The clinical samples were collected within the multicentricObelics trial (NCT01718873), which investigated different schedules of bevacizumab in combination with oxaliplatin plus fluoropyrimidines (FOLFOX-6 or OXXEL) regimens for treating metastatic colorectal cancer (mCRC) patients. Patient recruitment and sample collection were approved by the ethics committee of the National Cancer Institute of Naples – Fondazione G. Pascale. Written informed consent was obtained from all of the patients in accordance with the Declaration of Helsinki for the use of human biological samples for research purposes. Blood samples from the patients were obtained after overnight fasting. Plasma samples were retrospectively selected from patients in the Obelics trial enrolled at Pascale Institute who underwent surgery for resection of only liver metastases; patients with severe surgical complications were excluded. For thirty patients (15 in the experimental arm and 15 in the standard arm) with these characteristics, blood samples were collected at different time points (baseline and whenresectability was defined at the response evaluation). The patient characteristics are shown in Table 1 , and the patient inclusion process is reported in Figure 1 . The patients were subdivided into responder (R) and non-R (NR) according to 1-year disease-free survival (DFS): ≥ 1-year (R, n = 12) and < 1-year (NR, n = 18).

Blood samples were collected in plasma preparation Vacutainer tubes (BD Biosciences), centrifuged at 1500 × g for 10 min within 2 h of collection and then stored at -80°C until the day of analysis.

Vital tumor areas selected by a pathologist from the metastases resected from the liver of ten patients were frozen at -80°C until analysis. Normal liver tissues for the same patients were also collected and frozen.

Pathological tumor response was evaluated according to the 2010 American Joint Committee on Cancer (AJCC) TRG system: TRG 1, no viable cancer cells; TRG 2, single or few cancer cells; TRG 3, fibrosis predominating over residual cancer; TRG 4, predominant viable cancer cells outgrowing the fibrosis (39).

All plasma samples were prepared for NMR analysis by mixing 330 μL of plasma with 300 μL of PBS (containing 10% v/v H₂O) and 70 μL of reference standard D₂O solution containing 0.1 mM sodium 3-trimethylsilyl [2,2,3,3-2H4] propionate (TSP). Samples were inserted into an NMR tube, and all of the spectra were recorded using a Bruker Avance III HD (600 MHz) NMR spectrometer operated at a 599.97 MHz ¹H resonance frequency and equipped with a TCI cryoprobe. To attenuate the broad NMR signals from the slowly tumbling molecules in the lipids and proteins, a standard Carr−Purcell−Meiboom−Gill (CPMG) pulse sequence was used to record the ¹D spin−echo spectra. To suppress the water peaks, a CPMG presaturation pulse sequence was used with the equation -RD-90°-(t-180°-t) n - ACQ, where RD is the relaxation delay of 2 s; 90° and 180° represent the pulses that trip the magnetization vector; t is the spin−echo delay; n represents the number of loops; and ACQ is the data acquisition period. In our experiment, the data points were acquired using 256 transients.

Each 100 µL plasma sample was resuspended in 170 µL of H₂O and 700 µL of methanol. Then, 350 µL of chloroform was added, and the samples were mixed on an orbital shaker on ice for 10 min. Then, 350 μL of H₂O/chloroform (1:1, v/v) was added to each sample and centrifuged at 4000 rpm for 10 min at 4°C. Thereafter, the lipidic (apolar) phases were collected and evaporated. Then, these fractions were dissolved in 700 µL of deuterated chloroform containing 0.1 mM TSP and inserted into NMR tubes. ¹H-NMR spectra at 300 K were acquired using a Bruker Avance III HD (600 MHz) NMR spectrometer equipped with a TCI cryoprobe and zgesgp as the pulse sequence. The data points were acquired using 512 transients.

Frozen tumor tissue samples were cut to an appropriate size (mean weight: 10 mg) and placed in 50 μL disposable rotor insets filled with reference standard D₂O solution containing 0.1 mM TSP for the field lock. Inserts with frozen samples were transferred to 4 mm zirconium rotors. Samples were kept at 277 K to slow down tissue degradation. Spectra were acquired by a Bruker Avance III HD (600 MHz) NMR spectrometer equipped with a high resolution magic angle spinning (HRMAS) probe using a magic angle spinning rate of 4 kHz and CPMG presaturation pulse sequence. A total of 256 scans were collected.

All of the ¹H NMR spectra were manually phased and baseline-corrected and referenced to the CH₃ resonance of TSP at 0 ppm. The spectral 0.50-8.60 ppm region of the ¹H-NMR spectra was integrated in buckets of 0.04 ppm by the AMIX package (Bruker, Biospin, Germany). In detail, we excluded, in the case of the polar spectra, the water resonance region (4.5-5.2 ppm) during the analysis and normalized the bucketed region to the total spectrum area using Pareto scaling by the MetaboAnalyst v5.0 tool (40).

A large panel of cytokines, chemokines, and growth factors were evaluated in plasma collected when resectability was defined using LuminexxMAP-based multiplex bead-based immunoassay technology. In detail, the concentrations of β-NGF, CCL2 (MCP-1), CCL3 (MIP-1α), CCL4 (MIP-1α), CCL7 (MCP-3), CCL11 (Eotaxin), CTACK (CCL27), CXCL1 (GRO-α), CXCL9 (MIG), CXCL10 (IP-10), CXCL12 (SDF-1α), FGFbasic, G-CSF, GM-CSF, HGF, IFN-α2, IFN-γ, IL-1α, IL-1ß, IL-1ra, IL-2, IL-2Rα, IL-3, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, IL-10, IL-12 (p40), IL-12 (p70), IL-13, IL-15, IL-16, IL-17, IL-18, LIF, M-CSF, MIF, PDGF-ßß, RANTES, SCF, SCGF-ß, TNF-α, TNF-β, TRAIL and VEGF were determined using the Bio-Plex Pro™ Human Cytokine Screening Panel, 48-Plex assay and a Bio-Plex array reader (Luminex, Austin, TX, USA) that quantifies multiplex immunoassays in a 96-well format with very small fluid volumes. The analyte levels were calculated using a standard curve with software provided by the manufacturer (Bio-Plex Manager Software).

Pathway analysis of the modulated metabolites was performed using the Metaboanalyst 5.0 tool (40). In detail, we calculated the centrality through Pathway Impact, a combination of the centrality and pathway enrichment results. Metabolites were selected by evaluating both VIP values > 1 in the class discrimination and correlation values >0.8. Moreover, the Homo sapiens pathway library was chosen and analyzed using Fisher’s exact test for overrepresentation and relative betweenness centrality for pathway topology analysis.

The sparse partial least squares-discriminant analysis (sPLS-DA) algorithm was applied to explain the maximum separation between the defined class samples in the data. Score and loading plots were used to highlight and assess the role of X-variables (NMR signals and cytokine concentrations) in the classification models and, hence, to identify the top 10 significant NMR signals and cytokines. In detail, for the loading plot, we set H = K - 1, where H is the number of dimensions and k is the number of variables to select on each dimension (41). The significant NMR signals were assigned to metabolites and lipids using the reference metabolite spectra from the HMDB database (42).

The levels of proton signals were normalized to the total spectrum area using Pareto scaling with the MetaboAnalyst v5.0 tool (40). The average rate of change (Δ) values were obtained considering for each metabolite the ratio between [the average level of the proton signals at the response evaluation] and [the average level of the proton signals at baseline]×100 in the R and NR patient groups.

Receiver operating characteristic (ROC) curves were calculated for metabolites/lipids/cytokines that were found to be significantly correlated with DFS ≥ 1 year by the Biomarker Analysis tool of Metaboanalyst v5.0 (40). The area under the curve (AUC) was used to assess the accuracy. The 95% confidence intervals (CIs) were calculated to compute the optimal cutoffs for any given feature (significant metabolites, lipids and cytokines).

DFS was defined as the time from liver metastasis resection to the date of progression or death, whichever occurred first. Patients who did not progress were censored on the date of the last follow-up visit. OS was defined as the time from randomization to the date of death. Patients alive at the time of the final analysis were censored on the date of the last follow-up information available. DFS and OS curves were estimated according to the Kaplan−Meier method, and differences were evaluated with the log-rank test in MedCalc software (https://www.medcalc.org).

The Cox regression model was used to assess the role of the cutoff for metabolite parameters in predicting DFS and OS. Hazard ratios (HR) were derived from the Cox regression analysis, and their 95% confidence intervals (95% CI) were calculated using the proportional hazard model. Univariate analysis assessed the correlation of the baseline patient characteristics (sex, CEA, RAS status, TRG, treatment ARM and primary tumor location), metabolites, lipids and cytokines with DFS and OS. In all statistical tests, a p value less than 0.05 was considered significant. A multivariate analysis was performed using MedCalc software (https://www.medcalc.org) according to a backward elimination of factors showing a p value less than 0.05 in the univariate analysis.

Finally, biomarker analyses were performed on the basis of ROC curves for multiple biomarkers (metabolites, lipids and cytokines) using the support vector machine (SVM) algorithm by the module “Biomarker Analysis” in the Metaboanalyst 5.0 tool (40). The SVM classification algorithm aims to find a nonlinear decision function in the input space by mapping the data into a higher dimensional feature space and separating it by means of a maximum margin hyperplane (43).The input of an SVM is a training set S= (x₁, y₁)…,(x_n, y_n) of the vector of features (metabolites, lipids and cytokines) for each pt together with their known classes y_i∈ {R, NR}. The output of an SVM is a Model f: X → {R, NR} that predicts the class f(x) of any new pt (44). MetaboAnalyst’s SVM analysis is performed through recursive feature selection and sample classification using a linear kernel (45). Features are selected based on their relative contribution to the classification using cross validation error rates. The least important features are eliminated in the subsequent steps. This process creates a series of SVM models. The features used by the best model are considered to be important and are ranked by their frequencies of being selected in the model. In detail, in our study, different biomarker models were tested, and sample predictions were made. We evaluated 100 cross validations (CVs) to produce a smooth ROC curve, and the results were averaged to generate the plot. The average of the predicted class probabilities of each sample across the 100 cross-validations was produced (40).

Blood samples from a group of mCRC patients undergoing liver metastasis resection after first-line conversion oxaliplatin-based chemotherapy plus bevacizumab, enrolled within the Obelics trial (NCT01718873), were collected at baseline and at the time of response evaluation when resectability was established. Only those patients obtaining R0 resection without any severe surgical complications were considered, balanced between the two arms ( Figure 1 ). Blood samples available from thirty patients were analyzed by comparing, on the basis of disease-free survival (DFS) at 1 year: good responders (R), with a DFS ≥ 1 year, versus poor responders (NR), with a DFS < 1 year. The median DFS was 47.39 months (95% CI, 31.62-52.65) and 3 months (95% CI, 2.75-6.76) for R (n=12) and NR (n=18) patients, respectively. Of note, the median follow-up in this patient population was 39 months.

Baseline patient and tumor characteristics were well balanced between the two groups, although there was a statistically significant difference in the gender proportion ( Table 1 ). The median number of chemotherapy cycles administered before surgery was 6 (range 6-12) in both groups of patients; only 22% of NR and 25% of R patients received 12 cycles.

We first analyzed the plasma metabolic signature at the time of response evaluation when resectability was established. As reported in Figure 2 , sparse partial least squares discrimination analysis (sPLS-DA) (19.4% of the total variance), calculated on the ¹H NMR plasma spectra, clearly discriminated R from NR patients ( Figure 2A ), with a model accuracy of 63.3%, suggesting that the two study groups are distinctively different in terms of their plasma metabolic profiles. Moreover, the sPLS-DA score plot showed that the profiles of the R patients clustered together, whereas the NR patients were scattered, suggesting that patients with good outcomes may have similar metabolic profiles ( Figure 2A ).

An analysis of the PLS loading was then conducted to identify the metabolites found to be most relevant to the class separation (as reported in the Methods section). As shown in the loading plot of the top 10 NMR signals that were significantly different, it appears that R patients were characterized by lower plasma levels of isoleucine, 3-hydroxybutyrate, valine and hydroxyproline and higher levels of malate, histidine glutamate and gamma-aminobutyric acid (GABA) ( Figure 2B ).Notably, two NMR signals for both valine and histidine were reported, reinforcing the significance of their differential expression between the two patient groups.

Furthermore, these metabolites were used to perform a metabolite-set enrichment analysis. A complex interplay of several different metabolic pathways and metabolites was highlighted ( Figure 2C ; Table 1S ). In detail, aminoacyl-tRNA biosynthesis; butanoate metabolism; arginine and proline metabolism; alanine, aspartate and glutamate metabolism; valine, leucine and isoleucine biosynthesis; beta-alanine metabolism; nitrogen metabolism; valine, leucine and isoleucine degradation; histidine metabolism; synthesis and degradation of ketone bodies; and D-glutamine and D-glutamate metabolism emerged as playing a role in discriminating the plasma metabolic profiles of R from NR patients.

Next, to establish the optimal cutoff value for the metabolites selected by sPLS-DA, we performed ROC curve analysis, finding areas under the curve (AUC) values of the metabolites ranging between 0.63 and 0.74 ( Figure 1S ). Based on the metabolite parameter cutoff values, univariate and multivariate analyses were then conducted to evaluate metabolites potentially associated with DFS.

Univariate analysis demonstrated that sex (M vs. F) (HR, 2.90; 95% CI, 0.94–8.91; P=0.028) and tumor regression grade (TRG, 3-4 vs. 1-2) (HR, 2.80; 95% CI, 0.85–9.21; P=0.036) were significantly associated with DFS ( Table 2 ). No significant association was found between DFS and RAS status, CEA, primary tumor location, or treatment arm. Among the metabolites, both 3-hydroxybutyrate (HR, 4.35; 95% CI, 1.58–11.97; P=0.011) and histidine (HR, 0.23; 95% CI, 0.081–0-63; P=0.03) predicted DFS ( Table 2 , Figure 3A ). In detail, as shown by the Kaplan−Meier survival curves, only lower levels of 3-hydroxybutyrate (<cutoff) or higher levels of histidine (≥cutoff), evaluated before surgery, correlated with a more favorable DFS ( Figure 3A and Figure 2S ). These two metabolites were also the only parameters significantly correlated with overall survival (OS) in the univariate analysis ( Table 2S , Figure 3S ). Notably, in multivariate analysis, 3-hydroxybutyrate was the only parameter that significantly predicted DFS ( Table 2 ) and OS ( Table 2S ).

The prognostic role of TRG in patients with locally advanced rectal cancer treated with neoadjuvant chemoradiation has been explored and was previously confirmed as a predictor of disease-free survival in this setting (46, 47). Recently, TRG has been suggested as a useful prognostic factor in mCRC patients subjected to preoperative chemotherapy before metastasis resection (48). This observation was confirmed in our cohort of patients, where poor pathological responses (TRG 3-4) were associated with shorter DFS than complete and near-complete responses (TRG 1-2) ( Table 2 , Figure 3B ). When the metabolites histidine or 3-hydroxybutyrate were combined with TRG status, a striking separation of distinct categories was obtained ( Figure 3B ). Indeed, the two metabolites were far better predictors of DFS, with patients with either low 3-hydroxybutyrate or high histidine being associated with more favorable DFS outcomes, independent of TRG status ( Figure 3B ). Conversely, the patients with high 3-hydroxybutyrate or low histidine levels, although they had favorable prognostic TRG1-2, displayed a worse prognosis, further highlighting the powerful role of both metabolites in predicting DFS ( Figure 3B ). Similar data were also obtained for OS ( Figure 4S ). Kaplan–Meier curves of DFS and OS related to TRG in combination with other metabolites confirmed that only low 3-hydroxybutyrate or high histidine levels were associated with more favorable DFS/OS outcomes, independent of TRG status. The curves obtained for isoleucine were reported as representative example ( Figure S5 ).

When we considered the plasma metabolic signature at baseline, the sPLS-DA score plot demonstrated a less evident discrimination between R and NR patients ( Figure 6SA ). However, among the top 10 NMR signals contributing to class separation, the PLS loading plot again identified high levels of 3-hydroxybutyrate, hydroxyproline, and isoleucine as associated with NR patients ( Figure 6SB ), as also reported at the response evaluation time point for this group of patients ( Figure 2B ). In addition, high levels of 2-hydroxybutyrate, proline, trimethylamine and aspartate, and low levels of phosphoethanolamine and betaine, were among the most significant metabolites associated with NR patients ( Figure 6SB ). Overall, only a limited number of metabolic pathways, all included in the analysis reported in Figure 2C , were highlighted, confirming that the metabolic profiles of NR and R patients, at baseline, were less discriminated than those evaluated at the response evaluation time point ( Figure 6SC and Table 3S ).

Indeed, when we considered the normalized values of 3-hydroxybutyrate and histidine NMR signals at both baseline (B-R and B-NR) and at the response evaluation (R and NR), a clear rate of change in metabolic abundance over time from baseline was observed ( Figure 4 ). Notably, the average rate of change (Δ) for 3-hydroxybutyrate levels increased by 23.9% in NR and only 3.8% in R patients from baseline (p=0.042); conversely, the average Δ for histidine levels increased significantly by 71.8% in R vs. only 12.4% in NR patients from baseline (p=0.001) ( Figure 4 ). In other words, 3-hydroxybutyrate, whose high levels predicted a poor DFS at the response evaluation, increased during treatment only in the case of NR patients, whereas histidine, whose high levels at the response evaluation predicted a more favorable DFS, increased during treatment only in R patients. As reported above, in order to validate that only the 3-hydroxybutyrate and histidine were modulated in statistically significant way during treatment in NR or R patients, the plots obtained for isoleucine were reported as representative example ( Figure 7S ).

We conducted a parallel metabolomics investigation by ¹H HRMAS NMR analysis on 10 available patient-matched frozen resected liver metastasis tissues, 6 from NR and 4 from R patients. Notably, the sPLS-DA (48.3% of the total variance) calculated on the ¹H NMR tissue spectra clearly discriminated R from NR patients with a model accuracy of 80% ( Figure 5A ), suggesting that the two study groups are distinctively different in terms of their tissue metabolic profiles.

As shown in the loading plot ( Figure 5B ), the top 10 NMR signals were significantly different between the two patient cohorts. It is worth noting that higher levels of hydroxyproline and lower levels of GABA were observed in NR patients than in R patients, recapitulating the data obtained in the plasma samples (see Figure 2B ). We also conducted a targeted analysis in order to verify tissue levels of 3-hydroxybutyrate and histidine metabolites that were not among the top 10 NMR signals. Interestingly, although not statistically significant, higher levels of 3-hydroxybutyrate and lower levels of histidine were confirmed in NR tissues in line with the data reported on plasma samples ( Figure 8S ).

In addition, higher levels of methylamine, aspartate, proline, citrulline, lactate, 2-hydroxybutyrate, and threonine were observed in NR patients than in R patients.

We also compared the metabolic profiles of all liver metastases (LM) with the matched adjacent noncancerous tissues (NC), again demonstrating a clear separation into two distinct clusters ( Figure 9SA ), with high levels of proline, 2-hydroxybutyrate, aspartate and lactate associated with LM vs. NC ( Figure 9SB ), which interestingly were previously reported among the top metabolites discriminating NR vs. R in LM tissues ( Figure 5 ). Indeed, the metabolite-set enrichment analyses demonstrated three common pathways (aminoacyl-tRNA biosynthesis; alanine, aspartate and glutamate metabolism; valine, leucine and isoleucine biosynthesis) distinguishing both R vs. NR LM tissues and LM vs. NC tissues ( Figure 10SA ). Notably, some altered common pathways and metabolites (aminoacyl-tRNA biosynthesis; arginine and proline metabolism; alanine, aspartate and glutamate metabolism; and valine, leucine and isoleucine biosynthesis) distinguished R vs. NR patients in both plasma and metastatic tissue metabolomics, suggesting potential mechanistic correlations ( Figure 10SB ).

As suggested from all of the data reported above, plasma metabolic profiling at the response evaluation time point is able to discriminate R from NR patients, reflecting the impact of treatment. Therefore, further analyses conducted on plasma samples from our cohort of patients were limited to this time point.

To better define a metabolic signature predicting DFS, we also acquired ¹H NMR spectra on the lipidic fractions extracted from the thirty plasma samples. The sPLS-DA plot (54.8% of the total variance) grouped R and NR patients into two different clusters with a model accuracy of 65%, suggesting the presence of some lipidic proton signals with significantly different levels between the two patient groups ( Figure 6A ). The related loading plot showed that the NR group had lower levels of choline and higher levels of proton signals of fatty acids, cholesterol, triglycerides, omega-3 and phospholipids than the R group ( Figure 6B ), indicating increased plasma lipids in patients with poor DFS.

To determine the optimal cutoff value for the significant lipidic signals selected by sPLS-DA, we performed ROC curve analysis ( Figure 11S ), which showed AUC values ranging between 0.681 and 0.787. Based on the obtained parameter cutoff values, univariate analysis showed that lower levels (<cutoff) of cholesterol (HR, 9.72; 95% CI, 3.47–27.20; P=0.005), triglycerides (HR, 4.51; 95% CI, 1.59–12.81; P=0.003), or phospholipids (HR, 2.83; 95% CI, 1.01–7.90; P=0.042) were significantly associated with good DFS ( Table 2 ; Figure 7A ) and OS ( Table 2S ; Figure 12S ). Moreover, higher levels of choline (≥cutoff) were also found to correlate with good OS ( Table 2S ; Figure 12S ). None of these signals were statistically significant in the multivariate analysis for both DFS and OS ( Tables 2 and 2S ).

Most importantly, as also reported for histidine and 3-hydroxybutyrate ( Figure 3 ), we compared the DFS prediction potential of the lipid NMR signals with TRG, demonstrating that low cholesterol levels were a better predictor of DFS, independent of TRG status, and that both phospholipids and triglycerides plus TRG evaluation were better predictors of DFS than TRG alone ( Figure 7B ). Interestingly, the patients with high levels of either phospholipids or triglycerides, although they had favorable prognostic TRG1-2, displayed a worse prognosis, further highlighting the powerful role of lipid metabolites in DFS prediction ( Figure 7B ). Similar data for phospholipids and triglycerides were also obtained for OS ( Figure 13S ).

We evaluated a panel of 48 chemokines and cytokines in the patients’ plasma at the time of the response evaluation by a multiplex bead–based system. We applied sPLS-DA to analyze the results (17.4% of the total variance), again finding that the NR and R patients grouped into two distinct clusters with a model accuracy of 66.7% ( Figure 8A ). The loading plot showed that 9 out of the top 10 cytokines that were more statistically relevant for class separation, IL-6, CXCL9 (MIG), SCGF-β, IFN-α2, CXCL10 (IP-10), IL-12, IL-8, VEGF and MIP-1β, showed higher plasma levels in the NR relative to the R patients. Only the plasma levels of CTACK chemokine (CCL27) were higher in the R group Figure 8B .

Next, we performed ROC curves to determine the optimal cutoff value of these cytokines, reporting AUC values ranging between 0.471 and 0.773 ( Figure 14S ). Based on the obtained parameter cutoff values, univariate analysis showed that lower levels (< cutoff) of IL-6 (HR, 4.83; 95% CI, 1.68-13.83; P=0.002), SCGF-β (HR, 6.44; 95% CI, 1.83-22.63; P=0.034), and CXCL10 (HR, 7.98; 95% CI, 2.43-26.20; P=0.014) or higher levels (≥cutoff) of CTACK (HR, 0.22; 95% CI, 0.026-1.91; P =0.022) were significantly associated with good DFS ( Table 2 ; Figure 9A ). Similar results were found for OS ( Table 2S , Figure 15S ).

Again, we compared the DFS prediction potential of these four cytokines with TRG, demonstrating that only low IL6 levels were a clearly better predictor of DFS, independent of TRG status ( Figure 9B ). Similar data were obtained considering OS ( Figure 16S ).

Finally, taking advantage of all of the data accumulated on metabolomics, lipidomics and cytokinomics (as predictors of DFS) associated with the DFS of our cohort of patients, we analyzed all of the possible combinations of statistically significant variables that emerged to create a multiple biomarkers model using a support vector machine (SVM) algorithm. As shown in Figure 10 , we found that the best combination of circulating biomarkers to predict R (DFS ≥ 1 year) and NR (DFS < 1 year) patients in our cohort of metastases-resected cancer patients was represented by 3-hydroxybutyrate, cholesterol, phospholipids, triglycerides and IL-6, evaluated in plasma samples collected at the time of the response evaluation. In detail, ROC curve analysis for these combined biomarkers had an AUC value equal to 0.73 (95% CI: 0.083-0.0972) ( Figure 10A ). The combination of these five features was able to classify 15 R and 15 NR patients ( Figure 10B ) with a positive predictive value of 73% (probability of the correct identification of R) and a negative predictive value of 93% (probability of the correct identification of NR), overall correctly predicting the outcome of 83.3% (accuracy) of the patients.