*2.4. UHPLC-QTOF/MS-Based Non-Targeted Metabolomics Analysis*

Non-targeted metabolomics analysis was performed on an Infinity 1290 UHPLC system (Agilent Technologies, Santa Clara, CA, USA) coupled to an Agilent 6545 QTOF mass spectrometer (Agilent Technologies, Santa Clara, CA, USA). Chromatographic separation was carried out on a Zorbax Eclipse Plus C18 column (3.0 × 150 mm, 1.8 μm, Agilent Technologies, Santa Clara, CA, USA) at 40 ◦C. Water with 5 mmol/L ammonium acetate and methanol with 5 mmol/L ammonium acetate were used for mobile phases A and B, respectively, which were kept at a flow rate of 0.40 mL/min with a gradient elution profile. The proportion of solvent B was linearly applied as follows: 0–5 min, 0–12%; 5–15 min, 12–35%; 15–18 min, 35–45%; 18–26 min, 45–75%; 26–33 min, 75–95%; and 33–35 min, 95–5%. The injection volume was 3 μL. The post-time between each two consecutive injections was 3 min. Dual jet stream electrospray ionization (ESI) was performed under negative ionization mode. The parameters were as follows: nebulizer pressure: 35 psi; capillary voltage: 3.5 kV; gas flow rate: 8 L/min; fragmentator voltage: 130 V; sheath gas temperature: 350 ◦C; and sheath gas flow rate: 8 L/min. The TOF scan was set at an *m/z* of 100–1100 with an acquisition rate of 2 spectra per second. The auto MS/MS model was applied for compound identification with fixed collision energies (10 V, 20 V and 40 V). Reference ions with *m/z* 112.9856 and 1033.9881 were utilized for real-time mass calibration during both the TOF scan and the auto MS/MS scan.

The obtained raw data were imported into Masshunter Qualitative Analysis software (B.07.00 SP1, Agilent Technologies, Santa Clara, CA, USA) to extract all feature ions, then exported as .cef documents. These .cef documents were imported into MPP (Mass Profiler Professional software package, version B.14.5, Agilent Technologies, Santa Clara, CA, USA) for retention time correction using internal standards and subsequent peak alignment within the specified retention time window (±2.5%). The entities with an occurrence frequency >60% and a coefficient of variability (CV) < 25% were retained. After Pareto scaling and logarithmic transformation of the quantitative data, principal component analysis (PCA) was performed using SIMCA 14.1 (Umetrics AB, Umeå, Sweden) to provide an intuitionistic demonstration of an overall clustering pattern of the bee pollen samples.

Differential entities (*p* < 0.05 in analysis of variance, ANOVA) among the bee pollen samples from different botanical origins were identified using SPSS 20.0 (Chicago, IL, USA) for further analysis. Metabolite identification was performed by searching for exact mass and MS/MS spectra in the Metlin database (http://metlin.scripps.edu, accessed on 10 January 2021) and Human Metabolome Database (HMDB, https://hmdb.ca/, accessed on 10 January, 2021). The retention time and MS/MS spectra of putatively identified compounds were validated by authentic standards analyzed under the same conditions. To show the abundance differences in identified compounds among these bee pollen samples, heatmap visualization was carried out using MetaboAnalyst 4.0 [28] with Pareto scaling and logarithmic transformation. To improve the classification of CBP and non-CBP samples, orthogonal projections to latent structures discriminant analysis (OPLS-DA) was conducted in SIMCA 14.1. The OPLS-DA model was cross-validated by permutation tests with 200 iterations. The values of variable importance in projection (VIP) were used to rank the overall contribution of each compound to the OPLS-DA model. Compounds with VIP > 1.0, *p* < 0.05 according to Student's *t*-test and fold change (FC) > 1.5 were regarded as discriminating compounds driving the observed group separation.
