The study included 110 patients with chronic heart failure from 11 hospitals (The Second Affiliated Hospital of Beijing University of Chinese Medicine, Zhengzhou Hospital of TCM in Henan Province, The Affiliated Hospital to Changchun University of CM, Guang'anmen Hospital which was Affiliated to China Academy of Chinese Medicine Sciences, Hubei Provincial Hospital of TCM, Wuhan Hospital of TCM, Hospital of T.C.M.S Shijingshan District, Beijing Changping Hospital, Hospital of T.C.M.S Beijing Miyun County, Beijing Changping Hospital of Integrated Chinese and Western Medicine, TCM Hospital of Beijing Huairou and TCM Hospital of Tongzhou District) in China between May 2010 and December 2014. During the same period, 54 normal healthy participants of matched age were included from medical center of Dongzhimen Hospital as controls.

Diagnostic criteria of CHF: 2007 China Guideline for the Diagnosis and Treatment of CHF (Chinese Society of Cardiology of Chinese Medical Association and Editorial Board of Chinese Journal of Cardiology, 2007). Heart function standard: The Criteria for Diagnosis and Treatment of Heart Disease first published by the New York Heart Association (NYHA) (Feinstein, 1964). Syndrome differentiation of CHF patients followed the TCM differentiation standard: according to the Guiding Principles for the Clinical Study of New Drugs in Traditional Chinese Medicine released in 2002 (Zheng, 2002).

Primary heart disease in this research was Coronary Heart Disease (CHD), which could be diagnosed by coronary computed tomography, coronary angiography, limb-salvage Q wave for electrocardiogram (ECG), history of acute myocardial infarction, ECG test, radionuclide examination support, etc. The included patients also had no history of taking antihypertensive drugs and exhibited a blood pressure under 160/100 mmHg or of hypertension. Following symptoms and signs were observed with a history of CHD: fatigue, difficulty breathing, fluid retention (edema), left ventricular enlargement, NYHA functional classification II or III, and end systolic volume of left ventricular increase and left ventricular ejection fraction (LVEF) < 40. All subjects between 40 and 75 years old; If there was a “No” answer for any issue, the subject could not enter into this research.

Patients with one of the following diseases were excluded: (1) serious valvular heart disease; (2) cardiomyopathy; (3) pericardial disease; (4) congenital heart disease; (5) cardiac shock; (6) acute myocardial infarction (AMI) within 4 weeks; (7) acute myocarditis or serious arrhythmia with variation in hemodynamics. Patients who suffer from pulmonary artery hypertension caused by cor pulmonale, pulmonary embolism, or stroke within a half year were also excluded. Patients who suffer from serious hepatic or renal deficiency. Patients who suffer from diseases of blood system or malignant tumor. Patients who suffer from diabetes mellitus with serious complications, hyperthyrea, or hypothyrea. Patients who suffer from infection. Pregnant women or women in lactation. Patients with mental disorders. Patients with infectious disease or who join in other trials within 2 months of the present study.

All qualified participants had signed informed consent prior to the enrollment. This study was conducted according to the principles of the Declaration of Helsinki and the principles of Good Clinical Practice.

All patients formed the chronic heart failure group, and were further divided into 3 groups according to TCM syndrome differentiation, including group of Qi deficiency syndrome (QD), group of Qi deficiency and Blood stasis syndrome (QB), and group of Qi deficiency and Blood stasis and Water retention syndrome (QBW). The normal healthy participants composed the NH group.

Besides, according to previous research (Wang et al., 2013a), the most syndrome of CHF was Qi deficiency and Blood stasis syndrome, so 64 of the participants with CHF of Qi deficiency and Blood stasis syndrome (QB group) were enrolled in a randomized double-blind controlled clinical trial. According to the 2007 China Guidelines for the Diagnosis and Treatment of Chronic Heart Failure (Chinese Society of Cardiology of Chinese Medical Association and Editorial Board of Chinese Journal of Cardiology, 2007), all patients in the two groups received the standardized western medicine treatment, which includes angiotensin-converting enzyme inhibitors (ACEI) or angiotensin receptor blockers (ARBs), b-blockers, and diuretics. Except the standardized western medicine treatment, subjects with the CHM treatment were defined as the CHM group, and subjects with the placebo intervention were defined as the placebo group. The subjects were treated with CHM granules, made from Astragalus membranaceus (Fisch.) Bge. var. mongholicus (Bge.) Hsiao (huangqi, 60 g), Codonopsis pilosula (Franch.) Nannf (dangshen, 15 g), Salvia miltiorrhiza Bge (danshen 15 g), Paeonia lactiflora Pall (chishao, 15 g), Prunus davidiana (Carr.) Franch. (taoren, 10 g), and Carthamus tinctorius L. (honghua, 10 g) or placebo granules twice per day for 4 weeks. The CHM granules and placebo granules were offered by Beijing Kang Rentang pharmaceutical industry (Beijing, China). Finally, 31 patients were in CHM group, and 33 were in placebo group.

Plasma samples were collected on morning of the first and the twenty-eighth day after enrollment and were immediately frozen at −80°C. Six minutes of walking test (6 MWT) and echocardiography were also detected at the same time. Both the observer of the 6 MWT and the echo technician were blind to the allocation of the patients.

After several preliminary experiments by different mixtures of methanol, acetonitrile, and methanol/acetonitrile (1:1), finally the extraction solvent was acetonitrile. By mixing equal amounts of plasma (30 μL) from all samples, pooled quality control samples were prepared to ensure data quality for metabolic profiling. Detailed sample methods were shown in the Supplementary Material.

This research was approved by the Institutional Committee on Human Research of Beijing University of Chinese Medicine (No. 200807007), and this trial is registered with NCT01939236 (Luo et al., 2015).

The liquid chromatographic separation of processed plasma was conducted on a 100^*2.1-mm ACQUITY UPLC ° R BEH C18 1.7-um column (Lot No. 0252350221) using an ACQUITY Ultra Performance LC (Waters Corporation, Milford Massachusetts, USA), whereas mass spectrometry was performed on a SYNAPT G2 Quadrupole-Time of Flight system (Waters Corporation, Milford Massachusetts, USA). With a random-number generator in Excel (Microsoft, Redmond, Washington), 3 sequences for samples in discovery phase, validation phase and placebo group were assigned by the study administrator. During analyses of the sample sequence, after each 20 injections, then 1 quality control sample was run to check the stability of system. The acquired mass spectrometry data were exported to data format (.centroid) files by Masslynx Software (version 4.0, Waters Corporation). Data pre-treatment procedures, such as nonlinear retention time alignment, peak discrimination, filtering, alignment, matching, and identification, were performed in Progenesis QI (34 Maple Street, Milford Massachusetts, 01757, USA), and retention time and the mass-to ratio data pairs as the parameters for each ion. The UPLC-QTOF-MS characteristic chromatogram of CHM granules (6 herbs) was identified, and detailed method was shown in the Supplementary Material. Disturbed metabolites and metabolic pathways were identified by open database sources, including Human Metabolome Database (http://www.hmdb.ca/), METLIN (https://metlin.scripps.edu/), KEGG (http://www.genome.jp/kegg/pathway.html), and MetaboAnalyst (http://www.metaboanalyst.ca/faces/home.xhtml).

By SIMCA (version 14.1, Umetrics AB, Umea, Sweden), the orthogonal projection to latent structure-discriminant analysis (OPLS-DA) model was applied to compare the differences in metabolic profiles between CHF, QD, QB, QBW, and NH group, and P-values, from the Student's t-test or Welch's t-test, depended on whether the variances of two populations are assumed to be equal by Levene test. The significant metabolic changes after treatments of Chinese Herb Medicine or Placebo were identified by Partial Least Squares Discriminant Analysis (PLS-DA) and OPLS-DA, and P values were from paired samples t-test. R2Y (cum) was used to estimate the “goodness of fit” of the model, and Q2(cum) to estimate the ability of prediction. All P values for each metabolite in 6 comparisons were corrected by Bonferroni correction and the false discovery rate (FDR) method for multiple testing across all metabolites within each comparison. Matplotlib version 2.0.2 and SciPy version 0.19.1 were used to conducted the heat maps and hierarchical cluster analyses. A standard linear Support Vector Machine (SVM) Classification (Bledsoe et al., 2016): f(x)=sgn[(w∗·x)+b∗]=sgn{∑i=1Nai∗yi(xi·x)+b∗},                             (w∗x)+b→0f(x)={+1, x∈ω1−1, x∈ω2

And detailed method was shown in the Supplementary Material. And receiver operating characteristic (ROC) analysis were used for diagnosis of different CHF syndromes and validation of CHM treatment effects. To correct potential overfit of 2 OPLS-DA comparisons of Prior & Post-CHM group and Prior & Post-Placebo group, five-fold Cross Validation was conducted to verify the specificity and sensitivity of the classifier. The data set was divided into 5 approximately equal bins and the models were evaluated 5 times. Statistical analyses were performed using python software version 2.7 (https://www.python.org/) with scikit-learn version 0.18.2 (http://scikit-learn.org). An adjusted p < 0.05 was considered statistically significant.

Representative based peak intensity (BPI) chromatograms of 4 samples with NH, QD, QB, and QBW, and 5 samples with NH, Prior & Post-CHM group, and Prior & Post-Placebo group obtained are shown in Supplementary Figures 1, 2. After peak alignment and removal of missing values, 442 negative-mode features were detected. After the OPLS-DA test, in SIMCA software, the ions with variable importance in the projection (VIP) values >1.0 and p-value <0.05 were considered the potential differential metabolites.

Presently clinical diagnosis of different TCM syndromes of CHF depends on personal experience of clinical doctors (Xu et al., 2016), so accurate diagnosis of the different stages of CHF, based on objective biomarkers in blood, is fundamental and significant for personalized treatment. So CHF and NH was compared to identify and characterize special metabolites of patients with CHF in discovery phase. QD, QB, and QBW were compared with NH to identify and characterize special metabolites of different Chinese syndromes in discovery phase.

Clear differences were obtained for the following: CHF vs. NH, cumulative R2Y at 0.923 and Q2 at 0.892 (Figure 2A); QD vs. NH, R2Y at 0.976 and Q2 at 0.94 (Figure 2B); QB vs. NH, R2Y at 0.967 and Q2 at 0.94 (Figure 2C); and QBW vs. NH, R2Y at 0.957 and Q2 at 0.868 (Figure 2D). With VIP values >1.0 and adjusted p-value < 0.05, the metabolites for each comparison appear in Table 2. Six specific metabolic biomarkers distinguished CHF from NH, 4 for QD from NH, 7 for QB from NH, and 5 for QBW from NH. PCA (Principal Component Analysis) result among QD, QB, QBW, and NH was shown in Supplementary Figure 3, and QD was the predominate Zheng contributing to CHF progression far away from NH.

The ROC presentations, on the basis of Support Vector Classification (SVC) of Support Vector Machine (SVM) which trained by each biomarker panel from the discovery phase, appear in Figure 3; the areas under the curve (AUC), sensitivity, and specificity are 0.901, 97.1%, and 80.6% for CHF vs. NH (n = 106; Figure 3A); 0.867, 97.1%, and 80.0% for QD vs. NH (n = 49) (Figure 3B); 0.876, 97.1%, and 79.5% for QB vs. NH (n = 73; Figure 3C); and 0.964, 97.1%, and 88.9% for QBW vs. NH (n = 52; Figure 3D), respectively.

With highest prediction specificity and sensitivity of the ROC, the optimal cut-off values were 0.1684 for CHF vs. NH, 0.2934 for QD vs. NH, 0.3403 for QB vs. NH, and 0.4763 for QBW vs. NH in the discovery phase. Different stages of CHF in the verification phase were predicted by the trained classifier and cut-off values. Predictive value was 85.0% for CHF vs. NH in the verification phase (Figure 3E); 95.1% for QD vs. NH in the verification phase (Figure 3F); for 99.1% QB vs. NH in the validation phase (Figure 3G); and 90.1% for QBW vs. NH in the validation phase (Figure 3H).

There was no significant difference of 6 MWT and ejection fraction between Prior-CHM group and Prior-Placebo group, and 4 weeks later, both changed significantly between Post-CHM group and Post-Placebo group, as shown in Figure 4.

The UPLC-QTOF-MS characteristic chromatogram of CHM granules (6 herbs) was shown in Figure 5.

There were significant curative effects of CHM on different diseases, but usually with indistinct metabolic mechanisms (Lao et al., 2014). According to 2 PLS-DA comparisons, Prior-CHM vs. Prior-Placebo vs. NH, R2Y at 0.586, Q2 at 0.524 (Figure 6A), Post-CHM vs. Post-Placebo vs. NH, R2H at 0.514, Q2 at 0.39 (Figure 6C), and 2 OPLS-DA comparisons, Prior-CHM vs. Prior-Placebo, R2H at 0.505, Q2 at 0.0866 (Figure 6B), Post-CHM vs. Post-Placebo, R2H at 0.845, Q2 at 0.16 (Figure 6D), after the 4 weeks treatment, the signs of patients in CHM group left the Placebo group to the NH group. Then Post-CHM group and Prior-CHM group was compared to identify and characterize special metabolites of combined treatment. Post-Placebo and Prior-Placebo was compared to identify and characterize special metabolites of basic Western Medicine treatment. The metabolites with VIP values >1.0 and adjusted p-values <0.05 for 2 OPLS-DA comparisons appear in Table 3 and there were 11 metabolic biomarkers for distinguishing Post & Prior-CHM group, and 3 for Post & Prior- Placebo group. By comparison between the 11 and 3 metabolic biomarkers, there were 9 special metabolic biomarkers for Prior & Post-CHM group.

The ROC presentations, on the basis of Support Vector Classification (SVC) of Support Vector Machine (SVM) which trained by on the basis of the logistic regression of each biomarker panel, appear in Figure 6; the AUC, sensitivity, and specificity are 0.777, 64.5%, and 96.8% for Post-CHM vs. Prior-CHM (n = 62; Figure 6E); and 0.804, 63.6%, and 90.9% for Post-Placebo vs. Prior- Placebo (n = 66; Figure 6F), respectively.

Then five-fold Cross Validation was conducted to verify the specificity and sensitivity of the classifier OPLS-DA results with Python packages “sklearn,” the result showed the optimal cut-off value was 0.3631, and true positive rate (TPR), false positive rate (FPR) were 96.7 ± 4.2, 19.2 ± 2.1% for Post-CHM vs. Prior-CHM (n = 62; Figure 6G); and 0.2614, 90.9 ± 1.2%, and 24.5 ± 2.2% for Post-Placebo vs. Prior-Placebo (n = 66; Figure 6H), respectively.

At the same time, there existed a certain correlation between specific metabolites determined in CHF patients with QB Zheng (n = 7) and CHM efficacy in QB patients (n = 11), evaluated by Chi-squared test with P = 0.0557.

According to 4 OPLS-DA comparisons between Normal healthy group and different CHF groups, total 13 differential metabolites were identified. Average normalized quantities of the 13 differential metabolites were shown in the heat map Figure 7 among Normal healthy group and the CHF groups. Among them, 1 metabolites were elevated; 12 metabolites decreased in order according to the following groups: NH, CHF, QD, QB, and QBW.

Moreover, we constructed the metabolism-protein networks and identify 36 related proteins, such as Calcium-dependent phospholipase A2, Group IIF secretory phospholipase A2, and Cytosolic phospholipase A2 in the most important Arachidonic acid metabolism pathway (Figure 8). Up-regulated and down-regulated metabolites determined were shown with red and blue points, and undetected metabolites are shown as gray points. Direct related enzymes were shown in green square frames, and others were with no color. Detailed information of points in Figure 8 could be found at http://www.kegg.jp/kegg/tool/conv_id.html.

The total 36 related proteins are shown in the Supplementary Material. The detailed information of Proteins and Pathways are shown in Supplementary Table 2.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.