Standards of the methionine cycle-related metabolites (methionine, SAM, SAH and homocysteine) were obtained from Shanghai Macklin Biochemical Co., Ltd. The isotopic internal standard including Methionine-d3 (Met-d3), S-Adenosylmethionine-d3 (SAM-d3), S-Adenosylhomocysteine-d4 (SAH-d4), and Homocysteine-d4 (Hcy-d4) were purchased from MedChemExpress. LC-MS grade methyl alcohol and acetonitrile were obtained from MilliporeSigma. Ammonium formate and LC-MS grade formic acid were both obtained from Thermo Fisher Scientific, Inc.

A total of 146 patients with luminal breast cancer and 36 HCs were recruited for plasma sample collection at Jiangsu Cancer Hospital (Nanjing, China) between August 2023 and March 2024, including 94 (64.4%) patients with luminal A subtype and 52 (35.6%) patients with luminal B subtype. The inclusion criteria were as follows: i) Age >18 years; ii) histopathologically confirmed luminal A or luminal B breast cancer based on surgically resected or biopsy tissues; iii) no history of other concurrent malignant tumors or previous malignancies; and iv) patients with locally advanced or advanced luminal breast cancer who subsequently received first-line palbociclib-based therapy, as these patients were further included in the predictive model analysis for palbociclib treatment response. Breast cancer staging was performed according to established clinical staging criteria (18) and molecular subtypes were determined by professional pathologists based on molecular marker expression. Individuals that underwent physical examinations during the same period that had no clinical symptoms or abnormal findings upon examination were recruited as the HCs group. The present study was approved by the Ethics Committee of Jiangsu Cancer Hospital (Nanjing, China) and written informed consent was obtained from all participants.

For patients included in the palbociclib response prediction cohort, peripheral blood samples were collected prior to initiation of palbociclib-based therapy, at the time when patients were confirmed to receive treatment. Patients were subsequently followed longitudinally and classified according to clinical outcomes. Treatment response was evaluated based on the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. Progressive disease was defined as ≥20% increase in the sum of the diameters of target lesions relative to the nadir, with an absolute increase of at least 5 mm or the appearance of new lesions. Palbociclib resistance was operationally defined as RECIST v1.1-confirmed disease progression occurring ≤6 months after the initiation of palbociclib treatment, whereas patients who achieved complete response, partial response (PR) or durable stable disease with progression-free survival >6 months were considered treatment-responsive. Thus, the present study was designed as a predictive biomarker study based on pretreatment plasma samples rather than a cross-sectional analysis of established resistance.

Blood samples were collected into K2EDTA anticoagulant tubes, and plasma separation was performed <2 h after collection. Briefly, 5 ml of whole blood was gently inverted several times immediately after collection and centrifuged at 1,200 × g for 10 min at 4°C. The supernatant was collected as plasma and further centrifuged at 14,000 × g for 10 min at 4°C to remove residual cellular debris. The resulting plasma supernatant was aliquoted and stored at −80°C until further analysis.

In total, 200 µl of plasma was pipetted into a 1.5 ml microcentrifuge tube (EP tube), followed by the addition of 800 µl of precooled (−80°C) methanol containing an isotopically labeled internal standard (5.00 µM Met-d₃, 0.10 µM SAM-d₃, 0.05 µM SAH-d₄ and 5.00 µM HCY-d₄). The mixture was vigorously vortexed to ensure thorough homogenization and then incubated in a −80°C freezer for 4 h. After incubation, the samples were centrifuged at 14,000 × g for 10 min at 4°C. The resulting supernatant was transferred to a new EP tube and subjected to centrifugal concentration under 4°C. The concentrated sample was reconstituted with 50% methanol-water (methanol-water=1:1, v/v), thoroughly vortexed and centrifuged again at 14,000 × g for 10 min at 4°C. Finally, the metabolite-containing supernatant was transferred to a mass spectrometry vial for subsequent analysis.

Metabolite detection was performed using an LC-MS/MS system equipped with an Agilent 1260 Infinity LC system and an Agilent G6460 triple quadrupole mass spectrometer (Agilent Technologies, Inc.). An electrospray ionization (ESI) source was used for metabolite ionization and detection was conducted in positive ion mode. The mass spectrometry parameters were set as follows: Capillary voltage, 4,000 V; nebulizer pressure, 20 psi; drying gas flow rate, 11 l/min; and nitrogen gas temperature, 300°C. The multiple reaction monitoring (MRM) transitions were as follows: Methionine: 150.1>104.1/133.0; SAM: 399.1>136.2/250.2; SAH: 385.0>134.1/136.1; homocysteine: 136.1>56.2/90.1; methionine-d₃: 153.1>107.2/136.1; SAM-d₃: 402.0>136.2/250.1; SAH: 389.1>136.0/138.1; homocysteine: 140.0>60.1/94.0. A InfinityLab Poroshell 120 HILIC-Z column (2.1×100 mm; 2.7 µm) was used for metabolite separation with an injection volume of 5 µl. The column temperature was maintained at 40°C, and the flow rate was set to 0.4 ml/min. The aqueous phase (Phase A) consisted of a 5 mM ammonium formate aqueous solution containing 0.1% formic acid, while the organic phase was 5 mM ammonium formate in acetonitrile with 0.1% formic acid. The gradient elution program used was as follows: Phase A: 10% (0 min)>10% (1 min)>30% (4 min)>10% (4.1 min)>10% (5 min).

Stock solutions of metabolite standards were prepared using ultrapure water and stored for subsequent use. A series of standard solutions with gradient concentrations were prepared by serially diluting the metabolite stock solutions in a 5% bovine serum albumin (BSA) solution (used as a surrogate matrix). The concentration points of the calibration curve were set as follows: methionine: 1.00, 5.00, 10.0, 25.0 and 50.0 µM; SAM: 0.05, 0.10, 0.20, 0.50 and 1.00 µM; SAH: 0.01, 0.05, 0.10, 0.25 and 0.50 µM; homocysteine: 1.00, 5.00, 10.0, 25.0 and 50.0 µM. QC samples were prepared for the methionine cycle-related metabolites, including samples at the lower limit of quantification (LLOQ) concentration, low QC (3XLLOQ), middle QC (at the mid-concentration point of the calibration curve), and high QC (at 80% of the highest concentration point of the calibration curve).

Isotopically labeled analogs of the target metabolites were used as internal standards for quantitative analysis in the present study. A precise mass of each isotopic internal standard was accurately weighed, and then dissolved in ultrapure water to prepare 100 mM stock solutions. The internal standard stock solutions were serially diluted with water containing 0.1% formic acid to obtain working internal standard solutions, which were stored for subsequent use.

Acquisition and analysis of the raw mass spectrometry data were performed using the Agilent MassHunter Workstation Qualitative Analysis Software (version B.06.00; Agilent Technologies, Inc.). The mass spectrometry data was uploaded to Metabolomics Workbench (https://www.metabolomicsworkbench.org/; study_id: ST004922) (19).

Receiver operating characteristic (ROC) analysis of the combined metabolites was performed using multivariable logistic regression analysis. A total of five metabolites (methionine, SAM, SAH, homocysteine and SAM/SAH) were included as independent variables in the logistic regression model, with treatment response status as the dependent variable. The predicted probability generated from this multivariable logistic regression model was subsequently used as the combined metabolites score for ROC analysis. ROC analysis and the Delong test were performed with IBM SPSS Statistics version 27 (IBM Corp.).

ANCOVA was performed using SPSS (version 27.0; IBM Corp.). For each metabolite, concentration was entered as the dependent variable, disease status was entered as the fixed factor and age was entered as a covariate. The adjusted means and P-values for group differences after controlling for age were calculated. Statistical significance was set at P<0.05 (two-sided).

The Mann-Whitney U test was applied to compare metabolite level differences among three pairs of groups: Patients with luminal breast cancer vs. HCs, patients in the early-stage vs. patients in the late-stage and patients that are palbociclib-sensitive vs. patients that are palbociclib-resistant. Similarly, ROC analysis was used to evaluate the diagnostic efficacy of the independent/combined metabolites in these three comparisons. Least absolute shrinkage and selection operator (LASSO) regression was employed to screen for metabolites and clinical variables associated with palbociclib resistance. A logistic regression model was constructed for predicting palbociclib resistance, which was further visualized via a nomogram. The performance of the prediction model was assessed using ROC curves, calibration curves and decision curve analysis (DCA). Internal validation with bootstrap resampling (500 iterations) was performed to reduce the risk of overfitting. LASSO, logistic regression and bootstrap resampling were performed using RStudio (https://posit.co/products/open-source/rstudio).

The overall experimental workflow is illustrated in Fig. 1. First, patients with breast cancer and HCs were enrolled according to the inclusion criteria, with 94 patients with the luminal A subtype, 52 patients with the luminal B subtype and 36 HCs being recruited. An MS detection method for methionine cycle-related metabolites was optimized and established in advance. Subsequently, the metabolites were extracted from the plasma after collection and targeted MS detection was performed. Differential analysis of the metabolic landscape among groups was conducted, and the diagnostic efficacy of individual/integrated metabolites was analyzed. Subsequently, the clinical data of the patients was collected and integrated the metabolic landscape. LASSO was used to screen variables related to palbociclib resistance, and a logistic regression model was constructed and visualized with a nomogram. The performance of the prediction model was evaluated using calibration curves, ROC curves and DCA.

The clinical characteristics of the enrolled subjects are listed in Table I. The average ages of patients with luminal A/B breast cancer were 58 (31–83) years and 59 (37–82) years, respectively, and the average age of HCs was 51 (24–72) years. The differential analysis of the methionine cycle-related metabolites is presented in the subsequent sections.

Targeted metabolomics was employed for metabolite quantification (20,21). First, an MS detection method was established using metabolite standards. By optimizing parameters such as collision energy and fragment voltage, two MRM transitions were ultimately selected for each metabolite and its isotopic internal standard. The optimized mass spectrometry parameters are listed in Table SI. After validation, the methionine cycle-related metabolites exhibited excellent separation efficiency and mass spectrometric response within the 5-min gradient (Fig. 2A).

To achieve absolute quantification, calibration curves were constructed for the target metabolites. Specifically, calibration curves were constructed by plotting the ratio of the total MRM peak areas of target metabolites to those of their isotopic internal standards against metabolite concentrations. As shown in Fig. 2B, using a weighting factor of 1/x², all the calibration curves exhibited excellent linearity within the established concentration range (R² >0.99). Additionally, to verify the precision and accuracy of the quantification, a quality control (QC) assessment was performed for the method. The precision of the method, including intraday and interday precision, was quantified using the percent coefficient of variation (% CV). Method accuracy was evaluated by calculating the percentage bias (% bias), which reflects the discrepancy between the measured concentration of samples and their respective standard concentrations. For all QC samples, the resulting precision and accuracy values met the acceptance criteria of ≤15% (LLOQ ≤20%). The detailed results of the QC analysis are presented in Table SII.

To evaluate the potential clinical value of the methionine cycle, the present study first assessed the diagnostic efficacy of metabolites for luminal breast cancer. Given the significant difference in age between the luminal group and the HCs group (P=2.4×10⁻²), an analysis of covariance (ANCOVA) was performed to correct the impact of age. Compared with HCs, patients with luminal breast cancer had lower plasma concentrations of methionine (P=1.2×10⁻²), SAM (P=2.0×10⁻⁶), SAH (P=2.5×10⁻²) and homocysteine (P=2.0×10⁻²) (Fig. 3A). The SAM/SAH ratio, a common indicator of methylation level (22,23), did not differ (P=7.7×10⁻²; Table SIII). ROC analysis of individual metabolites revealed that SAM exhibited the best diagnostic efficacy (AUC=0.7578). Notably, the combined metabolites yielded improved diagnostic results when compared with any single metabolite alone (AUC=0.8130; Fig. 3B). To elucidate this difference, a DeLong test was conducted to compare the ROC curves of the combined metabolites and individual metabolites. The results revealed that the discriminatory performance of the combined metabolites was significantly improved when compared with that of the individual metabolites (Table SIV). Heatmap analysis also revealed a marked difference in metabolic landscape between patients with luminal breast cancer and HCs (Fig. S1).

To further investigate changes in the methionine cycle during the progression of luminal breast cancer, metabolite levels were compared between patients at the early-stage and late-stage. As shown in Fig. 4A, methionine (P<0.001), SAM (P<0.001), SAH (P<0.05) and the SAM/SAH (P<0.05) ratio were downregulated in patients at the late-stage, whereas homocysteine abundance did not vary with disease progression. ROC analysis revealed that SAM had the optimal performance in distinguishing between early-stage and late-stage patients with luminal breast cancer (AUC=0.8514), and the combined use of multiple metabolites resulted in an even greater diagnostic AUC of 0.9228 (Fig. 4B, Table SIV). Finally, to assess the value of the methionine cycle in palbociclib resistance, the present study compared the abundance of metabolites between patients who are palbociclib-sensitive and palbociclib-resistant. The abundances of SAM, SAH and homocysteine were lower in resistant patients, whereas methionine levels and the SAM/SAH ratio did not significantly differ between sensitive and resistant patients (Fig. 5A). ROC analysis demonstrated that the diagnostic efficacy of SAM (AUC=0.7029), SAH (AUC=0.7400) and homocysteine (AUC=0.7457) was favorable, and the combination of all metabolite indicators achieved an AUC of 0.9229 for distinguishing palbociclib-sensitive/resistant patients (Fig. 5B, Table SIV).

Given the limitations of current clinical methods for determining therapeutic resistance, the present study aimed to construct a palbociclib-resistance prediction model by combining the methionine cycle with clinical information. A total of 39 palbociclib-treated patients (25 sensitive vs. 14 resistant) were included in the model construction. After collecting clinical data, the correlations between the levels of methionine cycle-related metabolites and clinical characteristics were first analyzed. As shown in Fig. S2, except for a weak correlation (r=0.15, P<0.05) between methionine and age, no significant correlations were observed between the other metabolites and clinical indicators. ANCOVA further confirmed that there was no significant linear relationship between age and metabolites. These findings indirectly indicate that these metabolic indicators differ from traditional clinical indicators.

Subsequently, patients were divided into an internal training set and an internal validation set at a 7:3 ratio based on their resistance status. The training cohort included 18 sensitive and 11 resistant patients, while the validation cohort included 7 sensitive and 3 resistant patients. Here, LASSO regression was used to screen variables associated with resistance (Fig. 6A), and 10-fold cross-validation was employed for iterative analysis (Fig. 6B). The optimal model performance was achieved when the minimum criterion was applied (lambda.min=0.0424), with a total of 5 variables selected, including methionine, SAM, SAH, homocysteine and carbohydrate antigen 153 (CA153). On the basis of these variables, the present study developed a multivariate logistic regression model and further established a prediction model for the probability of palbociclib resistance. This prediction model was visualized as a nomogram (Fig. 6C), which calculates the probability of palbociclib resistance based on the total score corresponding to each variable. To reduce the risk of overfitting, the present study further performed internal validation using bootstrap resampling (500 iterations) for correction (Fig. S3).

Multiple metrics were employed to validate the model for accuracy and reliability. ROC analysis revealed an apparent AUC of 0.9795 (95% CI, 0.9409–1.0000) in the training set (Fig. S4). The mean optimism was 0.0724, yielding an optimism-corrected AUC of 0.9089, indicating that the model maintained good discrimination after correction. The internal validation set yielded an AUC of 0.8000, but its confidence interval was unstable because of the small sample size. The Hosmer-Lemeshow test (training set: P=7.64×10⁻¹; validation set: P=3.57×10⁻¹) confirmed that the correction was acceptable. Calibration curve analysis further confirmed the favorable agreement between observed and predicted results (Fig. S5). DCA demonstrated the model's clinical benefits (Fig. S6). In summary, all these results support the potential clinical utility of the prediction model.