situation (12). heart failure (MI-HF) study was performed in male Generation of mathematical m o d e l s . Finally, Sprague Dawley rats weighing 250 to 280 g, obtained artificial neural network (ANN)-based analyses were from Envigo (Huntingdon, United Kingdom). The used to identify the most relevant pathophysiological diabetic MI-HF study (MI-HF-DM) was performed in processes implicated in HF and DM. To train the male Wistar rats weighing 160 to 200 g obtained from mathematical models, this large collection of well- Envigo (Barcelona, Spain). established pathophysiological signals and clinical Animals were housed in groups under standard information relevant to the pathologies under study laboratory condition of 12L:12D cycle. All animals was reviewed. The molecular description of these received food and water ad libitum during the study input-output signals was investigated as follows: (n ¼ 7 to 12 per group). All experiments were con- ducted in accordance with U.S. National Institutes of 1. Model inputs, for example, information about Health Guide for the Care and Use of Laboratory An- drugs, pathologies, and protein/gene relationships imals and were approved by the Committee for Ani- provided by public databases to assess inhibition mal Experiments. or activation of 1 or more nodes of the protein Induction of diabetes mellitus in Wistar network (their targets) triggering a perturbation r a t s . DM was induced by a single intraperitoneal in- through the system. jection of streptozotocin (STZ) (50 mg/kg of body 2. Model outputs, for example, experimental RNAseq weight) (catalog number S0130, Sigma-Aldrich, St. data (upregulated or downregulated genes/pro- Louis, Missouri) dissolved in freshly prepared 0.1 M teins) and clinical information. citrate buffer (pH 4.5, catalog number 71402, Sigma- 3. Hidden mechanisms of action (MoAs). Using the Aldrich). In order to prevent hypoglycemia in the model’s input and output data, mathematical al- first 24 h after STZ injection, rats were allowed to gorithms were generated to trace a change or have free access to water with 5% glucose. Three days perturbation from 1 to another, elucidating the after STZ injection, rats with blood glucose levels MoAs that would explain this connection. >300 mg/dl were considered diabetic. The models are then able to weight the relative I n d u c t i o n o f m y o c a r d i a l i n f a r c t i o n . Initially, ani- value of each protein (node) relationship. mals were anesthetized using intraperitoneal keta- Therapeutic performance mapping system mine (75 mg/kg) and medetomidine (0.5 mg/kg) t e c h n o l o g y . Details about how therapeutic perfor- before being intubated and ventilated under 2% iso- mance mapping system (TPMS) (Anaxomics, Barce- flurane anesthesia. Rats were randomly allocated to lona, Spain) machine learning analyses work have MI or sham surgery, as described elsewhere (16). been reported previously (13). Briefly, TPMS incor- Left-sided thoracotomy was performed by a small porated massive amounts of biological information incision between the third and fourth intercostal relevant to both the disease(s) and the drug under spaces. The pericardial sac surrounding the heart was study, drawn from curated reports and from public cut open, but the heart was not exteriorized. The and private databases (Figure 1). Next, TPMS adopted infarction was performed by ligation of the left the gene expression dynamics stored in the truth anterior descending coronary artery. Visible blanch- table as a reference framework to shape all the bio- ing and cyanosis of the anterior wall of the left logical information. Finally, TPMS generated ventricle and swelling of the left atrium were taken as JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Iborra-Egea et al. 835 NOVEMBER 2019:831–40 Empagliflozin in Heart Failure indicative of successful ligation. Sham surgery was T A B L E 1 Mechanisms Involved in Heart Failure and Associated Conditions Considered in identical to the MI surgery but without any ligation. the Mechanistic Study E x p e r i m e n t a l d e s i g n a n d s t u d y p r o t o c o l . Two Pathology Pathological Mechanisms Proteins Implicated* weeks after surgery, nondiabetic rats (n ¼ 40) were Heart failure 106 fed either with standard rat chow (product R/M-H Heart hypertrophy 33 V1534-70, Ssniff-Spezialdiäten, Soest, Germany) or Cardiomyocyte cell death 47 chow containing empagliflozin (product BI 10773, in Inefficient myocardial fuel metabolism 7 concentration of 200 mg/kg per day to reach a dose of Oxidative stress 8 30 mg/kg per day), and this was continued for a Inflammation 11 period of 12 weeks. After 12 weeks, hearts were Obesity 140 rapidly excised, weighed, and processed for further Hyperphagia and dysregulated appetite 135 White adipose tissue formation 5 analysis. Animals were divided into 4 groups: group 1 Hypertension 42 (n ¼ 7) were sham rats receiving vehicle treatment; Retention of sodium 42 group 2 (n ¼ 10) were sham rats receiving empagli- Hyperuricemia 16 flozin; group 3 (n ¼ 12) were infarcted rats receiving Decreased renal excretion of uric acid 16 vehicle treatment; group 4 were infarcted rats Values are n. Bold numbers indicate the total number of proteins implicated in the general condition, later receiving empagliflozin. dissected by more specific motives. *Number of proteins implicated in each mechanism. Immediately after diabetes was induced, the Wis- tar rats were treated with empagliflozin (10 mg/kg per day) (n ¼ 34). This treatment was maintained for adjusted p value of <0.01 and log fold changes of 4 weeks before MI was induced and was continued for >0.25 were considered. In the in vivo analysis, gene a period of 4 weeks after MI. After 4 weeks, hearts expression was determined by correcting the refer- were rapidly excised, weighed, and processed for ence gene (36B4), and the calculated values were further analysis. Animals were divided into 4 groups: expressed relative to the control group per experi- group 1 (n ¼ 8) were diabetic sham rats receiving ment. The 2-DD CT method for comparing relative phosphate buffered saline (PBS) for 56 days; group 2 expression results between treatments in real-time (n ¼ 9) were diabetic sham rats receiving empagli- PCR was applied. To compare normally distributed flozin for 56 days; group 3 (n ¼ 7) were diabetic parameters, one-way analysis of variance (ANOVA) infarcted rats receiving PBS for 56 days; group 4 followed by Tukey post hoc test was used. When data (n ¼ 10) were diabetic infarcted rats receiving empa- were not normally distributed, a nonparametric gliflozin for 56 days. Kruskal-Wallis test followed by a Mann-Whitney R e a l t i m e P C R . Total RNA from tissues were isolated U test with correction for multiple comparisons was by the TRIzol RNA isolation protocol. cDNA was used. To compare empagliflozin with vehicle treatment, further isolated from RNA by using the QuantiTecht an independent t-test or a Mann-Whitney U test was RT kit (Qiagen, Venlo, the Netherlands), following the used where appropriate. Two-sided test results yielding manufacturer’s instructions. Quantitative real-time a p value of <0.05 were considered statistically signifi- polymerase chain reaction (qRT-PCR) was performed cant. SPSS statistics (version 23.0, IBM, Armonk, New in a model CFX384 real-time system (Bio-Rad, Phila- York) for Windows (Microsoft, Redmond, Washington) delphia, Pennsylvania) using SYBR Green mix was used to perform all statistical analysis. (Thermo Fisher Scientific, Waltham, Massachusetts) was used to determine the relative gene expression. RESULTS Gene expression was determined by correcting the reference gene (36B4), and the calculated values are EMPAGLIFLOZIN ACTS UPON A SPECIFIC expressed relative to the control group per experi- HF-RELATED PATHOLOGICAL SIGNATURE. First, ment. The primers for BIRC5 and XIAP qRT-PCR used the protein signatures of HFrEF and empagliflozin in this study can be found in Supplemental Table 2. were defined and an ANN analysis identified the DATA INTEGRATION AND STATISTICAL ANALYSIS. processes most likely associated with the beneficial The RNAseq data were compiled and processed using effects of empagliflozin observed in HFrEF the GEO2R tool, using the neqc method for normali- (Supplemental Table 1). Obesity, hypertension, and zation, allowing identification of differential expres- hyperuricemia contributed the most to the HF path- sion levels and calculation of fold changes. The p ological signature and encompassed the possible ef- values obtained for each RNA probe were adjusted fects of empagliflozin, with 140, 42, and 16 proteins of using the Benjamini-Hochberg false discovery rate at interest, respectively. For each of these 3 disorders, a a significance level of 0.01. Only genes with an series of mechanisms and the involved proteins were 836 Iborra-Egea et al. JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Empagliflozin in Heart Failure NOVEMBER 2019:831–40 HF: the most related motives are energy- T A B L E 2 Proteins Known to Be Modulated by Empagliflozin inefficient myocardial fuel metabolism, oxidative Gene Name Protein Name Reference (PMID) stress, and inflammation. STAT3 Signal transducer and activator of transcription 3 29311992 NHE1 IS RESPONSIBLE FOR THE CARDIAC EFFECTS NOS2 Nitric oxide synthase, inducible 29311992 OF EMPAGLIFLOZIN IN HFrEF WITH OR WITHOUT IL6 Interleukin-6 29311992 DM. By using the differentially expressed proteins BDH1 D-beta-hydroxybutyrate dehydrogenase, 27289126 mitochondrial found to be related to HF and the specific motives IFNG Interferon gamma 29311992 affected during the pharmacological treatment, the ALDH2 Aldehyde dehydrogenase, mitochondrial 29311992 mathematical models determined the molecular GCG Glucagon 26590679 mechanisms involved in the heart-focused beneficial INS Insulin 27289126 effects of empagliflozin relative to heart failure in DM ACE2 Angiotensin-converting enzyme 2 26880444 BDNF Brain-derived neurotrophic factor 25344694 and non-DM patients. HDAC1 Histone deacetylase 1 27829948 The most robust MoA identified in this study HDAC2 Histone deacetylase 2 27829948 involved the cardiac sodium/hydrogen exchanger 1 HDAC3 Histone deacetylase 3 27829948 (NHE1) with 94.7% accuracy (Figure 2, Supplemental HDAC8 Histone deacetylase 8 27829948 Table 3), which was similar for DM and non-DM pa- tients, indicating a DM-independent mechanism PMID ¼ unique identifier number used in PubMed for each article. (Table 3). This approach describes mainly the miti- gation of cell death but also the prevention of heart established (Table 1). The efficacy analysis identified hypertrophy and the improvement upon an ineffi- the 3 following complementary strategies: 1) Consider cient myocardial fuel metabolism as the potential sodium/glucose cotransporter 2, Na(þ)/glucose heart-mediated mechanisms by which empagliflozin cotransporter 2 (SGLT2); 2) consider sodium/ may counteract heart failure. hydrogen exchangers 1 and 3 (NHE1 and NHE3); and 3) Direct empagliflozin-driven NHE1 blockade considering empagliflozin bioflags; those proteins a m e l i o r a t e s c a r d i o m y o c y t e c e l l d e a t h . In this known to be modulated by the drug (Table 2). scenario, the amelioration of cardiomyocyte cell ANN EVALUATES THE RELATIONSHIPS AMONG death was established as the main driver of empa- PROTEIN SETS INSIDE THE NETWORK, PROVIDING A gliflozin beneficial effects in HFrEF. In the absence of PREDICTIVE SCORE. Once the pathological signa- empagliflozin, protein kinases B (AKTs; referred to as tures were identified, the possible relationships be- AKT1 to -3 in Figure 2) were identified as downstream tween empagliflozin and heart failure, including effectors of NHE1 activation responsible for the in- associated processes, were evaluated by ANN ana- duction of baculoviral IAP repeat-containing protein 2 lyses. Thus, the study of the MoA of empagliflozin (BIRC2), which in turn induced degradation of was focused on the following specific pathways proteasome-mediated X-linked inhibitor of apoptosis (motives) of the diseases affected by the treatment: (XIAP) and BIRC5 and prompted the progression of 1) None of the evaluated diseases and motives are HFrEF. In empagliflozin-enriched models, by inhib- strongly related to SGLT2 according to the estab- iting NHE1, repressed AKT1-3 and BIRC2 allowed the lished criterion (p > 0.3). expression of the antiapoptotic mediators XIAP and 2) Considering NHE1 and NHE3 as effectors, 2 condi- BIRC5, ultimately halting HFrEF progression. tions appear to be highly related (>75%) to empa- To validate these data obtained in silico, mRNA gliflozin’s complete target profile: levels of XIAP and BIRC5, as last effectors of the Hypertension: retention of sodium is the main pathway, were examined in vivo in an empagliflozin- motive identified, with a predictive score of 80%. treated MI-HF rat model. Confirming in silico find- HF: heart hypertrophy and cardiomyocyte cell ings, gene expression of XIAP and BIRC5 fell 45% and death are the main motives identified, with a 36%, respectively, in vehicle-treated MI-HF animals predictive score of 75.6%. compared to sham controls (p < 0.01 for both, 3) Three conditions appear to be highly related respectively). Remarkably, empagliflozin-treated MI- (>75%) to empagliflozin’s bioflags: HF animals without diabetes showed upregulation of Hyperuricemia: decreased renal excretion of uric XIAP and BIRC5 mRNA expression similar to that in acid is the motive most related to empagliflozin, sham controls and significantly higher than vehicle- with 77% of predictive score. treated MI-HF counterparts (p ¼ 0.003 and p ¼ 0.05, Obesity: the most related motives are hyperphagia respectively) (Figure 3). More importantly, these and dysregulated appetite and white adipose changes in gene expression occurred ubiquitously tissue formation. regardless of DM presence, further strengthening the JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Iborra-Egea et al. 837 NOVEMBER 2019:831–40 Empagliflozin in Heart Failure F I G U R E 2 NHE1-Identified Signaling Pathways in Cardiomyocytes in Heart Failure Each relationship represents a mechanism that may directly or indirectly (via downstream effectors) impact on HF, either through the activation (green arrows) or inhibition (red arrows) of downstream proteins. AKT1 ¼ RAC-alpha serine/threonine-protein kinase 1; AKT2 ¼ RAC-beta serine/threonine-protein kinase 2; AKT3 ¼ RAC- gamma serine/threonine-protein kinase 3; ATP ¼ adenosine triphosphate; BIRC2 ¼ baculoviral IAP repeat-containing protein 2; BIRC5 ¼ baculoviral IAP repeat- containing protein 5; LEP ¼ leptin; MAPK1 ¼ mitogen-activated protein kinase 1; MAPK3 ¼ mitogen-activated protein kinase 3; MC ¼ mitochondria.;NHE1 ¼ sodium/ hydrogen exchanger 1; NOS2 ¼ nitric oxide synthase ¼ inducible; PCr ¼ phosphocreatine. ; PDH ¼ pyruvate dehydrogenase; RPTOR ¼ regulatory-associated protein of mTOR; SR ¼ sarcoplasmic reticulum; TCA ¼ tricarboxylic acid; XIAP ¼ E3 ubiquitin-protein ligase XIAP. in silico predictions. Indeed, empagliflozin-treated exerted an effect on the myocardium independent of MI-HF-DM animals showed upregulation of XIAP its inhibition of SGLT2 in the kidneys, thereby and BIRC5 mRNA expression similar to sham controls benefiting HF patients (6). The exact mechanism re- and significantly higher than vehicle-treated MI-HF- mains unknown, but it was postulated to have a DM counterparts (p ¼ 0.01 and p ¼ 0.005, respec- direct action on the heart. The present study used tively) (Figure 3). Thus, empagliflozin-driven NHE1 artificial intelligence and machine learning, further inhibition ultimately counteracted reduced function validated in an in vivo animal model, to investigate of XIAP and BIRC5. whether empagliflozin could exert direct effects on Additionally, the analyses also suggested that the heart and whether these effects were the same for empagliflozin could further ameliorate car- patients with and without DM. diomyocyte cell death by inhibiting the AKT- This study has 2 main findings. First, these results dependent regulation- associated mTOR protein validate the fact that empagliflozin interacts with the (RPTOR) and by down-regulation of nitric oxide syn- cardiac Naþ/Hþ exchanger NHE1 directly and iden- thase (NOS2)-inducible actions (Figure 2). tifies a mechanistic pathway acting primarily by reducing cardiomyocyte cell death, the main effector DISCUSSION of its cardioprotective effects. Indeed, the activity of cardiac NHE1 seems to be increased in HFrEF pa- The EMPA-REG OUTCOME trial reported that empa- tients, and previous data in experimental models gliflozin exerted cardiovascular benefits which did (17–21) showed that NHE1 inhibition attenuated car- not depend on its effects on blood glucose. This diomyocyte injury, remodeling, systolic dysfunction, observation led to the hypothesis that empagliflozin and ultimately HF. Second, these data suggest that 838 Iborra-Egea et al. JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Empagliflozin in Heart Failure NOVEMBER 2019:831–40 XIAP expression in neonatal rat cardiomyocytes, T A B L E 3 Mathematical Models Generated in This Study* which inhibited caspase-3/7 activity, showing a Mathematical Models Stimulus Response Restrictions reduction of myocardial injury. Relative to BIRC5, Lee HF with DM Cardiac Cardiomyocyte cell death Truth Table et al. (26) found that overexpression of this pathway NHE1 Hypertrophy þ Inflammation GSE57345 subseries GSE57338 dataset may attenuate the progression of LV systolic HF without DM Cardiac Cardiomyocyte cell death Truth Table dysfunction in doxorubicin-induced cardiomyopathy. NHE1 Hypertrophy þ Furthermore, Levkau et al. (27) showed that cardiac- Inflammation GSE26887 dataset specific deletion of BIRC5 pathway resulted in pre- *The stimulus, response, and restrictions included in each model are shown. mature cardiac death caused by a dramatic reduction DM ¼ diabetes mellitus; GSE designations ¼ Gene Expression Omnibus accession and subseries; HF ¼ heart failure; NHE1 ¼ sodium-hydrogen exchanger-1. in total cardiomyocyte numbers. Again, using adenoviral vectors to restore this pathway (through restoration of survivin) expression in cardiomyocytes the cardiac effects of empagliflozin are independent inhibited doxorubicin-induced apoptosis, induced of the presence of DM. DNA synthesis, and promoted cell cycle progression Uthman et al. (22) reported that SGLT2i reduced (27). Following this rationale and, to further confirm cardiac cytosolic Na þ and cytosolic Ca 2þ concentra- the present study’s in silico data, mRNA levels of both tions by inhibiting NHE1 in mouse cardiomyocytes. XIAP and BIRC5 were measured in an in vivo MI-HF Cardiac NHE1 activity and cytosolic Na þ were rat model. In agreement with existing studies, measured in the presence of clinically relevant con- both expression levels of XIAP and BIRC5 were centrations of empagliflozin (1 m mol/l), dapagliflozin substantially lower in vehicle-treated animals with (1 m mol/l), and canagliflozin (3 m mol/l). All 3 SGLT2is dysfunctional myocardium. Remarkably, in the bound with high affinity to the extracellular Naþ- empagliflozin-treated groups, expression levels of binding site of NHE1 (23). Recently, Baartscheer both XIAP and BIRC5 returned to normal, thus they et al. (23) proposed a possible relationship between were able to exert their antiapoptotic effects. empagliflozin and NHE1 inhibition in rats and rabbits. The present study was designed to investigate In their study, an increase in extracellular glucose whether the MoA of empagliflozin in HF in patients produced an increase in cytosolic levels of Na þ and with DM was different from that in patients without Ca 2þ, an effect that was inhibited by empagliflozin. DM. Surprisingly, the most robust MoA identified The effects of empagliflozin were strongly reduced here was identical in both groups, suggesting that after cells were pre-treated with the NHE1 inhibitor empagliflozin may have value in treating HFrEF pa- cariporide. In addition, empagliflozin affected cyto- tients regardless of their glycemic status, acting to solic Na þ and NHE1 flux in the absence of extracel- ameliorate adverse cardiac remodeling. In other lular glucose, which demonstrated its putative effects words, empagliflozin, currently indicated only for the in normoglycemic conditions. treatment of DM, may eventually be repurposed to These mechanistic studies have the power to treat patients with HFrEF. Several ongoing random- address specific mechanisms in detail, but the ized clinical trials are testing the value of SGLTi in HF unravelling of the global molecular effects of empa- patients (both with reduced and preserved ejection gliflozin in HFrEF with or without DM requires the fraction) with and without DM. integration of complex big data, best handled by The mathematical models reported here were artificial intelligence and deep learning analyses. robust, and each link described is validated in the Considering all possible empagliflozin-derived car- existing research publications. However, there is an diac mechanisms, the present analysis pinpointed ongoing discussion about the mathematical algo- NHE1 modulation as the most likely effector of the rithms used in artificial intelligence, often termed beneficial effects. Moreover, amelioration of car- “black boxes,” because sometimes it may not be easy diomyocyte cell death was found to be the most to follow the determination of the output, no matter important effect of empagliflozin in HF. how good or reliable this may be. This study’s anal- Most importantly, the algorithm predicted a mod- ysis took a massive amount of collected information ulation of AKTs resulting in an inability of BIRC2 to and analyzed it for hidden patterns in the data that degrade the antiapoptotic factors XIAP and BIRC5. would be otherwise inaccessible to human stand- XIAP expression was found to be decreased in failing alone analysis, to understand how the drug could human hearts, suggesting that this potent inhibitor of lead to a specific clinical outcome. These algorithms apoptosis may play a key role in protecting the heart are widely used in other fields, and the techniques are from cell damage (24,25). Using an adenoviral- among the best understood and developed, with associated vector, Piacentino et al. (24) increased time-tested characteristics which ensure an accurate JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Iborra-Egea et al. 839 NOVEMBER 2019:831–40 Empagliflozin in Heart Failure F I G U R E 3 mRNA Levels of Antiapoptotic Proteins XIAP and BIRC5 In Vivo (A) Upper panels show nondiabetic rats. (B) Lower panels show diabetic rats. Data are mean  SEM. *p < 0.05 vs. empagliflozin-MI and Vehicle-Sham. One-way ANOVA followed by Tukey post hoc test was used for normally distributed data, whereas Kruskal-Wallis followed by Mann-Whitney U test was used for non-normally distributed data. Abbreviations as in Figure 2. application. A key differentiating feature of deep extrapolated to HF with preserved ejection fraction learning compared with other subtypes of artificial (HFpEF). Further research is required to better un- intelligence is the ability to operate without external derstand the MoA of empagliflozin in HFpEF. guidance to draw conclusions. The neural network is not designed by humans, rather, the number of layers CONCLUSIONS is determined by the data itself. Deep ANNs have used primarily supervised learning, with training from Deep learning in silico analyses, together with in vivo known patterns and labeled input data (truth table) to validation, allowed deciphering the MoA of empagli- prospect reliable insights in large datasets. At the flozin in HFrEF. Our data suggest that empagliflozin in- same time, all the outputs from the algorithm are teracts and blocks the NHE1 co-transporter at the restricted by the current scientific and medical stan- cardiomyocyte level, triggering a signaling cascade that dards. Taken together, all this limits human bias, and halts detrimental cell death. These results were inde- the algorithm can act as a hypothesis-free, indepen- pendent of DM comorbidity, suggesting that empagli- dent data-driven analysis (28). flozin may emerge as a new treatment in HFrEF. STUDY LIMITATIONS. However, this study has some ACKNOWLEDGMENTS The authors thank Boehringer limitations. The same premise that makes the models Ingelheim for providing animal control subjects and robust also serves as a potential downside. Specif- chow containing empagliflozin. ically, the information considered here has already been described or uploaded in public repositories, ADDRESS FOR CORRESPONDENCE: Dr. Antoni which in turn limits the capacity of the analyses to Bayes-Genis, Heart Institute, Hospital Universitari incorporate data that have not been yet collected. The Germans Trias i Pujol, Carretera de Canyet s/n, 08916 data reported here are valid for HFrEF and cannot be Barcelona, Spain. E-mail: abayesgenis@gmail.com. 840 Iborra-Egea et al. JACC: BASIC TO TRANSLATIONAL SCIENCE VOL. 4, NO. 7, 2019 Empagliflozin in Heart Failure NOVEMBER 2019:831–40 PERSPECTIVES COMPETENCY IN MEDICAL KNOWLEDGE: The MoA TRANSLATIONAL OUTLOOK: Empagliflozin, in of empagliflozin in HFrEF was deciphered by using deep addition to its renal effects, has direct cardioprotective learning in silico analyses together with in vivo validation. effects at the cardiomyocyte level. These effects were The most robust MoA involved the sodium-hydrogen independent of DM comorbidity, suggesting that exchanger-1 (NHE1), which was similar for diabetics and empagliflozin may emerge as a new treatment in HFrEF. nondiabetics. Notably, direct NHE1 blockade by empagliflozin ameliorated cardiomyocyte cell death. REFERENCES 1. Emerging Risk Factors Collaboration, et al. As- 12. Irving H. Peirce’s truth-functional analysis and heart failure after myocardial infarction in rats. sociation of cardiometabolic multimorbidity with the origin of the truth table. History and Philoso- Am J Physiol Heart Circ Physiol 2001;280: mortality. JAMA 2015;314:52–60. phy of Logic 2012;33:87–97. H738–74. 2. Gerstein HC, Bosch J, Dagenais GR, et al., for 13. Iborra-Egea O, Gálvez-Montón C, Roura S, 22. Uthman L, Baartscheer A, Bleijlevens B, the ORIGIN Trial Investigators. Basal insulin and et al. Mechanisms of action of sacubitril/valsartan et al. Class effects of SGLT2 inhibitors in cardiovascular and other outcomes in dysglyce- on cardiac remodeling: a systems biology mouse cardiomyocytes and hearts: inhibition mia. N Engl J Med 2012;367:319–28. approach. NPJ Syst Biol Appl 2017;3:12. of Naþ/Hþ exchanger, lowering of cytosolic Naþ and vasodilation. Diabetologia 2018;61: 3. Zinman B, Wanner C, Lachin JM, et al., for the 14. Romeo-Guitart D, Forés J, Herrando- 722–6. EMPA-REG OUTCOME Investigators. empagli- Grabulosa M, et al. Neuroprotective drug for nerve flozin, cardiovascular outcomes, and mortality in trauma revealed using artificial intelligence. Sci 23. Baartscheer A, Schumacher CA, Wüst RC, et al. type 2 diabetes. N Engl J Med 2015;373:2117–28. Rep 2018;8:1879. empagliflozin decreases myocardial cytoplasmic Naþ through inhibition of the cardiac Naþ/Hþ 4. Wiviott SD, Raz I, Bonaca MP, et al., for the 15. Enderton H. A Mathematical Introduction to exchanger in rats and rabbits. Diabetologia 2017; Sabatine MS; DECLARE-TIMI 58 Investigators. Logic. 2nd edition. New York: Harcourt Academic, 60:568–73. Dapagliflozin and cardiovascular outcomes in type 2001. 2 diabetes. N Engl J Med 2019;380:1881–2. 24. Piacentino V III, Milano CA, Bolanos M, et al. 16. Yurista SR, Silljé HHW, Oberdorf-Maass SU, X-linked inhibitor of apoptosis protein-mediated 5. Cavender MA, Norhammar A, Birkeland KI, et al. Sodium-glucose co-transporter 2 inhibition attenuation of apoptosis, using a novel cardiac- et al., for the CVD-REAL Investigators and Study with empagliflozin improves cardiac function in enhanced adeno-associated viral vector. Hum Group. SGLT-2 inhibitors and cardiovascular risk: nondiabetic rats with left ventricular dysfunction Gene Ther 2012;23:635–46. an analysis of CVD-REAL. J Am Coll Cardiol 2018; after myocardial infarction. Eur J Heart Fail 2019; 71:2497–506. 21:862–73. 25. Haider N, Arbustini E, Gupta S, et al. Con- current upregulation of endogenous proapopto- 6. Zelniker TA, Braunwald E. Cardiac and renal 17. Kilic A, Huang CX, Rajapurohitam V, tic and antiapoptotic factors in failing human effects of sodium-glucose co-transporter 2 in- Madwed JB, Karmazyn M. Early and transient hearts. Nat Clin Pract Cardiovasc Med 2009;6: hibitors in diabetes: JACC State-of-the-Art Re- sodium-hydrogen exchanger isoform 1 inhibition 250–61. view. J Am Coll Cardiol 2018;72:1845–55. attenuates subsequent cardiac hypertrophy and heart failure following coronary artery ligation. 26. Lee PJ, Rudenko D, Kuliszewski MA, et al. 7. Flores E, Santos-Gallego CG, Diaz-Mejia N, J Pharmacol Exp Ther 2014;351:492–9. Survivin gene therapy attenuates left ventricular Badimon JJ. Do the SGLT-2 inhibitors offer more systolic dysfunction in doxorubicin cardiomyopa- than hypoglycemic activity? Cardiovasc Drugs Ther 18. Aker S, Snabaitis AK, Konietzka I, et al. Inhi- thy by reducing apoptosis and fibrosis. Cardiovasc 2018;32:213–22. bition of the Naþ/Hþ exchanger attenuates the Res 2014;101:423–33. 8. Packer M, Anker SD, Butler J, Filippatos G, deterioration of ventricular function during pacing-induced heart failure in rabbits. Cardiovasc 27. Levkau B, Schäfers M, Wohlschlaeger J, et al. Zannad F. Effects of sodium-glucose cotransporter Res 2004;63:273–82. Survivin determines cardiac function by control- 2 inhibitors for the treatment of patients with ling total cardiomyocyte number. Circulation heart failure proposal of a novel mechanism of 19. Baartscheer A, Schumacher CA, van 2008;117:1583–93. action. JAMA Cardiol 2017;2:1025–9. Borren MM, Belterman CN, Coronel R, Fiolet JW. Increased Naþ/Hþ-exchange activity is the cause 28. Topol EJ. High-performance medicine: the 9. Waltemath D, Wolkenhauer O. How modeling of increased [Naþ]; and underlies disturbed cal- convergence of human and artificial intelligence. standards, software, and initiatives support cium handling in the rabbit pressure and volume Nat Med 2019;25:44–56. reproducibility in systems biology and systems medicine. IEEE Trans Biomed Eng 2016;63: overload heart failure model. Cardiovasc Res 1999–2006. 2003;57:1015–24. 20. Engelhardt S, Hein L, Keller U, Klämbt K, KEY WORDS empagliflozin, heart failure, 10. Greco S, Fasanaro P, Castelvecchio S, et al. Lohse MJ. Inhibition of Naþ/Hþ exchange pre- machine learning MicroRNA dysregulation in diabetic ischemic heart failure patients. Diabetes 2012;61:1633–41. vents hypertrophy, fibrosis, and heart failure in beta 1-adrenergic receptor transgenic mice. Circ 11. Liu Y, Morley M, Brandimarto J, et al., for the Res 2002;90:814–9. MAGNet consortium. RNA-Seq identifies novel A PPE NDI X For supplemental tables and a myocardial gene expression signatures of heart 21. Kusumoto K, Haist JV, Karmazyn M. Naþ/Hþ video, please see the online version of failure. Genomics 2015;105:83–9. exchange inhibition reduces hypertrophy and this paper.