Clinicopathological characteristics of patient cohort
Clinicopathological characteristics of patients are shown in Table 1. In this study cohort, tissue samples from 10 non-recurrent and 10 recurrent cases were analyzed. Among the characteristics of these two groups, differences in pathological stage (P = 0.033), lymph node metastasis (P = 0.033), and blood vessel invasion (P = 0.005) were statistically significant. The 1- and 2-year RFS rate of the recurrent group was 50% and 20% with median RFS time of 12.5 (range, 9–38) months, respectively (Additional file 1, Supplemental Fig. 1). The median follow-up time of the non-recurrent and recurrent groups was 68.5 (range, 60–77) and 42.5 (range, 21–60) months, respectively.
Table 1. Clinicopathological characteristics of the non-recurrent and recurrent groups.
Characteristics
|
Non-recurrent (n = 10)
|
Recurrent (n = 10)
|
P-value
|
Median age (range)
|
67.5 (49-75)
|
71.5 (67-89)
|
0.069
|
Gender (male/female)
|
7/3
|
7/3
|
1.000
|
Smoking history (+/-)
|
7/3
|
6/4
|
1.000
|
Pathological stage (I/II)
|
10/0
|
5/5
|
0.033
|
Median tumor size (mm) (range)
|
23.5 (9-29)
|
24 (9-37)
|
0.649
|
Adenocarcinoma subtype
|
|
|
0.293
|
Lepidic
|
2
|
0
|
|
Papillary
|
5
|
7
|
|
Acinar
|
1
|
2
|
|
Solid
|
2
|
1
|
|
Lymph node metastasis (+/-)
|
0/10
|
5/5
|
0.033
|
Pleural invasion (+/-)
|
2/8
|
5/5
|
0.350
|
Lymphatic vessel invasion (+/-)
|
1/9
|
6/4
|
0.057
|
Blood vessel invasion (+/-)
|
2/8
|
9/1
|
0.005
|
Micropapillary component (+/-)
|
5/5
|
8/2
|
0.350
|
Spread through air space (+/-)
|
2/8
|
4/6
|
0.628
|
Driver gene mutation
|
|
|
|
EGFR (+/-)
|
2/8
|
6/4
|
0.170
|
ALK (+/-)
|
0/8
|
0/7
|
-
|
Surgical procedure
|
|
|
1.000
|
Lobectomy
|
10
|
9
|
|
Wedge resection
|
0
|
1
|
|
Adjuvant chemotherapy
|
|
|
1.000
|
Indication (Stage IA3-IIB)
|
8
|
9
|
|
Received
|
4
|
4
|
|
Recurrent style
|
|
|
-
|
Locoregional
|
-
|
3
|
|
Distant
|
-
|
8
|
|
Abbreviations: ALK, anaplastic lymphoma kinase; EGFR, epithelial growth factor receptor.
Trend analysis between the non-recurrent and recurrent groups
The frozen tissue samples were subjected to LC–MS/MS, and the total lipid level of cases was calculated by accumulating normalized intensities of lipids. Notably, the average total lipid level of the recurrent group was 1.65 times higher than that of the non-recurrent group (P = 0.026) (Fig. 1). A total of 2,595 lipid species were identified and quantified by analyzing the mass spectral data using a LipidSearch™ software (the full list of identified 2,595 lipid species is presented as Additional file 2), which were also subjected to PCA. The PCA plot did not show clear separation between the recurrent and non-recurrent groups; however, the recurrent group exhibited partial separations between the first three principal components (Additional file 1, Supplemental Fig. 2). These results suggested differences of lipidome between the recurrent and non-recurrent groups, which urged us to screen lipids to distinguish the two groups.
Screening of candidate lipids for recurrence prediction
To screen lipids with different levels between the two groups, volcano plots of the identified lipids were described first, and lipidomes between the non-recurrent and recurrent groups were compared (Fig. 2). The volcano plot identified 207 lipid species, with relative amounts significantly different between the two groups (folding change, ≥2.0 or ≤0.5; P-values, <0.05). The number of lipids that increased and decreased in the recurrent group was 203 and 4, respectively. These increased or decreased lipid species consisted of various head groups (Additional file 2, increased lipid species; shown in red, decreased lipid species; shown in green). Then, based on prominent distributions of the volcano plot, we narrowed the 203 candidate lipids increased in the recurrent group to the following 9 molecules (Fig. 2, blue arrows pointing to red plots): biotinyl-phosphoethanolamine (BiotinylPE)(30:3), ceramide (Cer)(d42:0), sphingomyelin (SM)(d35:1), Cer(d18:0_24:0), PC (41:2), monoether phosphatidylcholine (MePC)(34:6e), cholinesterase (ChE)(24:1), MePC (40:8e), and ChE(20:1). As for the lipids that decreased in the recurrent group, the following four molecules were annotated (Fig. 2, blue arrows pointing to green plots): monohexosylceramide (Hex1Cer)(t42:1+O), triglyceride (TG)(15:0_14:0_14:0), PC(18:2_18:2), and lysophosphatidylcholine (LPC)(12:0).
The relative amounts of these lipid species were evaluated with their distributions by comparing the two groups (Fig. 3A and B). In all tested lipids, distributions between the two groups were well separated enough to establish the cut-off values, whereas only few marked outliers were found.
We next calculated the cut-off values and AUC of these 13 lipids to evaluate their discrimination ability for disease recurrence, and the following final candidates with top three AUC were selected: SM(d35:1), 0.90; Cer(d42:0), 0.90; and TG(15:0_14:0_14:0), 0.90 (Table 2) (respective lipid species can be found in Additional file 2 with the following identical numbers: SM(d35:1), 2201; Cer(d42:0), 122; and TG(15:0_14:0_14:0), 2354). These three final lipid candidates were annotated as the following ions [SM(d35:1)+H]+, [Cer(d42:0)+HCOO]-, and [TG(15:0_14:0_14:0)+NH4]+ in the LipidSearch™ software (Additional file 2).
Table 2. AUC rank of candidate lipid predictors determined by ROC curve.
Rank
|
Species
|
Cutoff value
|
AUC (95% CI)
|
1
|
SM(d35:1)
|
1866710.893
|
0.91 (0.773 - 1.000)
|
2
|
Cer(d42:0)
|
127504.392
|
0.90 (0.769 - 1.000)
|
3
|
TG(15:0_14:0_14:0)
|
3788045.717
|
0.90 (0.766 - 1.000)
|
4
|
Cer(d18:0_24:0)
|
521665.875
|
0.85 (0.673 - 1.000)
|
5
|
PC(18:2_18:2)
|
81938569.45
|
0.84 (0.654 - 1.000)
|
6
|
ChE(24:1)
|
52345.314
|
0.83 (0.650 - 1.000)
|
7
|
PC(41:2)
|
33392.237
|
0.83 (0.645 - 1.000)
|
8
|
BiotinylPE(30:3)
|
6185556.894
|
0.83 (0.602 - 1.000)
|
9
|
LPC(12:0)
|
379006.021
|
0.79 (0.577 - 1.000)
|
10
|
Hex1Cer(t42:1+O)
|
854682.452
|
0.79 (0.562 - 1.000)
|
11
|
MePC(40:8e)
|
7939029.972
|
0.78 (0.531 - 1.000)
|
12
|
ChE(20:1)
|
66948.94
|
0.77 (0.549 - 0.991)
|
13
|
MePC(34:6e)
|
1029943.584
|
0.77 (0.536 - 1.000)
|
Lipids with top three AUC were selected as final candidate predictors (boldfaced notations).
Abbreviations: AUC, area under the ROC curve; CI; confidential interval; ROC, receiver operating characteristic.
MS/MS for [SM(d35:1)+H]+, [Cer(d42:0)+HCOO]-, and [TG(15:0_14:0_14:0)+NH4]+ demonstrated product ion peaks corresponding to phosphocholine, several fragments compatible with fragmentation of Cer(d42:0) with concomitant oxidation reaction, two fragments produced by neutral loss of fatty acid (FA)(14:0) or FA(15:0) from TG(15:0_14:0_14:0), respectively (Additional file 1, Supplemental Fig. 3). Consequently, the annotations of the final candidates by LipidSearch™ software were consistent with the results of MS/MS.
Among these three candidate predictors, SM(d35:1) was found to be positively correlated with Cer(d42:0) (Spearman’s rank correlation coefficient [rS] = 0.621, P = 0.004), TG(15:0_14:0_14:0) was inversely correlated with SM(d35:1) (rS = −0.553, P = 0.013), and TG(15:0_14:0_14:0) was weakly inversely correlated with Cer(d42:0) (rS = −0.353, P = 0.127) (Additional file 1, Supplemental Fig. 4).
Validation of recurrence prediction ability among the final lipid candidates
Table 3 shows the sensitivity, specificity, and accuracy of the final candidate lipid predictors compared with the conventional pathological prognostic factors, lymph node metastasis, and blood vessel invasion, which were identified as significant recurrent factors in this cohort. Sensitivity of all three candidate lipid predictors is superior to that of lymph node metastasis. Patients with lymph node metastasis (all of them were hilar or lobar lymph node metastasis) corresponded to those in stage II. Among the recurrent group in this study cohort, half of the study population had stage I, whereas the other half had stage II. As lymph node metastasis can be detected among stage II cases, the sensitivity of lymph node metastasis was consequently lower than those of three candidate lipid predictors, which detected both stage I and stage II. Hence, these three predictors were superior to lymph node metastasis for patient screening. When comparing the candidate lipid predictors and blood vessel invasion, only SM(d35:1) showed prediction abilities higher or equal to those of blood invasion in all validation points. Therefore, we propose SM(d35:1) as the most hopeful candidate for recurrence prediction.
Table 3. Comparison of sensitivity, specificity, and accuracy among the three final candidate predictors and conventional histopathological prognostic factors.
Predictors for recurrence
|
Sensitivity
|
Specificity
|
Accuracy
|
Candidate lipid predictors
|
|
|
|
SM(d35:1)
|
1.00
|
0.80
|
0.90
|
Cer(d42:0)
|
0.90
|
0.70
|
0.80
|
TG(15:0_14:0_14:0)
|
1.00
|
0.70
|
0.85
|
Pathological prognostic factors
|
|
|
|
Lymph node metastasis
|
0.50
|
1.00
|
0.75
|
Blood vessel invasion
|
0.90
|
0.80
|
0.85
|
SM(d35:1) showed the most excellent prediction ability.
Abbreviations: Cer, ceramide; SM, sphingomyelin; TG, triglyceride.