A new rapid diagnostic system with ambient mass spectrometry and machine learning for colorectal liver metastasis

This is a prospectively planned study using retrospectively obtained tissues. Patients who underwent surgical resection of CRLM at The University of Tokyo Hospital during February 2014 and October 2017 were potential candidates for this study. The study comprises two mass spectrometry experiments using resected specimens. First, we investigated the diagnostic accuracy of PESI-MS and machine learning for CRLM. Next, we used LC-ESI-MS to examine the key molecules that distinguish cancer from non-cancer. This study was approved by the Institutional Ethics Committee of The University of Tokyo, and written informed consent was obtained from all participants.

Of all candidates, patients whose maximum tumor diameter exceeded 5 × 5 × 5 mm were included in the further analysis. We obtained a block of approximately 5 × 5 × 5 mm from surgically resected CRLM and an equivalent block of non-cancer liver parenchyma. If multiple CRLMs were resected, the largest nodule was chosen. If the remaining specimen size following this sample procedure precluded accurate histological analysis, the patients were excluded from this study. Patients with specimens that showed gross necrosis were also excluded. Analysis of liver parenchyma was waived in patients with impaired liver function––Child–Pugh class B and indocyanine green retention rate at 15 min (ICGR15) of > 20.0%––and only the tumor specimen was analyzed. Obtained specimens were immediately frozen in liquid nitrogen and stored at − 80 °C until analysis. LC-ESI-MS was applied for patients in whom both the tumor and the liver parenchyma were analyzed using PESI-MS.

The assessment of the diagnostic accuracy of the logistic regression-based diagnostic algorithm consisted of two steps. First, the mass spectra of each cancerous or non-cancerous tissue were acquired by PESI-MS. Second, all mass spectra were learned by logistic regression to discriminate the blind samples. Furthermore, the discriminant accuracy of diagnostic algorithm was validated using 20 independent specimens that were obtained during surgery undertaken in 2018.

PESI-MS measurements are described in more detail in a previous report [13]. Briefly, for sample preparation before PESI-MS, 2.5 mg of tissue was homogenized in 100 μl of 50% ethanol using a disposal pestle (Argos Technologies, Vernon Hills, IL, USA). The homogenate was centrifuged at 15,000×g for 5 min, and the supernatant was diluted by 50% ethanol to 4-fold for positive ion mode and 2-fold for negative ion mode. Nine microliters of sample solution were placed in the sample plate (Shimadzu Corp., Kyoto, Japan) to perform PESI-MS.

Ambient ionization unit (DPiMS-8060; Shimadzu Corp.) was used for PESI combined with a triple quadrupole mass spectrometer (LCMS-8060; Shimadzu Corp.) for direct MS, and the analyses were performed as previously described [12]. Analyses were performed for both positive and negative ion mode. The probe needle with a tip radius of < 1 μm was moved downward to touch the sample solution and then upward to apply high voltage (2.3 kV for positive and − 2.0 kV for negative ion mode) for ESI. This movement was repeated, and generated ions were introduced into the mass spectrometer. Figure 1a shows the procedure of PESI-MS analysis. Acquisition of scanning data was completed within approximately 10 min after beginning the sample preparation. Representative mass spectra of each sample were generated using LabSolutions (Shimadzu Corp.). The abscissa indicates mass-to-charge ratio (m/z), and the ordinate shows ion intensity (label-free quantification).

Logistic regression analysis was applied to each mass spectrum pattern and corresponding tissue type, that is, CRLM or non-cancerous liver. The expression levels of analyzed lipids obtained from each sample were individually normalized by the median value. Normalized datasets of CRLM and normal liver parenchyma were learned by logistic regression, a type of machine learning, and blinded samples were classified as cancer or not. The possibility of cancer was indicated as the value of probability (0.0–1.0). The detailed procedure, mathematical formula of data processing and discriminant analysis were previously described [14]. To evaluate the discriminative accuracy of the algorithm, leave-one-out cross validation (LOOCV) and 10-fold cross validation (10-fold CV) were applied. LOOCV procedure is summarized as follows. One sample is left out of all samples; this one is considered as the validation set and the remaining samples are assumed as the training set. This cycle is repeated until all samples enter the test set [15]. The 10-fold CV procedure was as follows. Original samples were randomly separated into 10 equally sized subgroups. Of the 10 subgroups, a single subgroup was retained as the blind data for testing the model, and the remaining nine subgroups were used as training data. The procedure was repeated for all subgroups, with each of the 10 subgroups used exactly once as the blind data. The 10 results were then averaged (or otherwise combined) to produce a single estimation. The 10-fold CV for randomly constructed subgroups was performed 10 times. Additionally, the discriminant accuracy of the diagnostic algorithm was further validated using specimens from additional patients with 40 data sets (20 CRLMs and 20 non-cancerous tissue).

In LC-ESI-MS, high-pressure liquid chromatography (Nexera X2; Shimadzu Corp.) and ESI unit were installed to LCMS-8060 (Fig. 1b). For analysis of tissue components, LC/MS/MS Method Package for Phospholipid Profiling (Shimadzu Corp.) was applied in accordance with the manufacturer’s instructions. A Kinetex C8 column (Kinetex C8, 150 mm × 2.1 mm i.d., 3.6-μm particle size; Phenomenex, Torrance, CA, USA), mobile phase A (20 mM ammonium formate in water) and mobile phase B (acetonitrile: isopropanol 1:1 v/v) were used for LC separation. The concentration of mobile phase B was programmed as 20% (0 min) – 20.0% (1 min) – 40.0% (2 min) – 92.5% (25 min). The oven temperature was 45 °C. Data processing and molecular identification/quantification were performed automatically using LabSolutions software (ver. 5.82 SP1; Shimadzu Corp., Kyoto, Japan).

PESI-MS presents mass spectra associated with molecules of up to 2000 m/z, which include most phospholipids. A total of 457 ion intensities of phospholipids with definite numbers of carbon atoms and double-bonded acyl groups were analyzed.

A flow chart of patient selection and sample selection is shown in Fig. 2. Among 202 patients who underwent surgical resection of CRLMs, 99 were excluded because they met either of the following criteria: (1) a tumor smaller than 5 × 5 × 5 mm or a small tumor volume that did not allow qualitative histological examination after sampling for MS (n = 74), or (2) a necrotic area that occupied most of the tumor (n = 25). Therefore, we made 103 CRLM and non-cancer background liver specimens as candidates of this study. Of these 103 non-cancer liver specimens, 23 were excluded from the analysis because they were obtained from patients with impaired liver function (Child–Pugh class B or ICGR15 of > 20.0%). Finally, PESI-MS was performed on 103 CRLM specimens and 80 non-cancer liver specimens. Next, the matched CRLM and non-cancerous samples from each patient were selected to apply LC-ESI-MS. Hence, 75 CRLM and normal liver parenchyma could be re-analyzed. The demographics and clinicopathological features of the 103 patients with CRLM are summarized in Table 1. Forty-four patients (43%) underwent neoadjuvant chemotherapy.

Normalized mean mass spectra of CRLM and non-cancer liver parenchymal tissue are shown in Fig. 3. The mass spectral pattern of each category differed in positive and negative ion modes. These spectral patterns were learned and analyzed using logistic regression and validated using LOOCV method. Among 103 CRLM samples, 102 samples were correctly diagnosed. Meanwhile, all 80 non-cancer samples were correctly diagnosed. Specificity, sensitivity and accuracy of logistic regression were 100, 99 and 99%, respectively. Ten-fold CV demonstrated that the total accuracy rate was 99.0%. (Supplementary Table 1). ROC curve of logistic regression for discriminating CRLM is shown in Fig. 4, in which the area under the curve was 0.9999.

In the independent validation set, 18 CRLMs were correctly diagnosed, while all 20 non-cancerous samples were correctly diagnosed. Specificity and sensitivity were 100.0 and 90.0% respectively (Supplementary Table 2).

We identified 457 species of phospholipids, among which 146 ion (31.9%) species showed significantly different ion intensities between CRLM and non-cancer liver parenchyma (P < 0.01). Table 2 shows top 10 phospholipids among these 146 species. The total ion intensity of each phospholipid, including phosphatidylcholine, phosphatidylethanolamine, phosphatidylserine, phosphatidylinositol, sphingomyelin, lysophosphatidylcholine, lysophosphatidylethanolamine, lysophosphatidylinositol and lysophosphatidylglycerol, was compared between CRLM and non-cancer liver parenchyma (Fig. 5). Notably, the expression of phosphatidylcholine and phosphatidylethanolamine in the non-cancer liver parenchyma (mean: 2.32 × 10⁸ arbitrary unit [AU] and 6.19 × 10⁷ AU, respectively) was higher than CRLM (mean: 1.46 × 10⁸ AU and 1.11 × 10⁷ AU; P < 0.01, respectively).

As shown in Table 2, phospholipid species that are upregulated in non-cancer liver tissues are more likely to contain fatty acids (no double bonds exist in the carbon chain) and polyunsaturated fatty acids (more than two double bonds exist in the carbon chain), while those increased in CRLM are predominantly monounsaturated fatty acids (one double bond exists in the carbon chain). The total ion intensities of saturated, monounsaturated, and polyunsaturated fatty acids were calculated in CRLM tissue and non-cancer liver parenchyma. Monounsaturated fatty acids accounted for 37% of total phospholipids in CRLM and 11% of those in non-cancer liver specimens (P < 0.01). Conversely, the total ion intensity of the other two saturated fatty acids was higher in non-cancer liver tissue than in CRLM (Table 3).