Baseline characteristics
The brief study designs are shown in Fig.1, A-B. The study first recruited 360 patients (cohort 1, 144 men and 216 women, median age: 55.5 years; age range: 17–87 years), including 120 HCs, 120 patients with UCA, and 120 age- and sex-matched patients with RCA. The baseline characteristics of the human subjects in the three groups, based on their diagnostic status, are shown in Extended Data 1 Table S1. There were no significant differences in sex, age, hypertension, diabetes mellitus, hyperlipemia, coronary heart disease, smoking, alcohol consumption, and body mass index (BMI) among the three groups of participants. The duration of plasma storage at measurement did not differ among the three groups. There were 147 aneurysms in the 120 patients with UCA and 145 aneurysms in the 120 patients with RCA.
Next, we enrolled 72 men and 108 women for trend testing of the primary results from the 180 patients (cohort 2), including 60 HCs, 60 patients with UCA, and 60 age- and sex-matched patients with RCA. Baseline characteristics of the human subjects in the three groups are shown in Extended Data 2 Table S1.
Overview of the distribution of lipid species and subclass intensities in the three groups
To enable comprehensive plasma lipidomic profiling of CAs, lipidomic analysis was performed with an untargeted LC-MS method using a CSH C18 column and using the same LC-MS with consistent quality control in a total of 360 participants from one center and 68 QC samples (Extended Data 1 Fig. S1).
After QC and support vector regression (SVR), LC-MS detected 1312 lipids (972 ESI+ and 340 ESI-) covering 8 lipid categories and 29 subclasses in all three groups, in which TG was the most abundant lipid in all three groups, followed by phosphatidylcholine (PC), sphingomyelin (SM), phosphatidylethano- lamine (PE), and ceramides (Cer) (Fig.2 A). In the UCA and RCA samples, the numbers of identified lipids were the same as those identified in the HC samples after case-by-case review. We then performed PCA to analyze the lipid data set and identify the characteristics of each group. QC samples, shown as calamus ellipses, were center clustered, which indicated good reproducibility of the instruments and stability during the lipidomics study (Fig. 2B).
The lipid intensity of every sample accompanied by sex and age is shown in Fig. 2 C. The total lipid intensity comparisons of the three groups, with and without sex and/or age data, is shown in Fig. 2 D-I. The trend of total lipid intensity decreased from HC to UCA to RCA, and a significant difference among the three groups was observed (Fig. 2 D, ANOVA test, p < 0.001). The same decreased trend was also observed in the F (female) or M (male) subgroups (Fig. 2 G, ANOVA test, p < 0.001) and age subgroups (Fig. 2 I, ANOVA test, p < 0.05) among the three groups. There were no significant differences between M and F in each of the three groups (Fig. 2 E), among the four age subgroups in each of the three groups (Fig. 2 F), and among the four age subgroups in the F or M subgroups (Fig. 2 H). In addition, the same decreased trend was observed in 26 of the 29 subclasses among the three groups (Fig. 2 J, ANOVA test, p < 0.05).
The same trend was well tested in cohort 2 (Extended Data 2 and Extended Data 2 Fig. S2.)
Plasma lipidomic profiling of HCs, UCAs, and RCAs
To investigate the lipidomic changes associated with CAs, three paired comparisons were performed. The differential lipids that satisfied the criterion of variable importance in the projection (VIP) of > 1.0 and p value < 0.05 were considered as potential differential lipids. An Orthogonal partial least squares discriminant analysis (OPLS-DA) model was employed to further investigate lipid changes and differential lipids. OPLS-DA score plots revealed that all three groups could be discriminated (Fig. 3 A-C). Parameters for the explained variation (R2), an indicator of model robustness, and the cross-validated predictive ability (Q2) were obtained, as shown in Fig. 3A-C. The heatmap depicts the relative abundance of all lipids in all three groups (Fig. 3 D). Among these three paired comparisons, the lipid profiles of RCA were better distinguished from those of the HC group (Fig. 3 D).
As summarized in Extended Data 1 Table S2-3, 75 and 130 differential lipids were identified from the comparisons of UCA vs. HC and RCA vs. HC participants, respectively. Compared with the HC group, 5.5% (75) of the identified lipids were significantly different in the UCA group (p value < 0.05; VIP > 1.0), while, 9.9% (130) of identified lipids were significantly altered in the RCA group (p value < 0.05; VIP > 1.0), and the number of altered lipids were greater in the RCA group than in the UCA group (c2 test, p < 0.05). From this trend, we found that the number of altered lipids correlated positively with the severity of the clinical CA status (UCA and RCA). We hypothesized that increased lipid alteration in patients with a CA resulted in increased likelihood of CA rupture.
Interestingly, compared with HCs, except for 10 lipids in UCA group, the other altered lipids were under-represented, with lower levels in the UCA and RCA groups (Extended Data 1 Table S2-3). This indicated that most altered lipids were decreased in the plasma of the CA group, and the degree of the decreased lipids was much lower in the RCA group than that in UCA the group. Using a combination of our results and published research (13,14), we hypothesized that the decreased lipids in plasma probably accumulated in the CA and normal artery wall, which resulted in the formation, development, and rupture of CAs. This trend was also tested in cohort 2 (Extended Data 2 and Extended Data 2 Fig. S3).
In addition, as shown in Extended Data 1 Tables S2-5, there were 35 differential lipids including 4 lipid subclasses (1 LPC, 4 SMs, 7 PCs, and 23 TGs) that were significantly altered among the three groups (Fig. 3E), which could be used as potential biomarkers to diagnose or to discriminate CAs from HCs. The relative levels of these 35 differential lipid profiles were presented as a heatmap and a differential lipid profile was observed when comparing UCA to HC, RCA to HC, and RCA to UCA (Fig. 3 F). The lipids exhibited the same decreasing trends from HC to UCA to RCA, and a significant difference was observed among the three groups and between each of the two groups (Fig. 3 G, p < 0.001).
TGs were the distinct profile of lipids in patients with UCA and RCA as compared with HCs
The large number of samples in our cohort allowed us to investigate the association between lipids shifts and CA development according to clinical classification (UCA and RCA). As shown in Extended Data 1 Table S2 and Fig. 4 A, 75 differential lipids were identified from the comparisons of UCA vs HC, of which, 65 were downregulated and 10 were upregulated. 13 out of 75 altered lipids were significantly decreased (Fold change (FC) < 0.75 or log2FC < -0.42) and no lipids were significantly increased (FC > 1.5 or log2FC > 0.59) in patients with UCA compared with HCs Extended Data 1 Table S2 and Fig. 4 A). Of the 13 significantly decreased lipids, all were TGs. Furthermore, the 9 significantly decreased lipids including TG(56:2)+NH4, TG(54:1)+NH4, TG(52:0)+NH4, TG(54:2)+NH4, TG(56:4)+NH4, TG(56:5)+NH4, TG(53:2)+NH4, TG(53:3)+NH4 and TG(54:3)+NH4 were further decreased in patients with RCA as compared with HCs.
In patients with RCAs, all the differential lipids were decreased as compared with HCs (Extended Data 1 Table S3 and Fig.4B). 75 out of 130 altered lipids were significantly decreased (FC< 0.75 or log2FC < -0.42), and 24 lipids showed a less than 0.5-fold decrease (log2FC < -1). Of the 75 significantly decreased lipids, 86.7% (65/75) were TGs, and all the 24 lipids with a less than 0.5-fold decrease were TGs (Extended Data 1 Table S3, and Fig.4B). Therefore, although there were many lipids altered from UCA to RCA, we only identified one distinct lipid subclass (TGs) with respect to clinical classification.
In addition, as shown in Extended Data 1 Table S4 and Fig.4C, there were 119 identifiable lipids exhibiting statistically significant differential abundance between RCAs and UCAs, and the majorities (115 lipids) were decreased in RCAs as compared with UCAs. Of these, 46 out of 119 altered lipids were significantly decreased (FC< 0.75 or log2FC < -0.42) and 16 lipids showed a less than 0.5-fold decrease (log2FC < -1). Of the 46 significantly decreased lipids, 95.7% (44/46) were TGs, and all the 16 lipids were TGs (Extended Data 1 Table S4 and Fig.4C). Of note, as compared with HCs, we observed that TG became the predominant and distinct profile of altered lipid, which indicated that TG metabolism was severely disrupted in patients with CAs.
In addition, TGs were also the distinct profile of lipids in patients with UCA and RCA as compared with HCs in cohort 2 (Extended Data 2 and Extended Data 2 Fig. 4.)
Lipid-based diagnostic prediction model for CA vs. HC
To investigate the lipid-based diagnostic prediction model for CAs (UCAs + RCAs), we first discriminated CAs from HCs, and then differentiated UCAs from RCAs. We assigned cohort 1 (n = 360; HCs = 120, CAs = 240) to the training cohort, and cohort 2 (n = 180; HCs = 60, CAs =120) to the validation cohort. Therefore, the ratio of samples in the training and validation sets was 2:1. The baseline characteristics of subjects in the two cohorts are shown in Extended Data 1 Table S6.
To discriminate CAs from HCs, the two cohorts were subjected to an independent and comprehensive analysis to discover biomarkers. First, the lipid profiles between the HCs and the patients with CAs were compared in the two cohorts, which identified 61 differentially abundant (variable importance in projection (VIP) of > 1.0 and p value < 0.05) candidates in both cohorts. Secondly, we used random forest algorithms and the least absolute shrinkage and selection operator (LASSO) to decrease the number of lipid biomarkers, which produced 12 overlapping biomarkers from the two algorithms.
To make the model concise and consistent, we selected four lipids (PE(20:1p/18:2)+H, CerG1(d40:4)+NH4, TG (18:0p/16:0/16:1)+NH4, and TG (54:2e)+NH4) that exhibited consistent differential abundance in the UCA vs. RCA groups as our final biomarker. Then, we constructed a nomogram (Fig. 5A) and a diagnostic prediction score (dp-score) that was obtained according to the coefficients from generalized linear models (regressions) (Extended Data 1 Table S7). The risk score model was developed as follows: 0.6243+ (-0.4460*PE (20:1p/18:2)+H) + (-0.3263* CerG1(d40:4)+NH4) +(-0.5968* TG(18:0p/16:0/16:1)+NH4) + (-1.2855* TG(54:2e)+NH4).
First, the training cohort was used to train the 4-lipid prediction model. In the training cohort, the nomogram’s calibration plot demonstrated good agreement between observation and prediction (Fig. 5B). The Hosmer-Lemeshow (HL) test statistic was not significant (P = 0.244), indicating a good fit to the model. The receiver operator characteristic (ROC) analysis for the nomogram of the four biomarkers yielded areas under the ROC curve (AUCs) of 0.814 (95% confidence interval (CI): 0.767–0.861, Fig. 5 C, D) for the training cohort. The dp-score showed specificity and sensitivity of 60.0% and 87.5, respectively (Fig. 5D), using the best cutoff value.
Next, the four-lipid signature and the same statistical model were applied in the validation cohort (60 HC and 120 CA cases) to assess the accuracy of the signature. In the validation cohort, the lipid biomarkers showed excellent diagnostic accuracy to the identify patients with CAs. As it did in the training cohort, the nomogram showed favorable calibration in the validation cohort (Fig. 5E). The HL test result was not significant (p = 0.387), and in the validation cohort, the AUCs were 0.803 (95% CI: 0.735–0.871, Fig. 5F, G) for the nomogram. The specificity and sensitivity of the dp-score were 85 % and 69.2%, respectively (Fig. 5G).
Next, decision curve analysis (DCA) was used to compare the lipid performance of the model in the training and validation cohorts. To diagnose CAs in the validation and training cohorts, the developed model showed the highest net benefit within the ranges of most of the potential thresholds (Fig. 5H). Based on the lipid nomogram, the net benefits were similar, with several overlaps, within this range. This suggested the possibility of using the nomogram in clinical practice for the diagnosis and prediction of CAs.
A lipid-based combination diagnostic prediction model for CA vs. HC
Sex, age, and hypertension are known to be associated with CAs; therefore, whether a model combining our lipid signature with these three preoperative clinical features could improve the diagnostic accuracy to detect CA in clinic was assessed. The results showed that the diagnostic accuracy for CA of the combined signature was slightly better in both the validation and training cohorts (AUC values of 0.802 and 0.836, respectively, Fig. 5 I, J, and Extended Data 1 Fig. S2). Moreover, compared with the preoperative clinical features alone, including gender, age and hypertension, the combination signature demonstrated significantly improved diagnostic accuracy. Finally, using the cutoffs determined using the Youden index from this four-lipid signature model, all patients were categorized into high- and low-risk groups. The univariate and multivariate logistic regression analyses results are shown in Extended Data 1 Table S8. In both clinical cohorts, multivariate analysis showed that the four-lipid signature was an independent predictor to discriminate CAs from HCs, (training cohort: odds ratio [OR], 10.13; 95% CI, 5.93–17.74; p < 0 .001; validation cohort: OR, 12.66; 95% CI, 5.84–30.28, p < 0.001, Extended Data 1 Table S8).
Lipid-based diagnostic prediction model for UCA vs. RCA
To differentiate UCAs from RCAs, we assigned cohort 1 (n = 240; UCAs = 120, RCAs = 120) to the training cohort, and cohort 2 (n = 120; UCAs = 60, RCAs = 60) to the validation cohort. The baseline characteristics of cohort 1 and cohort 2 subjects are shown in Extended Data 1 Table S9. Consistent with the model for CAs vs. HCs, we still used the four biomarkers as the diagnostic prediction model for UCAs vs. RCAs. We constructed a nomogram and a dp-score that was obtained according to the coefficients and the constant derived from multinomial logistic regression (Extended Data 1 Table S10). The risk score model was obtained as follow: -0.0412+ (0.0320*PE (20:1p/18:2)+H) + (-0.7768* CerG1(d40:4)+NH4) +(-0.3816* TG(18:0p/16:0/16:1)+NH4) + (-1.2405* TG(54:2e)+NH) .
Fig. 6A shows the nomogram for UCAs vs. RCAs with the four biomarkers. The calibration plot of the nomogram that agreement between observation and prediction agreed well in the validation and training cohorts (Fig. 6B, E). The training cohort revealed a calibration slope of 1.0, with a nonsignificant HL test statistic of 0.176, and the test cohorts revealed a calibration slope of 0.744, with a nonsignificant HL test statistic of 0.345. The AUC values for the training and test cohorts were 0.775 (95% CI, 0.699–0.934, Fig. 6 C, D) and 0.721 (95% CI, 0.692–0.955, Fig. 6 F, G), respectively.
The DCA for the lipid nomogram of UCAs vs. RCAs is presented in Fig. 6 H. The decision curve indicated that at a threshold probability of 20–85% for a patient or doctor, then using the developed lipid nomogram for the diagnosis and prediction UCAs from RCAs would add more benefit than either the treat-none scheme or the treat-all-patients scheme. The net benefit of the test cohort was somewhat lower than the net benefit of the training cohort based on the lipid nomogram within this range. This suggested possibility of using the nomogram in clinical practice to diagnose and predict UCAs from RCAs
A lipid-based combination diagnostic prediction model for UCAs vs RCAs
To further increase the accuracy of diagnosis of URCs or RCAs in the clinic, we also assessed a combination the three preoperative clinical features and the developed model lipid signature. In the validation and training cohorts, the combination signature demonstrated slightly better diagnostic accuracy for URCs or RCAs (AUCs of 0.719 and 0.78, respectively, Fig. 6 I, J and Extended Data 1 Fig. S3). Moreover, in terms of diagnostic accuracy, compared that that of the preoperative clinical features sex, age, and hypertension, the combination signature demonstrated a significant improvement. Finally, using the Youden index-derived cutoffs from the four-lipid signature model, all patients were categorized into high- and low-risk groups. The results of for the univariate and multivariate logistic regression analyses results are shown in Extended Data 1 Table S11. In both cohorts, multivariate analysis identified the four-lipid signature as an independent predictor to discriminate UCA form RCA (training cohort: OR, 6.44; 95% CI, 3.68–11.56; p < 0.001; validation cohort: OR, 4.65; 95% CI, 2.10–10.84, p < 0.001, Extended Data 1 Table S11).
Lipid-based subtyping of RCA
A lower lipid intensity in RCA was associated with patients with severe RCA
To explore the lipid-defined specific subtypes within CA, stratification analysis was carried out using non-negative matrix factorization (NMF) consensus-clustering 19,20. NMF consensus-clustering was first performed for RCA in cohort 1, and two major lipid subtypes (R-I and R-II) were identified among the RCA samples (Fig. 7 A, Extended Data 1 Fig. S4), with 65 cases belonging to R-I and 55 belonging to R-II. The total lipid intensity of R-II was significantly lower than that of R-I (Fig. 7 B, p < 0.001). The relative abundance of the two subtypes according to age, sex, etc., is shown in Fig. 7C.
To explore the clinical characteristics between the R-I and R-II subgroups, the baseline characteristics of the subjects in the two subgroups were compared. There were no significant differences in sex, age, hypertension, diabetes mellitus, hyperlipemia, smoking, alcohol consumption, and BMI between R-I and R-II (Extended Data 1 Table S12). Subsequently, we compared the aneurysm characteristics associated with RCA, such as aneurysm size, location, aneurysm neck, single or multiple, regular or irregular, bifurcation or sidewall aneurysm, modified fisher grade (MFG) 21, Glasgow Coma Scale (GCS), the Coma number at onset, ventricular drainage (VD), and hospital day. There were significant differences in MFG, GCS, and the number of Coma at onset between R-I and R-II (Fig. 7 D); no other significant differences were observed (Extended Data 1 Table S12). Fig. 7 E-F show typical cases in the R-I subtype, and Fig. 7 G-H show typical cases in the R-II subtype. These results indicated that patients in the R-II subgroup were associated with severe RCA.
Diagnostic prediction potential of the four-lipid signature for patients with severe RCA
As mentioned-above, patients in R-II are often associated with severe or poor outcomes in the clinic; therefore, we next evaluated the diagnostic predictive potential of the four-lipid biomarker signature for patients with severe RCA. First, the Youden index-derived cutoffs from the four-lipid signature were used to separate the patients with RCA into high- and low-risk groups, which were then analyzed as independent predictors using univariate and multivariate logistic regression. Multivariate logistic regression analysis identified the four-lipid signature as an independent predictor of detection severe RCA, in the two clinical cohorts (training cohort: OR, 6.46; 95% CI, 2.71–16.55; p < 0.001; validation cohort: OR, 4.33; 95% CI, 1.30–16.06, p = 0.01, Extended Data 1 Table S13). Thus, in addition to diagnosing RCA, the four-lipid signature could also predict patients with severe RCA.
Lipid-based subtyping of UCA
UCA subtypes predict slow or rapid aneurysm growth
Next, NMF consensus-clustering was performed for UCA in cohort 1, and two major lipid subtypes (U-I and U-II) were identified among the UCA samples (Fig. 8A, Extended Data 1 Fig. S5) with 54 belonging to subtype U-I and 66 cases belonging to U-II. The total lipid intensity of U-II was significantly lower than that of U-I (Fig. 8B, p < 0.001). The relative abundance of the two subtypes by age, sex, and etc., is shown in Fig. 8C.
To explore the clinical characteristics between U-I and U-II, the baseline characteristics of the human subjects and the aneurysm characteristics in the two subgroups were compared. However, there were no significant differences in sex, age, hypertension, diabetes mellitus, hyperlipemia, smoking, alcohol consumption, BMI, aneurysm size, location, aneurysm neck, single or multiple, regular or irregular, bifurcation or sidewall, and hospital day between the U-I and U-II subtypes (Extended Data 1 Table S14).
We next checked the patients using MRA follow-up. Although, most patients with UCAs were treated when detected, we still enrolled 23 UCA patients in this study, among whom 20 were followed-up using MRA for more than seven years. According to NMF consensus-clustering, 10 patients were classified in the U-I subgroup and 10 were in the U-II subgroup. Other than the lipid intensity of U-II being significantly lower than that of U-I, we noticed that seven of ten UCAs (70%) in the U-II subtype were enlarged, while only two of ten UCAs (20%) in the U-I subtype were enlarged. Therefore, the aneurysm growth rate was higher in the U-II than in the U-I subtype (c2 test, p < 0.05), which indicated that the lower lipid intensity in the U-II subtype was associated with rapid CA progression, when compared to those in the U-I subtype. Fig. 8 D-E show the typical cases in U-I, and Fig. 8 F-G show the typical cases in U-II. These illustrated that rapid aneurysm growth in the U-II subtype group increased the chance of CA rupture, while the slow growth observed in the U-I subtype group might be associated with slower aneurysm progression and resistance to CA rupture.
Diagnostic prediction potential of the four-lipid signature for high-risk patients
Patients in the U-II subtype group are associated with rapid aneurysm growth and the chance of CA rupture; therefore, we next evaluated the diagnostic predictive potential of our lipid biomarkers for patients with UCA with rapid growth. First, the Youden index-derived cutoffs from four-lipid signature model were used to separate the patients with RCA into high- and low-risk groups, which were then assessed as independent predictor using univariate and multivariate logistic regression analyses. In both cohorts, multivariate logistic regression analysis identified the four-lipid signature an independent predictor to detect patients with UCA with rapid growth (training cohort: OR, 6.04; 95% CI, 2.73–14.05; p < 0.001; validation cohort: OR, 145.3, 95% CI, 21.3–3203.3, p < 0.001, Extended Data 1 Table S15).
In addition, when we extracted 20 UCA cases from cohort 1 and analyzed them separately. Surprisingly, using the Youden index-derived cutoffs from the dp-score, we observed a high consistency between NMF subtype and the risk groups. Among the 20 patients, multivariate analysis identified the 4-lipid signature as an independent predictor to detect patients with UCA with rapid growth (OR, 12.91; 95% CI, 1.61–180.5; p < 0.05, Extended Data 1 Table S16). These results showed that the lipid signature had a significant predictive potential to detect patients with UCA with rapid growth (high-risk UCA patients).