Skeletal muscle reprogramming by breast cancer regardless of treatment history or tumor molecular subtype

Increased susceptibility to fatigue is a negative predictor of survival commonly experienced by women with breast cancer. Here, we sought to identify molecular changes induced in human skeletal muscle by BC regardless of treatment history or tumor molecular subtype using RNA-sequencing and proteomic analyses. Mitochondrial dysfunction was apparent across all molecular subtypes, with the greatest degree of transcriptomic changes occurring in women with HER2/neu-overexpressing tumors, though muscle from patients of all subtypes exhibited similar pathway-level dysregulation. Interestingly, we found no relationship between anti-cancer treatments and muscle gene expression, suggesting that fatigue is a product of BC per se rather than clinical history. In vitro and in vivo experimentation confirmed the ability of BC cells to alter mitochondrial function and ATP content in muscle. These data suggest that interventions supporting muscle in the presence of BC-induced mitochondrial dysfunction may alleviate fatigue and improve the lives of women with BC.

molecular subtypes, with the greatest degree of transcriptomic changes occurring in women with HER2/neu-overexpressing tumors, though muscle from patients of all subtypes exhibited similar pathway-level dysregulation. Interestingly, we found no relationship between anti-cancer treatments and muscle gene expression, suggesting that fatigue is a product of BC per se rather than clinical history. In vitro and in vivo experimentation confirmed the ability of BC cells to alter mitochondrial function and ATP content in muscle. These data suggest that interventions supporting muscle in the presence of BC-induced mitochondrial dysfunction may alleviate fatigue and improve the lives of women with BC.

INTRODUCTION
Muscle dysfunction in individuals with cancer is commonly thought to be a consequence of muscle atrophy, which is a major component of the paraneoplastic syndrome known as cancer cachexia 1,2 . While studies in men with cancer support the claim that muscle functional capacity is dependent on muscle size, women with cancer report a significant degree of muscle dysfunction despite typically remaining weight-stable [3][4][5] . Muscle dysfunction in breast cancer (BC) commonly presents as a persistent, severe fatigue that frequently contributes to dosereduction or treatment cessation and is an independent predictor of survival in a variety of cancer types, including BC [6][7][8][9][10][11] . Thus, it is probable that improving muscle fatigue will improve both quality of life and survival in BC.
At present, there are several purported contributory factors for cancer-related fatigue, including immunological responses to tumor growth; side-effects of cancer therapies; depression and/or emotional distress; anemia; hormonal, nutritional, and metabolic disturbances; and inadequate physical activity 8,12 . Pharmacological treatments for cancer-related fatigue exhibit limited and inconsistent success, in part because determining the mechanisms contributing to fatigue in a given patient can be quite challenging, particularly in patients with early-stage disease and those not receiving anti-cancer treatments 13 . Identifying mechanisms of BC-related fatigue that are intrinsic to skeletal muscle, generalizable across BC subtypes, and independent of treatment status could significantly aid in the development of appropriate therapies, which would be applicable to a large number of patients experiencing BC-associated fatigue.
Our laboratory has recently reported that skeletal muscle of women with BC exhibits a distinct gene expression signature that is not dependent on molecular subtype 14 . Furthermore, we have identified signaling via the metabolic regulators of the peroxisome-proliferator activated 5 receptor (PPAR) family as potential key mediators of fatigue in women with BC and female mice bearing BC patient-derived orthotopic xenografts (PDOXs) 15 . Our previous analyses did not include women with primary tumors that overexpressed HER2/neu in the absence of estrogen receptor (ER) and progesterone receptor (PR) expression. In the current study, we have expanded our analyses into all molecular subtypes, including both transcriptomic and proteomic analyses of muscle biopsies from patients with HER2/neu-overexpressing tumors, and significantly increased our sample size to create, to our knowledge, the largest study of transcriptomic and proteomic changes in muscle of women with BC. We tested the hypothesis that BC induces a common molecular response in skeletal muscle that is independent of the molecular subtype of the tumor and the patient's treatment history.

Patient characteristics. A total of 51 BC patients representing 4 breast tumor subtypes and 20
non-cancer controls provided pectoralis major muscle biopsies and/or detailed clinical information for use in the present study. There were no differences in mean body mass index (BMI) between non-cancer controls and BC patients; average BMI in the control group was categorized as Class I Obesity and in the BC patients was categorized as Overweight.
Additionally, there were no significant differences in the percent (%) change in BMI, body fat %, or lean body mass between controls and BC patients. There were no differences in BMI between any of the 4 breast tumor subtypes (Table 1).
Skeletal muscle gene expression profiles. Skeletal muscle biopsies from patients diagnosed with BC (n=33) and non-cancer controls (n=10) were used for RNA-seq, and BC patients were classified based on molecular subtype of their primary tumor, as follows: luminal (ERPR)positive for estrogen receptor (ER) and progesterone receptor (PR) without overexpression of HER2/neu; HER2 -overexpression of HER2/neu in the absence of ER and PR expression; triple negative (TN)-absence of ER, PR, and HER2/neu expression; and triple positive (TP)presence of ER and PR expression, and overexpression of HER2/neu (ERPR n=10, HER2 n=5, TN n=9, TP n=9). Unsupervised clustering analysis indicated a significant degree of clustering based on molecular subtype, particularly with regard to patients with tumors overexpressing HER2/neu in the absence of ER or PR ( Figure 1A). Multidimensional scaling (MDS) analysis in 3 dimensions revealed that the gene expression profiles of skeletal muscle from patients with ERPR, TP and TN tumors were similar, while the profile from skeletal muscles from patients with HER2/neu-overexpressing tumors was significantly different ( Figure 1B). Overall fit of the MDS model was dramatically improved by the use of BC subtype as a covariate ( Figure 1B, adjusted R 2 =0.44, p=0.0001) rather than a model including only binary disease status ( Figure   1C, adjusted R 2 =0.20, p=0.008), which was already a significant improvement over the null model ( Figure 1D).
To address the question of whether this obvious difference in overall muscular gene expression between groups was inherent to differences in the primary tumor type or the myriad of clinical characteristics that could potentially differ between groups, we assessed the relationship between clinical characteristics and skeletal muscle gene expression in the context of a multivariate linear regression model, using the 3 dimensions of the MDS dissimilarity matrix as response variables.
Among the various treatment types, body composition, serum albumin, and changes in body mass, the only assessed variable to yield statistical significance at α=0.05 when used as a single independent variable was patient group (i.e. Control, ERPR, HER2, TN, TP), with serum albumin nearing statistical significance (Table 2). Notably, chemotherapy, radiotherapy, and hormonal treatments did not correlate with overall gene expression patterns, nor did the patient's trend of weight change over time. Using forward selection, a final model including patient group and serum albumin was identified as the best-fitting model for predicting muscular gene expression from clinical data (Table 3).
Because serum albumin appeared to provide predictive value for skeletal muscle gene expression data, and because serum albumin is commonly used by oncologists to monitor patients' nutritional status, the relationship between serum albumin and changes in BMI over time were assessed in the group of patients that provided biopsies. There was no correlation observed between serum albumin at the date of biopsy collection and the individual's rate of weight change over time (Figure2A). Similar results were obtained in a retrospective chart review of 3,001 patients with BC. While there was a statistically significant correlation between a patient's first record of serum albumin and the patient's rate of weight change in this large sample ( Figure   2B), the effect size may well be clinically insignificant (R = 0.094). The average daily weight change was negligible in women with normal serum albumin as well as those with low serum albumin (< 3.4 g · dL -1 ), with both groups having means within one standard deviation of 0 ( Figure 2C). For a patient to be considered cachectic by traditional standards, an average daily weight loss of at least 0.027% would be required to lose 5% of their weight in 6 months 1,2 . In our large cohort, a logistic regression analysis was conducted using a threshold of 3.4 g · dL -1 serum albumin to predict whether a patient would exhibit this rate of weight change. In this analysis, omnibus model fit was significant at ɑ < 0.05, though effect size as determined by Nagelkerke's pseudo-R 2 was very small and indicates a very weak predictive value (R 2 = 0.08).
Additionally, a threshold of 3.4 g · dL -1 serum albumin was only 40% sensitive to identifying this level of weight change and only yielded a positive predictive value of 17.3%. In other words, 60% of BC patients exhibiting a rate of weight loss consistent with cachexia had normal serum albumin, while BC patients with serum albumin < 3.4 g · dL -1 only had a 17.3% chance of exhibiting a rate of weight change consistent with cachexia.

Differential gene expression analysis by subtype. Differentially expressed genes (DEGs)
within skeletal muscle were first identified by comparing BC patients by subtype to control.
Considerable overlap of DEGs was observed between the 4 breast tumor subtypes. Of the 3,468 genes identified as differentially expressed in at least one subtype, only 7 (0.2%) were unique to ERPR patients, 173 (5.0%) to TN patients, and 80 (2.3%) to TP patients. However, 2,410 genes (70%) were unique to HER2 patients, and only 8 (0.23%) were differentially expressed in all patient groups (Figures 3A and 3B). These observations were quantified and reveal that there was an approximately 2-fold fewer-than-expected number of unique DEGs in muscle of patients with ERPR, TP and TN tumors if the DEGs were independent of subtype. Specifically, one would expect 14, 186, and 508 unique DEGs in each subtype, respectively, whereas we actually observed 7, 80, and 173 DEGs in these groups. Further, HER2 patients' muscle exhibited 2-fold fewer-than-expected DEGs shared with any combination of two other subtypes (Observed = 5 + 2 + 84 = 91; Expected total = 203; overall χ 2 on 6 degrees of freedom = 1224, p = 3 x 10 -261 ).
This indicates that the ERPR, TP, and TN groups share a greater number of DEGs in skeletal muscle than one would expect if the DEGs were independent of subtype and in contrast, muscle from HER2 patients does not exhibit the same similarity to the other subtypes in terms of shared DEGs ( Figure 3C). Collectively, these data demonstrate that transcriptional responses in skeletal muscle of patients with ERPR, TP and TN tumors are highly similar, in support of previous data from our laboratory 14,15 . Furthermore, the transcriptional responses in muscles from patients with HER2/neu-overexpressing tumors partially overlap with the other subtypes, but exhibit a significant contrast to the other 3 subtypes, suggesting that this tumor type is associated with a unique transcriptional adaptation within skeletal muscle. DEG pathway analysis by subtype. Qiagen's Ingenuity Pathway Analysis (IPA) software was used to infer pathway-level dysregulation based on trends in transcriptomic changes. Identified pathways were shared between all breast tumor subtypes to a greater extent than individual genes. Of the 368 pathways identified by IPA as dysregulated in any subtype, 43 (11.7%) pathways were identified as significantly dysregulated in all 4 breast tumor subtypes relative to control. 23 (6.3%) pathways were uniquely dysregulated in muscle from patients with TP tumors, 25 (6.8%) in patients with ERPR tumors, 29 (7.9%) unique to TN patients, and 62 (16.8 %) uniquely dysregulated in patients with HER2/neu-overexpressing tumors ( Figures 4A, 4B,   4C). Chi-square analysis to test whether the number of pathways shared between subtypes differed between groups was non-significant (overall χ 2 on 6 degrees of freedom = 8.10, p = 0.23), indicating that the four BC subtypes share a common core of dysregulated pathways.
When the 43 overlapping pathways were ranked by how similarly all subtypes appeared in terms of magnitude and directionality of dysregulation, LXR/RXR signaling stood out in its strong, consistent inhibition across all BC subtypes ( Figure 4D), and the related TR/RXR pathway was also identified as consistently significantly dysregulated. These pathways are closely related to and often integrated with the PPAR/RXR signaling pathway, which our laboratory has previously identified as a likely upstream regulator of muscle fatigue in BC patients 14,15 .
Additional dysregulated pathways were related to inflammation and immune system processes. identified as being the two most significantly dysregulated canonical pathways by IPA ( Figure   5F).

Subtype-independent DEG analysis and experimental validation.
To identify potential mechanisms of muscle fatigue that are consistent across all molecular subtypes of BC, pathwaylevel enrichment analysis was conducted using a restricted dataset of genes that were significantly differentially expressed when comparing all BC patients to control (FDR < 0.10). In To test this prediction of altered mitochondrial function in tissues distant from the primary tumor in vivo, we generated 6 female BC-PDOX mice and isolated live mitochondria from their skeletal muscle at euthanasia for quantification of ATP content. We found a significant reduction in ATP content within both of the two major skeletal muscle mitochondrial subpopulations, the interfibrillar mitochondria (IFM) and subsarcolemmal mitochondria (SSM), relative to female control animals ( Figure 6C). In vitro assays were conducted to determine whether mitochondrial dysfunction in muscle is a direct response to BC-secreted factors or an indirect response mediated by other tissues. In support of a direct response, conditioned media from the luminal EO771 BC cell line and the HER2/neu overexpressing NF639 line significantly repressed aerobic ATP production in differentiated C2C12 myotubes, whereas media conditioned by either the normal mammary epithelial cell line EpH4-EV or C2C12 myoblasts did not ( Figure 6D). To validate the role of the PPAR signaling proteins in mediating the systemic response to BCsecreted factors, conditioned media was isolated from EO771 and NF639 cells and applied to HEK293 cells stably expressing a PPAR-responsive promoter driving GFP expression. Media conditioned by both BC cell lines significantly repressed GFP intensity relative to reporter cellconditioned media, while media conditioned by a normal mammary epithelial cell line did not alter GFP signal relative to control ( Figure 6E). These data indicate that BC cells secrete a substance that is capable of directly influencing metabolic function in skeletal muscle, which may be related to a repression of PPAR-mediated transcriptional activity.

DISCUSSION
BC-induced muscle dysfunction is a common problem of unclear etiology with few therapeutic options. Here we sought to identify possible mechanisms of fatigue that are generalizable across BC subtypes and are independent of treatment status by assessing statistical relationships between patients' clinical characteristics and overall skeletal muscle gene expression.
In support of our laboratory's previous publications 14, 15 , we found that women with three molecular subtypes of BC, those being ERPR, TN, and TP BC, exhibit overall similarity in muscular gene expression. Remarkably, patients with tumors overexpressing HER2/neu in the absence of ER and PR expression exhibited a markedly different muscular gene expression profile. However, gene expression data from all subtypes pointed to similar pathway-level dysregulation, indicating that the mechanisms leading to muscle fatigue in BC patients may indeed be generalizable across subtypes. In assessing the commonalities across patients, we observed significant dysregulation of metabolic pathways in all groups of BC patients, and proteomic analysis of patients with HER2/neu-overexpressing tumors showed decreased abundance of nearly all proteins involved in the mitochondrial electron transport chain. Because mitochondrial density in skeletal muscle has been shown to correlate with abundance of ETC complex proteins and skeletal muscle oxidative capacity 16 , it is likely that BC patients also have decreased mitochondrial density in their skeletal muscle as well as decreased oxidative capacity.
We propose then that BC-secreted factors induce muscle dysfunction by abrogating oxidative capacity via alteration of mitochondrial biogenesis, mitophagy, or fission/fusion dynamics. Both in vitro and in vivo assays confirm that factors from BC cells alter skeletal muscle ATP content and/or aerobic ATP production, perhaps via dysregulation of the PPAR-signaling pathway.
Our laboratory's interest in the PPAR family of proteins arose from our previous RNAsequencing analysis of muscle from BC patients and PDOX-bearing mice, which our group found to recapitulate the clinical phenotype of increased muscle fatigue without muscle atrophy or bodyweight loss 15 . These proteins have demonstrated roles in whole-body energy regulation, are critical regulators of mitochondrial function in multiple tissues, and are targets of multiple FDA-approved agents in the treatment of type 2 diabetes and hyperlipidemia. Among the three PPAR isoforms, we identified PPARG as a key regulator in BC-induced muscle dysfunction observed in PDOX mice 15 . PPARG is a ligand-activated nuclear receptor that, upon activation by a variety of endogenous and synthetic lipids, forms complexes with retinoid X receptor (RXR) and cofactors such as the peroxisome proliferator-activated receptor-γ coactivator 1α (PGC1α), and stimulates transcription of downstream genes. This relationship with PGC1α is particularly relevant, as PGC1α is known to be a master regulator of mitochondrial biogenesis in several tissues, including skeletal muscle 17 . PGC1α also participates in the regulation of other metabolic processes, including gluconeogenesis, muscle fiber-type specification, and control of antioxidant expression [18][19][20] . Therefore, the interaction between PPARG and PGC1α clearly has the potential to impact muscle function through mitochondrial mechanisms. In addition to the potential development of muscle-intrinsic mitochondrial dysfunction, a systemic consequence of dysregulated PPARG is the development of insulin resistance (IR) 21 , which is commonly associated with type 2 diabetes, obesity, and metabolic syndrome, and appears to have a bidirectional relationship with BC, with individuals with IR at greater risk of BC and BC survivors at an increased risk of IR 22,23 . We propose that the development of mitochondrial dysfunction and IR, secondary to muscular PPARG downregulation by BC, creates an environment that facilitates the development of muscle fatigue through several mechanisms, including decreased mitochondrial ATP production as well as dysregulated glucose and lipid metabolism. Pharmacological restoration of PPARG function results in the induction of a number of genes involved in insulin signaling, as well as glucose and lipid metabolism, and PPAR agonist drugs including the TZDs have shown a remarkable ability to restore insulin sensitivity in insulin-resistant conditions. Because our predictions have implicated many related metabolic pathways, we hypothesize that the development of BC induced muscle fatigue constitutes a pathology similar to type 2 diabetes/metabolic syndrome. If repression of PPAR signaling is indeed central to BC-induced muscle dysfunction, the numerous FDA-approved PPAR-agonists could address this unmet need in clinical oncology.
A particularly surprising result in the present study was that no clinical data aside from BC molecular subtype exhibited significant correlation with skeletal muscle gene expression, including TNM staging and history of chemotherapy, radiation, or immunotherapy. Additionally, patients in the HER2 group consistently exhibited a decreased abundance of mitochondrial proteins in their muscle tissue, despite significant differences in prior treatments, and importantly, we observed these responses consistently across all five biopsies from this patient group, despite significant differences in anti-cancer treatment history. At the date of biopsy collection, one patient was entirely treatment naïve, one patient had completed chemotherapy for BC 10 years prior, and the remaining 3 patients received different multi-agent neoadjuvant chemotherapy in the months prior to surgery. Therefore, we propose that changes in skeletal muscle physiology seen in BC are due to tumor-derived factors rather than side-effects of therapies or other patient-specific factors. This hypothesis is supported by our in vitro conditioned media experiments where tumor-derived factors directly repressed mitochondrial respiratory capacity in differentiated muscle cells. Additionally, neither weight loss nor body composition were predictive of skeletal muscle gene expression, suggesting that gene expression changes observed in BC patients are unrelated to muscle atrophy or cachexia and may instead be reflective of muscle dysfunction.
Because albumin is often used by physicians as a measure of patients' nutritional status and mortality risk [24][25][26][27] , we assessed serum albumin in our patient cohort and assessed its relationship to skeletal muscle gene expression and weight change. In our analysis, serum albumin was not found to be predictive of weight change, and in a larger sample was found to be only weakly predictive of cachexia risk. Yet, it was unexpectedly predictive of skeletal muscle gene expression. This indicates that serum albumin may be a useful biomarker of muscle function in the absence of cachexia, a possibility supported by previous literature connecting serum albumin with muscle strength 28 and insulin resistance 29 in other clinical contexts.
Prospective studies directly addressing the relationship between skeletal muscle function, gene expression, and serum albumin would be required to validate the clinical utility of serum albumin in identifying those at risk of cancer-induced muscle dysfunction.
In line with our previous publications 14, 15 , we report that transcriptional responses to the ERPR, TP and TN subtypes of BC are similar in terms of skeletal muscle gene expression, while muscle biopsies from patients with tumors overexpressing HER2/neu in the absence of ER and PR exhibit an unusual degree of uniqueness in terms of gene expression. We identified a strong signal for BC-induced mitochondrial dysfunction in BC patients, PDOX-bearing animals, and in in vitro assays, and proteomic analysis showed decreased protein abundance of nearly all components of the mitochondrial electron transport chain in muscle biopsies from patients with HER2/neu-overexpressing tumors relative to control. Further we found no relationship between various BC-related treatments (surgery, chemotherapy, radiotherapy) and changes in skeletal muscle gene expression. However, serum albumin was predictive of skeletal muscle gene expression without being predictive of weight loss, suggesting that serum albumin may be a useful indicator of BC-induced skeletal muscle dysfunction. Overall, these data indicate that all BC subtypes induce dysfunction in mitochondrial respiration in skeletal muscle independent of molecular subtype, and this effect appears to be independent of anti-cancer treatments. These findings call for prospective studies assessing interventions to support skeletal muscle function in the presence of BC-induced mitochondrial dysfunction. In a separate analysis of de-identified electronic medical records (EMRs) from female BC patients, body mass and standing height data were acquired for 5,201 individuals for calculation of BMI. Within this population, final analyses were completed in 3,001 patients having at least two measurements for weight in addition to at least one record for serum albumin.

RNA-Sequencing. Pectoralis major muscle biopsies were acquired intraoperatively and stored
in Invitrogen RNAlater Solution (Thermo Fisher, San Jose, CA) overnight at 4 o C and then at -80 o C until processing. RNA was isolated, assessed for quality, and utilized to construct libraries for RNA-Seq as previously reported 15 . Completed libraries were sequenced on one lane of the HiSeq 1500 with PE50 bp reads. Subsequently, Salmon was used for transcript-level abundance estimation, with both gcBias and seqBias set, and libType A 30 . Transcript-level abundance estimates were summarized to the gene-level using tximport 31 . The resulting counts matrix was scaled to library size using edgeR 32 and filtered to remove genes without a counts-per-million (CPM) value > 1 in 3 or more samples. Log-transformed CPM values for the 10,000 genes with highest variance were used as input for heatmap creation using gplots 33 . CPM values were logtransformed and scaled prior to distance matrix computation. The resulting distance matrix was then input for classical multidimensional scaling with k=3 and visualized in 3-dimensions using car 34 . Differential gene expression analysis was conducted using DESeq2 35 . PDOX-bearing animals were euthanized approximately 30 days after reaching a tumor volume of 200 mm 3 . Control female NSG mice of similar age (n=4) were euthanized at the same time as tumor-bearing animals and tissues were processed identically in both groups. Immediately after death, both quadriceps muscles from each mouse were quickly removed and interfibrillar (IFM) and subsarcolemmal (SSM) mitochondria were isolated separately according to previously described methods [39][40][41][42]

In vitro conditioned media (CM) metabolic analysis. C2C12 cells were plated into Agilent
Seahorse XF24 (Agilent Technologies, California, USA) plates and differentiated by confluence for 3 days. Meanwhile, EpH4-EV, EO771, NF639, and C2C12 cells were plated at approximately 15% confluence in separate 10cm dishes for 72 hours. The 72-hour CM was then removed from all cell lines, centrifuged at 1,500 RPM for 10 minutes, and then the supernatants were collected, diluted 1:3 in fresh growth media, and applied to the differentiated C2C12 cells in Seahorse assay plates for 48 hours (n=10 wells per treatment condition) prior to conducting the Agilent Seahorse XF Cell Mito Stress Test protocol according to manufacturer's instructions.
In vitro CM PPAR-reporter assays. HEK293 cells were transfected with PPRE-H2b-eGFP 43 (Addgene #84393) using Invitrogen Lipofectamine 3000 (Thermo Fisher), selected with 500 ng·uL -1 Gibco geneticin (Thermo Fisher) for 20 days, and flow-sorted to select the cells expressing GFP. The resulting HEK293-PPRE-H2b-eGFP cell line was plated at approximately 15% confluence in a 24-well plate. The following day, cells were imaged using the BioTek Cytation 5 Cell Imaging Multi-Mode Reader (Agilent Technologies) to collect baseline GFP intensity, and 72-hour conditioned media from HEK293, EpH4-EV, EO771, and NF639 well lines was applied to the 24-well plate, using individual wells as biological replicates (n=6 wells per treatment condition). Cells were then incubated in the conditioned media under normal culture conditions for 24 hours, at which point GFP intensity was measured using identical imaging settings as the baseline collection. Mean cellular GFP intensity per well was calculated using Gen5 Microplate Reader and Imager Software (Agilent Technologies) after background flattening and thresholding, which were set consistently across all images.

Statistical Analyses. Clinical information:
Patients' trends of weight change over time were calculated for each individual patient by fitting a simple linear regression line to their weight at each date in the EMR, normalized such that each patient's first weight record equaled 100, resulting in a slope representing their approximate percentage of body weight change per day. 19 patients were identified as outliers in terms of daily weight change. 13 of these patients were excluded due to having a limited observation period (< 15 days). Each patient's first albumin measurement was then obtained and the rate of daily weight change was regressed on the patients' first albumin measurements. Pearson's correlation coefficients and p-values were calculated and plotted using ggpubr 44 in R v3.6.1 45 .
Logistic regression analysis was utilized to determine whether serum albumin was predictive for a rate of weight change consistent with cachexia (i.e. < -0.027% per day to reach 5% weight loss in 6 months 1,2 ). Omnibus model fit was assessed by Chi-square test and effect size was calculated using Nagelkerke's pseudo-R 2 . The receiver operating characteristic curve (ROC) test was conducted, with preferred sensitivity and specificity > 0.7. No point on the ROC satisfied these conditions. The fitted logistic regression model was used to predict whether each patient would exhibit a rate of weight loss consistent with cachexia, and these predictions were compared to the actual data to create a confusion matrix for determination of sensitivity, specificity, and positive predictive value. Differential gene expression analysis was conducted using DESeq2 35 . Input data consisted of transcript-level abundance estimates from Salmon summarized to the gene level using tximport 31 . Two differential expression analyses were run: one comparing the group of BC patients to control patients, and another comparing the BC patients by subtype to control patients, with the null hypothesis rejected at FDR < 0.10. Because molecular subtype was the only variable that yielded statistical significance for predicting gene expression in the multivariate regression model described above, no clinical characteristics were assessed in the differential expression analysis.

RNA-Seq and clinical correlates:
Proteomics: Differential protein expression analysis was conducted using DEP 46 .
Protein expression values were first filtered to remove known contaminants and proteins with missing expression values in more than 1 sample per group, then normalized, and then background-corrected using variance stabilizing transformation. Remaining missing values were imputed using k-nearest neighbor, after determining that the small number of missing values were likely missing at random. Differential expression analysis using linear models and empirical Bayes statistics was then conducted on the imputed dataset, with the null hypothesis rejected at FDR < 0.05. In vitro and in vivo validation assays: ATP content in each mitochondrial subpopulation in the PDOX muscle were compared to control using the Mann-Whitney U test.
The null hypothesis was rejected at p < 0.05. In the conditioned media metabolic experiments, one-way ANOVA was used to compare the rate of oxygen consumed in ATP production as a percentage of basal oxygen consumption between the four conditioned media treatment groups followed by two-tailed Student's t-tests with Bonferroni correction comparing each treatment group to CON-Muscle. The null hypothesis was rejected at p < 0.05. This experiment was conducted as reported twice with similar results, and results from the first analysis are reported.
In the PPAR-responsive reporter assays, mean GFP intensities in each well were normalized to account for differences in baseline GFP expression between wells. Normalized GFP intensities at 24 hours were compared to normalized baseline measurements using a paired samples two-tailed t-test by treatment group with Bonferroni correction for multiple comparisons. The null hypothesis was rejected at p < 0.05. This experiment was conducted as reported twice with similar results, with results from the first analysis reported. This analysis includes n=5 for NF639-treated cells due to a technical problem during baseline image capture that resulted in the loss of one image.         Num-numerator; Den-denominator; Df-degrees of freedom; LBM %-Lean body mass as a percentage of total body mass; Ave-Average; T-tumor; N-lymph node; M-metastatic; Txscourses of treatments; * p < 0.05, o p < 0.10