NRF2-related transcriptomic alterations as a long-term effect of doxorubicin treatment for childhood acute lymphoblastic leukemia


 The use of doxorubicin is associated with an increased risk of acute and long-term cardiomyopathy. Despite that the number of cancer survivors is growing constantly, little is known about the transcriptional mechanisms which progress in time leading to severe cardiac outcome. It is also unclear whether long-term transcriptomic alterations are related to an acute response to doxorubicin. We have sequenced miRNA from total plasma and extracellular vesicles (EVs) from 66 acute lymphoblastic leukemia (ALL) survivors treated with doxorubicin and 61 healthy controls (254 samples in total). We identified 94 and 33 TFs regulating differentially expressed miRNA in plasma and EVs compartments, respectively. For total plasma we found: HEY1, NRF2, HIF1A, NOTCH1, and for EVs: TGFB, ZEB1, ASCL2, PELP1, SIP1, TWIST1. Analysis of the data from patients with dilated and idiopathic cardiomyopathy revealed similarities with our data from EVs, especially in the activity of TFs related to epithelial-to-mesenchymal transition (EMT). To verify if similarities exist between acute and long-term response to doxorubicin, we performed experiments on cultured as well as we studied the role of NRF2, previously considered as an important player in acute response, in transcriptomic network upon doxorubicin treatment). KEGG analysis of miRNA targets for EVs indicates development of cardiomyopathy, whereas among GO process we found terms related to cardiac septum. In vitro experiments revealed that NRF2 is co-regulated with NOTCH effectors from HEY family, with known role in muscle regeneration. NRF2 and doxorubicin treatment contribute also to the dysregulation of TWIST family, important for the process of EMT.


Introduction
Anthracyclines, including doxorubicin, have contributed to improved survival in childhood acute lymphoblastic leukemia (ALL) from less than 10-90% and are still the most widely used antineoplastic drugs worldwide 1,2 . However, because of the lack of speci city for cancer cells, anthracyclines can also damage healthy, non-cancer cells, causing severe complications including cardiotoxicity during chemotherapy, as well as many years after treatment cessation. Lipschultz et al. proved that more than 50% of doxorubicin-treated ALL survivors exhibit abnormalities of left ventricular afterload or heart muscle contractility 3 several years after treatment cessaation. Moreover, heart disease is the most common non-cancer related cause of death among cancer survivors. The central dogma of anthracyclines evoked cardiomyopathy, based on acute doxorubicin action, points to oxidative stress caused by excessive amounts of reactive oxygen species (ROS) produced due to severe functional disruption of mitochondria 4,5 . Doxorubicin action involves also massive DNA damage including 8oxoguanine formation, DNA intercalation, and topoisomerase 2 poisoning, with downstream doublestrand breaks (DSBs) formation 6-8 . As DNA lesions are repaired only partially 7 , these should have consequences at the transcriptomic level and may be considered as the cause of long-term treatment side effects manifested at distant time points. However, knowledge on long-term transcriptomic processes leading to health complications due to anthracycline use is very limited, despite the fact that the number of cancer survivors is constantly increasing 9 . Furthermore, studying the molecular aspects of doxorubicin action is mostly con ned to acute effects or eventually short-term studies in cultured cardiomyocytes or animals. Thus, it is still not clear whether acute transcriptomic alterations are related to long-term effects that are linked to cardiomyopathy development.
Here, we build our approach based on the assumption that doxorubicin-induced dysregulation of the transcriptional network in target cells is responsible for sustained long-term side effects of anthracyclines. As the analysis of TF expression or activity in cardiac tissues is not feasible in human subjects, we aim to analyze the product of TF activity -miRNAs, secreted into the extracellular environment, which provides a potential for investigating the processes occurring in cells.
The latest reports on the role of extracellular vesicles (EVs) in intercellular communication underscore the stability of miRNAs in EVs and its usefulness as indicators of diverse processes 10 . Additionally, the selective characteristic of miRNA sorting into EVs and the speci city of EVs targeting due to the presence of particular surface proteins interacting with the membrane of recipient cell [11][12][13] were reported to re ect processes ongoing in tissues. Therefore, we searched for regulators (transcription factors (TFs) and miRNA) of the gene expression network that could be related to the development of cardiac complications in acute lymphoblastic leukemia survivors using information on miRNA expression in two compartments, namely the total plasma and EVs. On the other hand, we aimed to check whether the common denominator for long-term and acute doxorubicin transcriptomic effects exists. Thus, we analyzed transcriptomic data from experiments on cultured cardiomyocytes also in cells with downregulated NRF2, which we found as dysregulated in ALL survivors.

Results
The summary of the rst part of the study regarding long-term effects of doxorubicin use is presented in Fig. 1.

Characteristics of the studied groups
There was no statistically signi cant difference in sex and there was a borderline nonsigni cant difference in age between the study groups (as presented in Supplementary Table 1). Subsequently, we used the results of complete blood count and lipid panel tests to compare the two study groups. We applied a feature selection algorithm based on cross-validation (see Methods section) to identify variables with the largest discriminative power and subsequently tested for signi cance. These results are summarized in Supplementary Table 2.

Extracellular vesicles characteristics
Western blot experiments showed signi cant enrichment for tetraspannins CD63, CD81, and CD9 (Fig. 2). Measurements on Nanosight instrument revealed that median of vesicles size was 69,75 (Q1 = 65,05 and Q3 = 73,35 ) and concentration 9,17E + 12 (Q1 = 5,58E + 12 and Q3 = 1,42E + 13) Differential expression of miRNAs in blood plasma and exosomes We sought to nd miRNAs that were differentially expressed between cases and controls in blood plasma and EVs separately. Due to an insu cient RNA amount, several samples were not included in sequencing. All further analyses were performed on 59 controls and 61 ALL patients.
We detected the following: (A1) 201 miRNAs that were differentially expressed between ALL cases versus controls in blood plasma; and (A2) 49 miRNAs that were differentially expressed in exosomes (top 10 results are presented in Table 1, the full list of differentially expressed miRNAs is available as   Supplementary Table 3 (A1) and 4 (A2). related to cardiomyopathy or strongly associated with it. Among the enriched pathways, we found tyrosine kinases, MAPK, ErbB, and neurotrophin signaling, which have been previously linked to heart function or anthracycline action [22][23][24][25][26][27] . However, only in the case of A2 analysis the term "dilated cardiomyopathy" appeared. These ndings ensured us that transcriptomic changes, which in a long perspective lead to cardiomyopathy development, can be seen as early as at the time point when no severe symptoms are clinically manifested. This also suggested that miRNA in EVs may have particularly signi cant meaning.
GO term enrichment analysis of targets of differentially expressed miRNAs Next, for the two sets of differentially expressed miRNAs de ned above, we performed GO term enrichment analysis. Brie y, we rst mapped the miRNAs to targets, subsequently annotated targets to GO terms, and performed enrichment tests. For the miRNAs in the (A1) set, we detected over 600 enriched terms, whereas for (A2), we detected over 500 enriched terms (with FDR < 0.05). The most enriched terms in (A1) included processes associated with neural and epithelial cell development, mesenchymal cell differentiation, and histone modi cations. The terms most enriched in (A2) included renal, kidney, and respiratory system development as well as regulation of mRNA splicing and cardiac septum development. (Supplementary Table 5).
Association of circulating extravesicular miRNAs with interventricular septum thickness Detailed echocardiographic examination was performed for a subset of ALL survivors. We decided to gain further insight into the association between the circulating miRNAs and the thickness of the interventricular septum (IVS), which was present among GO terms related to the A2 subset. For this purpose, we performed an association analysis of miRNAs (in EVs) versus the measured thickness of the IVS. The results are summarized in Table 2. Mir-323, mir-134, and mir-199a were previously reported as differentially expressed in patients with chronic heart failure at early stages characterized by reduced catecholamine sensitivity 28 . Association of miRNA-speci c regulatory variants with CHF.
Since we were investigating the long-term effects of chemotherapy -in particular, its relation to the cardiotoxicity of anthracyclines -therefore, we asked whether there exist common regulatory variants which are associated with the expression of A1 and/or A2 miRNAs and, at the same time, were found to be linked to HD. This analysis allows us to further prioritize the molecular pathways which may lead (in a long time horizon) to heart failure. To this aim, we used two external datasets: (1) the recent CHF GWAS results 29 and (2) miRNA eQTLs identi ed by the Huan et al 30 . For the miRNAs found in A1 and A2 (separately), we searched for cell-type adjusted eQTLs and subsequently queried the CHF GWAS results to nd the relevant variants. For the A1 set, we were only able to nd variants with signals for two miRNAs: mir-100 and mir-339, whereas for the A2 set we only identi ed relevant variants in mir-31.  Tables 6 and 7). In A1, we found p53 and other TFs related to its pathway, which were previously linked to DNA damage and oxidative stress also upon doxorubicin treatment [32][33][34] , as well as NOTCH1. The second most signi cant TF is HEY1 involved in cardiac septum development 35 . We also found NRF2 and other TFs that are associated with short-term response to environmental stress (for instance, xenobiotics), e.g., HIF1A

Enrichment analysis of TFs
expression is directly regulated by NRF2 36 . In the A2 analysis, we found TFs related to epithelial-to-Page 9/45 mesenchymal transition (EMT) regulation: TGFB1, ZEB2, ASCL2, PELP1, SIP1, TWIST1, SLUG (SNAI2), JAG1.  Co-expression between differentially expressed miRNAs and selected TFs To further investigate the association of differentially expressed miRNAs and TFs, we expanded our analysis and tested whether, for selected TFs, we observed a strong mutual relationship between the extracellular content of the differentially expressed miRNAs and the expression of a given TF in leukocytes derived from peripheral blood. For this purpose, we used the data available in the Framingham Heart Study cohort for individuals in the Offspring Cohort. For most of the selected TFs, we found that differentially expressed miRNAs were strongly co-expressed with them. For this analysis, we considered only the top 0.25% of all miRNAs as being strongly co-expressed with a given TF. We used this approach instead of the standard notion of statistical signi cance, as we wished to present this result only as a proof-of-principle, thus avoiding the extensive multiple testing correction. For each TF of interest, we present in Supplementary Table 8a list of co-expressed miRNAs that are in A1 and A2. Additionally, for each set of differentially expressed miRNAs (A1 and A2), we rank the miRNAs by the number of TFs coexpressed with them (Table 5). Interestingly, two miRNAs correlated with the highest number of TFs are involved in NRF2 pathway through targeting NRF2 inhibitors (miR-193a -targets NRF2 antagonist BACH2 37,38 , miR-141 -NRF2 repressor KEAP1 39 . Additionally, our analysis of FHS data also shows the correlation of NRF2 with the high number of miRNAs differentially expressed in ALL survivors. Comparison with TFs enriched in regulation of miRNAs associated with cardiomyopathy. To assess whether our data on circulating miRNAs re ect processes active in people with dilated cardiomyopathy (DCM) or idiopathic cardiomyopathy (ICM), which are manifested by changes in their circulating and/or target tissue miRNAs, we analyzed the data of Akat et al 40 . First, we compared the differentially expressed miRNAs in A1 and A2 sets with miRNAs differentially expressed between the serum of healthy individuals (HC) and patients with DCM/ICM (prior to any surgical intervention). We obtained the following results: 14 miRNAs in A1 that were also differentially expressed between HC and ICM in blood plasma; 13 miRNAs in A1 that were also differentially expressed between HC and DCM; 14 miRNAs in A2 that were also differentially expressed between HC and ICM; and 8 miRNAs in A2 that were also differentially expressed between HC and DCM. Of these, 9 miRNAs were common between ICM and DCM for A1 subset and 4 were common between ICM and DCM for A2 subset (  Fig. 6. Subsequently, we analyzed the tissue mRNA expression data from the same study. Again, we considered only samples from HC, DCM, and ICM (excluding the fetal hearts from the analysis). We identi ed two sets of differentially expressed probes (under FDR < 0.05) -between HC and DCM as well as between HC and ICM. Subsequently, we used the "RcisTarget" package to evaluate the enriched motifs of TF binding in 50-kb window around the differentially expressed genes. Among the signi cant TFs, we found that many were present in our data from ALL survivors, for instance, NFE2, STAT1, STAT5A, NFKB, EP300, GATA1, GATA2, and GATA3 (data not presented).   Transcription factors co-expression in cardiomyocytes With the results given above, we hypothesized that NRF2, important for short-term response to doxorubicin, might play also a signi cant role in the development of cardiomyopathy. Therefore, we decided to study NRF2 function in terms of its co-expression network. The schematic view of this part of study is depicted in Fig. 7.
To verify the functional effects of NRF2 on the transcriptional response to doxorubicin in human cardiomyocytes, we performed silencing experiments on immortalized cardiomyocytes (IHC).
Subsequently, we identi ed four sets of differentially expressed mRNAs: B1 -between shCTRL and shNRF2 cells with no dox treatment, B2 -between shCTRL and shNRF2 cells with dox treatment, B3between no-dox and dox treatment in shCTRL cells, B4 -between no dox and dox treatment in shNRF2 cells. The results of these comparisons are summarized in Supplementary Tables 9-12. Due to the limited sample size, we did not focus on the direct interpretation of these results, instead, we performed a transcription factor enrichment analysis based on the ChEA3 tool. In this way, we aimed to gain further insight in the mutual relations of the TFs with NRF2 in the context of dox treatment. For each of the sets B1-B4, we used the top 50 differentially expressed mRNAs and obtained scores for TFs which are likely to be regulating these mRNAs. We decided to use only the top 50 mRNAs as we wanted to avoid the bias associated with the different numbers of differentially expressed genes between the considered conditions. The complete results of this analysis are summarized in Supplementary Tables 13-16. Subsequently, we detected the TFs with the largest fold changes between the study conditions. In other words, we identi ed four sets of TFs: T1 -with the highest (positive) fold change between scores for B1 and B2; T2 -with the lowest (negative) fold change between scores for B1 and B2; T3 -with the highest (positive) fold change between scores for B3 and B4; T2 -with the lowest (negative) fold change between scores for B3 and B4. Therefore, the set T1 represents TFs likely to be regulating B2 and not B1 mRNAs, T2 -TFs likely to be regulating B1 and not B2 mRNAs, T3 -TFs likely to be regulating B4 and not B3 mRNAs, T4 -TFs likely to be regulating B3 and not B4 mRNAs. Results are summarized in Table 9 and in Supplementary Tables 17-20.

Co-expression of NRF2 and other TFs in the GTEx Heart Left Ventricle RNA-sequencing data
To further re ne these results, for each set T1-T4, we took the top 50 TFs (Supplementary Tables 17-20) and tested the co-expression between the mRNAs of these TFs (and the NRF2 mRNA) in the GTEx Heart Left Ventricle RNA-sequencing data (for the co-expression analysis TFs with TPM > 0.5 were chosen). The results are presented in Figs. 8-11. Interestingly, in this way we were able to identify a cluster of: 13 TFs for T1, 1 TFs for T2, 29 TFs for T3 and 15 TFs for T4 which are co-regulated with NRF2 (with a correlation coe cient above 0.5). The list of TFs for each subset is presented in Table 10. Hence, both doxorubicin treatment and NRF2 de ciency cause alterations in HEY1 functioning. Doxorubicin treatment of NRF2 de cient cells causes changes in the functioning of TWIST1, TWIST2, SNAI2, which could be related to the difference in GRHL1 action (T2 comparison -sh control vs sh NRF2 cells upon treatment).

Discussion
Here we present diverse lines of evidence that long-term molecular effects of doxorubicin action in ALL survivors include persistently altered transcription factor activity, which results in changes of miRNA presence in the circulation that may contribute to cardiomyopathy development, the major life-threatening long-term effects of anthracycline treatment 41,42 . We demonstrate that especially miRNA circulating in EVs may be a useful source of information on transcriptomic processes that are changed due to anticancer treatment.
Indeed, our KEGG and GO enrichment analyses for differentially expressed vesicular miRNA targets indicate that processes including dilated cardiomyopathy and related processes (e.g. ERBB signalling 43,44 ) are ongoing in ALL survivors. Hence, we decided to search for TFs that regulate the expression of differentially expressed miRNAs.
It is noteworthy that TGFB expression can be increased in the heart tissue many weeks after doxorubicin treatment 58 .
Moreover, TGFB1-related factors including TWIST1 and SLUG (SNAI2), were previously reported as related to cardiac morphogenesis including atrioventricular cushion formation or cardiac septation 59,60 .
Recent data revealed that TWIST2 is expressed in speci c populations of interstitial heart cells, which contributes to cardiac homeostasis and regeneration 61 . Additionally, TWIST1 as well as TWIST1regulated mir-199a (miRNA correlated with IVS thickness in our study) were reported as down-regulated in the heart tissue of patients with severe cardiomyopathy 62 . Dysfunction within this system may result in poor cell renewal and proliferation, as was shown for stem cells 63 .
Additionally, our analysis of miRNA expression data from individuals with dilated and idiopathic cardiomyopathy revealed that except for some similarities in miRNA differential expression, vesicular miRNAs in the former ALL patients share Tfs regulators with miRNA related with cardiomyopathy (TGB, ZEB2, SNAI1, TWIST1).
Moreover, one of the miRNA common for these analyses, mir-31, signi cantly upregulated in extracellular vesicles of ALL survivors, and which we found to be involved in the genetics of heart failure is also the top miRNA correlated with the highest number of Tfs regulating miRNA in A2 analysis, including factors well known for signi cant role in both response to doxorubicin and cardiomyocyte functioning (NRF2, TP53, HIF1A) as well as EMT process (ZEB1, SNAI1). Mir-31 has been previously reported to be upregulated in cardiomyocytes due to hypoxia or oxidative stress and in hearts after myocardial infarction. Its silencing upregulated troponin T in cardiomyocytes, while administration of mir-31 inhibitor in rats post-myocardial infarction improved heart function 64 . Moreover, a recent paper has shown that doxorubicin treatment induced mir-31 expression in cultured cardiomyocytes as well as in heart tissue, causing downregulation of RNA binding protein quaking playing a signi cant role in doxorubicin-induced cardiotoxicity 65 , what imply that pathways related to mir-31 might also be valuable targets for future studies.
Surprisingly, the results of our analyses support the role of NRF2 as one of the key regulators of the transcriptomic network in ALL survivors treated with doxorubicin. NRF2 is associated with doxorubicin resistance and short-term effects of chemotherapy 66,67 . Down-regulation of NRF2, which physiologically plays the main role in the protection against oxidative stress through the activation of transcription of antioxidative response genes, causes excessive doxorubicin-induced cardiotoxicity 68,69 . Additionally, prolonged oxidative stress and unbalanced NRF2 activity may induce senescence and premature age-related diseases, which have been attributed to the long-term effects of anticancer treatment 70,71 . To investigate the detailed role of NRF2 in the transcriptional network upon doxorubicin treatment, we performed NRF2 silencing experiments on cultured cardiomyocytes. It reveals that even without drug exposure, down-regulation of NRF2 is accompanied by changes in the functioning of NOTCH1 (present also among TFs regulating plasma miRNAs) and TGFB-related TFs (present among TFs regulating miRNAs in EVs') involved in EMT: TGFB, TWIST1, TWIST2, ZEB1, SNAI2, which have been previously reported as related to cardiac morphogenesis including atrioventricular cushions formation or cardiac septation 59,60 . We also observed signi cant alterations of NOTCH signaling targets: HEY2, HEYL, and HES factors are required for e cient chromatin binding by HEY1 72 , which we show is further dysregulated in shNRF2 cells upon doxorubicin treatment. HEY1, HEY2, and HEYL together with HES family of TFs are crucial for the formation of valves and septa during embryogenesis 59 . HEY1, the second signi cant TF regulating plasma miRNA expression, turns up as differentially active upon NRF downregulation as well as during doxorubicin stimulation in control cells. It was previously shown that inactivation of HEY1 and HEYL causes congenital heart defects including ventricular septal defects 35 .
Whereas in aging muscle stem cells, HEY1 rescues cells from death due to mitotic catastrophe during muscle regeneration 73 . Moreover, in the case of NRF2 loss, HEY1 downregulation was accompanied by a delay in repair after chemical injury 63 . Second of Hey genes -HEY2 is speci cally expressed in the interventricular septum, ventricular compact myocardium, the atrioventricular canal out ow tract, at the base of trabeculae, and in epicardial cells 74 . It was also postulated that HEY2 is necessary to maintain ventricular identity by mature cells 74 . Indeed, cardiomyocyte-speci c deletion of HEY2 results in the transcription of atrial genes in the ventricular myocardium accompanied by impairment of cardiac contractility and changes in the morphology of the right ventricle 75,76 , and its de ciency causes septal defects and cardiomyopathy in mice 75 .
Additionally, TWIST1 as well as TWIST1-regulated mir-199a (miRNA correlated with IVS thickness in our study) were reported as down-regulated in the heart tissue of patients with severe cardiomyopathy 62 .
Notably, GRHL1, a regulator of gene expression upon doxorubicin treatment in the case of diminished NRF2 activity, is a member of the grainyhead family of TFs, which acts as a suppressor of TGFβ-induced, and TWIST-induced EMT 77 . The nding of co-regulation between NRF2 activity and NOTCH signaling in ALL survivors remains in line with previous reports suggesting reciprocal regulation between these these Tfs 78 . Dysfunction within this system may result in poor cell renewal and proliferation, as was shown for stem cells 63 .
Here we also show that doxorubicin treatment of cardiomyocytes alters the activity of RBPJ, an important transcriptional regulator of NOTCH 79 . Additionally, since TWIST, HEY, and HES proteins all belong to the basic helix-loop-helix (bHLH) transcription factors family 80,81 , it seems reasonable to hypothesize that the co-regulation of these TFs with NRF2 may be related to the 60 amino acid region which contains two highly conserved domains, considered as involved in interactions with TFs 82 .
In summary, our approach allowed to identify dysregulated mechanisms, including altered NRF2 activity, that might be involved in late cardiac complications due to doxorubicin treatment.
However, to the best of our knowledge, it is still unclear how the damage caused by anthracyclines, often during childhood, can be propagated until adulthood. Doxorubicin-induced DSBs are especially frequent around active promoters 83 and are known to cause transcriptional repression 84 . Gene expression silencing was shown to be converted to long-term stable silencing through modi cations like CpGs or histone methylation, which may be heritable 85,86 . Thus, NRF2 gene with high transcriptional activity upon exposure to oxidative stress, could be at particularly high risk of modi cation during doxorubicin treatment. Although further investigations are required to prove such possibility.
Our approach using differential expression of circulating miRNA allowed to identify dysregulated mechanisms, which may be involved in the development of long-term consequences of doxorubicin treatment in ALL survivors. Additionally, microRNA encapsulated in extracellular vesicles signi cantly indicated EMT-related processes which seem to be related to cardiomyopathy development, which although needs further study.
In summary, we present data proving that NRF2-related molecular changes initiated by doxorubicin treatment may cause alterations in the functioning of transcription factors related to cardiac cells identity and regeneration. Among them, especially HEY as well as TWIST Tfs could be considered as potential targets in future studies.
Our approach using differential expression of circulating miRNA allowed to identify dysregulated mechanisms, which may be involved in the development of long-term consequences of doxorubicin treatment in ALL survivors. Additionally, microRNA encapsulated in extracellular vesicles signi cantly indicated EMT-related processes which seem to be related to cardiomyopathy development, which although needs further study.  20 , using a two-pass method. The uniquely aligned reads mapped to genes were enumerated using HT-Seq-count v.0.11.2 21 . After quality assessment of alignment and enumerated gene expression levels, the raw expression values were uploaded into R for further analysis.

Statistical methods
All statistical analyses and ltering steps were performed in R (v3.5.2). Out of 1986, the detected miRNAs, 1452 (with at least ve reads observed per sample) were considered as positive hits and were used in the statistical analysis. The differential expression was analyzed by the edgeR package with two different experimental designs: model #1. Xp ~ compartment + status:compartment and model #2. xp ~ status + compartment:status.
Additionally, the results of differential expression analysis were further ltered, and only results with FDR < 0.05 and logCPM > 1 were considered signi cant. Elastic net regression was performed as implemented in the package "glmnet" with default settings for cross-validation-based feature selection. GO enrichment was performed using the "topGO' package, and the KEGG enrichment analysis was performed using the RbiomirGS package. eQTL enrichment analysis was performed using Fisher's exact test. For correlation tests, Pearson's product-moment correlation coe cient was used. TransmiR v2.0 software was accessed through www.cuilab.cn.
The analysis of mRNA expression was performed by means of the edgeR package in R. In short, raw reads were normalized for library size and subsequently dispersion was estimated (common, tagwise, and trended). For target differential expression analysis, genes with at least 10 reads in at least one sample were used. Differentially expressed genes were detected by means of the Likelihood Ratio Test in a linear model with suitably chosen contrasts.
The ChEA3 tool was used as available through. For the enrichment analysis, the top 50 differentially expressed genes were used.

Analysis of the correlation between miRNA expression and echocardiography
To test in an unbiased fashion whether the expression in plasma/exosomes of miRNA correlates with selected echocardiographic parameters, we used the following approach. Using pseudocounts, we rst selected features (miRNAs) which are differentially variable between cases and controls (using Levene's test). For further analyses, we used only the ones which remain signi cantly differentially variable with FDR < = 0.05. Subsequently, we reduced the dimension of the data to three using the unsupervised UMAP method. We did this separately in plasma and exosomes (differentially variable miRNAs were selected regardless of compartment). In what follows, we used the LASSO model to test which echocardiographic parameters correlate with the three coordinates of the embedding (here the set of covariates is de ned by the selected parameters of the echocardiographic test and the dependent variable is three-dimensional and represented by the coordinates of the UMAP embedding)

Declarations
Con ict of interest: The authors report no con ict of interest