Lorlatinib (>99.9%) was obtained from MedChem Express (United States). Methanol, HPLC-grade, was purchased obtained from Fisher Chemicals (Pittsburgh, PA, United States). Acetonitrile, HPLC-grade, was obtained from Merck (Darmstadt, Germany). Purified water was produced by Millipore’s ultrapure water system (Millipore, Bedford, MA, United States). All other chemicals and reagents were of analytical grade unless otherwise indicated.

All the animal-related experiments were conducted in accordance with guidelines of Institutional Experimental Animal Ethical Committee. SPF grade KM and ICR mice (weight: 18–20 g, age: 8 weeks) were obtained from the Beijing HFK Bioscience Co., Ltd. (License No. 11401300092657). All mice were given free access to normal diet and water during the experiment with an exception that mice were fasted for 12 h prior to drug administration. The experiment was conducted under standard breeding conditions with a temperature regime of 26°C day/18°C night, a relative humidity level of between 50 and 70 percent and a 12-h light/12-h dark photocycle. Mice weighing more than 21 g or less than 18 g were excluded from the analysis. Additionally, mice that suffered accidental injury and/or bleeding during the study were excluded from the analysis and finally, mice that died unexpectedly during the study were excluded from the analysis.

After 3 days of acclimatization, KM mice (weight: 18–20 g, age: 8 weeks) acquired for this study were weighed and randomly distributed into 2 groups: a lorlatinib group and a non-lorlatinib group. The mice in the non-lorlatinib group were orally administrated with physiological saline solution and the mice in the lorlatinib group were orally administered with 10 mg/kg lorlatinib (the concentration of lorlatinib solution: 1 mg/ml). Blood was collected from mice in both groups at 0.5, 1, 2, 4, 8, and 24 h after administration. Serum was exacted from the collected blood and stored at −80°C for further pretreatment and analysis.

Blood samples were collected from each mouse via orbital sinus at 0.5, 1, 2, 4, 8, and 24 h after lorlatinib administration and transferred to a non-heparinized tube. The blood was allowed to clot at room temperature before being centrifuged to separate serum, which was then stored at −80°C until further sample preparation.

Methanol (150 μL) with an internal standard, 2-chlorophenylalanine (20 mg/ml), was added to 50 μL serum samples in 1.5 ml centrifuge tubes followed by vortexing for more than 30 s. The mixture was centrifuged at 14,000 rpm for 10 min at 4°C. 120 μL of supernatant was collected from the centrifuged mixture and spin-dried in a centrifuge tube. Sixty μL of 75% methanol was used to re-dissolve the sample, which was then centrifugated at 12,000 rpm for 10 min to separate 15 μL of supernatant as the final sample that was analysed using mass spectrometry.

We have previously developed a rapid liquid chromatography-tandem mass spectrometry (LC-MS/MS) method for analysis of the concentration of lorlatinib in mouse serum (Chen et al., 2019). Methanol was used to precipitate protein in samples and lorlatinib was separated on a C18 column by gradient elution (0.1% of formic acid and methanol) and detected in the positive-ion mode with m/z 407.28 [M + H]⁺.

Waters Xevo G2-XS QTOF/MS (Waters, Manchester, United Kingdom) was utilised for chromatographic analysis. A reverse phase column, UPLC HSS T3 C18 (100 mm × 2.1 mm, 1.8 μm), was used for chromatographic separation with the column temperature set to 40°C. The detection wavelength was set at 275 nm. The optimal mobile phase consisted of ultrapure water with 0.1% formic acid as solvent A and acetonitrile with 0.1% formic acid as solvent B. The gradient conditions were as follows: 0–1 min, 95 to 95% A; 1–9 min, 95 to 60% A; 9–19 min, 60 to 10% A; 19–21 min, 10 to 0% A; 21–25 min, 100 to 100% B. The sample injection volume was 4 µL. To verify the accuracy and reproducibility, the sample run sequence was randomized and quality control (QC) samples were prepared and analyzed every 10 samples. All samples were maintained at 4°C during the experimental period.

For mass spectrometry profiling, Waters Xevo G2-XS QTOF/MS equipped with an electrospray ionization sources (ESI) (Waters Corporation, Manchester, United Kingdom), in which both positive and negative ESI was produced and detected. All mass scans were acquired under MSE mode (specifically, ESI Continuum mode). Mass detection was operated using the following setting parameters: drying gas (N2); flow rate, 800 L/h; gas temperature, 350°C; capillary voltage, 2.2 kV (ESI-) and 2.5 kV (ESI+); skimmer, 40 V; collision energy, 10–40 EV. Leucine enkephalin (m/z 556.2771 in ES+ and 554.2615 in ES-) was used as the external standard substance to perform online mass calibration for all the detection runs. Masslynk 4.1 software was used to collect data, with detected molecular weights ranging from 50 to 1,200 Da.

The mass data were preprocessed by MetaboAnalyst 3.0 (www.metaboanalyst.ca/) to generate a normalized data matrix. For multivariate analysis, the data matrix was introduced into SIMCA-P 14.1 software (Umetrics, Umea, Sweden). Unsupervised Principal Component Analysis (PCA) was employed to describe and identify the differences and relationships between samples. The supervised Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) model was constructed to mine for different metabolites. S-plots were created to confirm the result, as such, to avoid false positives: according to whether the variables were distributed in the neutral position, it could be determined whether there were significant alterations. OPLS-VIP parameter was applied to the metabolomics profiles of the experimental animal groups to achieve an improved certainty of the variables with the most significant contribution. Variables representing metabolites with a vip of more than 1, if the |p (corr)| ≥ 0.5, p value < 0.05 and folder change>2 or<0.5 at the same time, were considered as potential biomarkers. Molecules representing the potential biomarkers were identified by the online Human Metabolome database (https://hmdb.ca/) search engines based on the accurate mass data. The list of compound labels was uploaded to MetaboAnalyst 5.0 (http://www.metaboanalyst.ca/) and the pathway enrichment analyses were performed by the Pathway Analysis module to identify the most relevant pathways involved in the conditions of the study.

In our previous study, SD rats were randomly divided into groups (Chen et al., 2020). After cardiac perfusion with saline, the brain tissue of rats in the control group and in the lorlatinib administration group were taken for sequencing, which was completed at the BGI-Shenzhen. The library preparation included the following steps: mRNA isolation, RNA fragmentation, cDNA strand synthesis, ends reparation, A-tailing, adapter ligation, linker addition, PCR reaction and purification of products. The data obtained from sequencing, namely raw reads, was subjected to quality control (QC) to determine whether the sequencing data was suitable for subsequent analysis. After passing the quality control, the filtered clean reads were compared to the reference genome. On this basis, according to the statistical comparison rate and the distribution of reads on the reference sequence, it was judged whether the comparison result passed the second quality control (QC of alignment). Finally, gene quantitative analysis and various analyses of gene expression levels were carried out, these included: principal component, correlation, differential gene screening, etc., and the differentially expressed genes were subjected to GO function enrichment analysis and pathway enrichment analysis.

ICR mice were randomly divided into three groups, with 3 mice in each group. The animals were fasted overnight without water before administration, and lorlatinib was given at a dose of 10 mg/kg (0.1 ml/10 g) after fasting. Brain tissue samples were taken at 30 min, 2 h, and 4 h after administration. Finally, the expression levels of OPN and related key proteins was detected by Western blotting.

The brain tissue of mice was lysed with lysis buffer (Solarbio, whole protein extraction kit, cat. no. BC3710) to extract total protein. The protein concentration was determined by using a BCA assay (NCM biotech, BCA protein assay kit, cat. no. WB6501). After treatment with protein loading buffer, a 10% PAGE precast gel was used for protein electrophoresis. Subsequently, the protein was transferred to a PVDF membrane (Millipore, IPVH00010). After blocking with 7% fat-free milk at room temperature for 1 h, the membrane was incubated with primary antibody overnight at 4°C. The primary antibodies used in this study were OPN (1:1,000, abcam, cat. no. ab8448) and β-actin (1:1,000, Bioss, cat. no. bs-0061R). After washing 5 times with TBST containing 0.01% Tween-20 at room temperature (3 times for 5 min, 2 times for 10 min), the blot was incubated with goat anti-rabbit IgG horseradish peroxidase (1:10,000, ZSGB-BIO, cat. no. ZB-2301) at room temperature for 1 h. After washing again with TBST, the immunoblots was visualized with a chemiluminescence reagent (APPLYGEN, Super ECL Plus, cat. no. P1050), and the gray value of the immunoblots was semi-quantified using ImageJ software.

In previous studies, the blood and brain concentrations of lorlatinib in 48 mice were measured by Liquid Chromatography-Mass Spectrometry after administration (Chen et al., 2019). Based on this data, we further calculated the drug brain/blood distribution coefficient for each mouse. After determining the median, the mice were divided into high-coefficient level and low-coefficient level groups based on the comparison of cerebral blood distribution coefficient and the calculated median; these groups were represented by 1 and 0 respectively. Taking the abundance of metabolic markers as the independent variable, a neural network was constructed to predict the size of the blood-brain distribution coefficient. 70% of the data was selected randomly to be part of the training set and the remaining 30% data was used in the test data set.

The untargeted mass data collected by LC-IT-TOF/MS in positive and negative ion modes were analyzed using PCA to investigate the differences between the principal components of the control group and the lorlatinib group. PCA score scatter plots were illustrated in Figure 1A (ESI + mode) and Figure 1B (ESI- mode). The tightly grouped distribution characteristics of the quality control samples shown in both two figures indicated that the instrument was stable throughout the analytical process. Data generated on analysis of serum samples from the control group and the lorlatinib group gathered in distinct areas of the PCA score scatter plots, indicating substantial differences at the metabolite level between two groups.

To further investigate the potential differential metabolites between the two groups, the supervised Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) model was established in order to identify the relationship between metabolite expression level and sample group and to make predictions regarding the sample category. As shown in the OPLS-DA scores plot for data generated in the ESI + mode (Figure 2A) and the ESI- mode (Figure 2B), the two sample groups clustered in different areas of the figure, indicating that the model could predict the classification of the two samples groups. The evaluation parameters R^2Y and Q2 of the OPLS-DA model were 0.997 and 0.984, respectively, in the ESI + mode and 0.989 and 0.935, respectively, in the ESI- mode. With the R^2Y and Q2 being greater than 0.5, this suggested that not only did the model have a satisfactory interpretation rate of the matrices, but also that the model could fit and predict accurately.

An S-plot (Figure 2C and Figure 2D), as an implement for visualization and interpretation of OPLS discriminate analysis, was carried out to identify statistically significant metabolites based on their reliability and contributions to the model. The variables appearing at the top or bottom of the S-plot had a significant contribution to modeled class designation, while those appearing in the middle were considered to contribute less. Variables were classified according to their explanatory power. Predictors with a VIP of larger than 1 were the most relevant for explaining classification and were marked in red in the S-plot if, at the same time, the absolute values of their p (corr) were greater than or equal to 0.5.

Four-hundred and ninety-one (491) potential biomarkers were obtained for further analysis by refining the above result based on the additional filtering by the criteria of p < 0.05 and fold change (FC) > 2 or < 0.5. The accurate mass charge ratio of all potential biomarkers were entered into a search of the online Human Metabolome database (https://hmdb.ca/) for putative identification of biomarkers. After converting the biomarker from HMDB ID to KEGG ID, 360 biomarkers were enriched in the KEGG pathway and mapped to the metabolic pathways in the metabolomics data analysis platform, MetaboAnalyst 3.0. As shown in Figure 3, we identified 56 biomarkers related to metabolic pathways, of which the most relevant pathways were selected after comprehensive consideration of impact factors and raw_p; these were Sphingolipid metabolism, Glycerophospholipid metabolism, Thiamine metabolism and Synthesis and degradation of ketone bodies. These metabolic pathways hit 9 significant metabolites, namely: Acetoacetyl-CoA (S)-3-Hydroxy-3-methylglutaryl-CoA, Dihydroceramide, Sphingosine, L-Cysteine, Thiamin diphosphate, CDP-ethanolamine, Phosphatidylcholine and Choline, as depicted by the schematic diagram of the metabolic pathways related to lorlatinib (Figure 4).

In the preliminary experiment, we sequenced the RNA of the control group and the lorlatinib group mice (Chen et al., 2020). By using DEGseq algorithm, |log2Fold Change| ≥1 and Adjusted p value ≤0.001 as the screening criteria, 126 differentially expressed genes were obtained. Among them, there were 70 genes that were significantly up-regulated and 50 genes that were down-regulated (p < 0.01). Volcano plots (Figure 5A) were created to quickly identify meaningful changes from within a very large set of genes. According to the GO and KEGG annotation results and the official classifications, we classified the differential genes by function before using the phyper function in the R software package for enrichment analysis. These differentially expressed genes are involved in 23 biological processes, which mainly affect cellular processes, biological regulation, and multicellular organismal processes. The molecular functions of the identified genes mainly involve binding, catalytic activity, and signal transducer activity. The remaining 12 pathways were those of cellular components. Twenty-four pathways were significantly enriched, 20 of which are shown in Figure 5B, these included Neuroactive ligand-receptor interaction, RNA polymerase, Herpes simplex infection, Pyrimidine metabolism and Epstein-Barr virus infection.

The expression of OPN protein in the brain tissue of mice gradually decreased with increasing time after administration of lorlatinib. Claudin-5 protein levels did not change significantly within 4 h after lorlatinib administration, and vegf protein was up-regulated within 4 h after administration. Finally, TGF-b was significantly down-regulated after drug administration (Figure 6).

Combining metabolomics with transcriptomics, a previously undescribed Metabolite-Reaction-Enzyme-Gene interaction network was constructed by searching for correlations between genetic expression profiles and metabolite accumulation profiles. As shown in Figure 7, the Metabolite-To-Gene interaction network consisted of 13 metabolites which were identified in this study and 5 genes which had been revealed to be important in previous research (Chen et al., 2020). These networks also served to identify and validate a select number of genes and metabolites likely to contribute to combatting drug-resistant tumors and promoting blood-brain barrier permeability.

Mean serum concentration-time curves, upon which the pharmacokinetic parameters and the tissue distribution calculations were based, have been published previously (Chen et al., 2019). The plasma concentration curve shows two-compartment pharmacokinetic characteristics. The ratio of brain lorlatinib concentration to blood concentration in 48 samples was calculated, giving an average of 0.70 (standard deviation of 0.20) and a 90th and 10th percentile of 0.90 and 0.39, respectively. These findings indicated that there was significant individual variation in the distribution of lorlatinib in brain.

An artificial neural network (Figure 8A) was created with 9 inputs, one hidden layer, and one output layer. The hidden layer had 6 nodes. The output layer had 2 nodes since we needed to implement a binary classification of the blood-brain distribution coefficient, where there could only be a high-coefficient level or low-coefficient level. The hyperbolic tangent function, a nonlinear activation function that outputs values between −1.0 and 1.0, was used for connection between the input layer and the hidden layer. The sigmoid function, which can transform the range of combined inputs to a range between 0 and 1, was used as the Output layer activation function. This neural network architecture is more suitable for the nonlinear boundaries formed by complex metabolic processes. The classification table (Table 1) shows the practical results of using the neural network. In Figure 8B, we provide the importance of independent metabolic biomarkers as different measures of the extent to which the network’s model-predicted classification of brain-blood distribution coefficient is altered for different values of the independent metabolic biomarker. Normalized importance is simply the importance value divided by the importance values of (S)-3-Hydroxy-3-methylglutaryl-CoA, which had the largest importance value among all metabolic biomarker variables.