Expression of 5-Fluorouracil and platinum related markers in LAGC patients who received curative surgery in the SRRSH cohort and TCGA cohort
We constructed a Multiple Tissue Array (MTA), including all patients enrolled, as we previously reported (15). RRM1, RRM2, RRM2B, DUT, TS, TP, POLH, and Ki67 were stained on the slides of MTAs. All demographic data and contingencies with biomarkers in the SRRSH cohort were listed in Table S2. Among all the GC cases, positive expression of RRM1, RRM2, RRM2B, POLH, DUT, KI67, TS, TP were 110(45.4%), 135(55.6%), 137(55.1%), 154(63.6%), 140(57.4%), 155(63.8%), 141(57.5%), 59(23.4%), respectively. All the biomarkers were higher in elderly and male patients except TYMP, compared to younger and female patients. RRM1, DUT, and TYMS were preferentially higher in stage III, low tumor grade and proximal GC patients, whereas RRM2, TYMS and TYMP were associated with poor differentiation.
The baseline data and contingencies of biomarkers of the TCGA cohort were listed in Table S3. All the mRNA expression of selected biomarkers was dichotomized with median levels. In 279 LAGC cases, positive expression of RRM1, RRM2, RRM2B, POLH, DUT, MKI67, TYMS, TYMP were 136(49.0%), 140(50.0%), 126(45.0%), 150(54.0%), 146(52.0%), 135(48.0%), 136(49.0%), 155(56.0%), respectively. POLH, DUT was associated with high tumor grade and younger age.
5-Fluorouracil and platinum related biomarkers predict the efficacy of chemotherapy in LAGC
Univariate COX Proportional analysis indicated that adjuvant chemotherapy was beneficial to high RRM1 (HR, 0.52, 95%CI 0.33–0.84), high DUT (HR,0.65, 95%CI 0.43-1.00), low RRM2 (HR,0.51, 95%CI 0.30–0.84), low RRM2B (HR, 0.57, 95%CI 0.35–0.97), low POLH (HR,0.34, 95%CI 0.19–0.60), low KI67 (HR,0.52, 95%CI 0.29–0.92), low TYMS (HR,0.55, 95%CI 0.31–1.01) or low TYMP (HR,0.60, 95%CI 0.40–0.89) LAGC patients (Table 2, Fig. 1). To avoid the confounding effect, we further conducted a multivariate Cox proportional analysis. In low RRM2B, low RRM2, low POLH, low KI67 or low RRM1 patients, chemotherapy remained to protect factors in LAGC (P < 0.05), whereas other biomarkers showed a consistent trend with the univariate model.
Table 2
Univariable and Multi-Variable Cox proportional hazard analysis in SRRSH cohort
| | Univariable Model | | Multi-Variable Model | |
Gene signatures | Treatment | HR (95%CI) | P | †HR (95%CI) | P |
RRM1 | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.74(0.45–1.23) | 0.25 | 0.72(0.42–1.23) | 0.23 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.52(0.33–0.84) | 0.007 | 0.66(0.40–1.10) | 0.11 |
RRM2B | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.57(0.35–0.97) | 0.04 | 0.55(0.31–0.99) | 0.045 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.70(0.45–1.09) | 0.12 | 0.85(0.72–1.92) | 0.51 |
RRM2 | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.51(0.30–0.84) | 0.008 | 0.54(0.32–0.89) | 0.02 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.78(0.49–1.24) | 0.3 | 0.97(0.57–1.66) | 0.92 |
POLH | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.34(0.19–0.60) | 0.0003 | 0.37(0.20–0.68) | 0.001 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.92(0.60–1.42) | 0.71 | 1.05(0.66–1.67) | 0.84 |
dUTPase (DUT) | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.63(0.35–1.11) | 0.11 | 0.63(0.35–1.31) | 0.13 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.65(0.43-1.00) | 0.05 | 0.82(0.51–1.31) | 0.41 |
Ki67 | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.52(0.29–0.92) | 0.02 | 0.49(0.25–0.95) | 0.034 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.74(0.48–1.13) | 0.17 | 0.85(0.55–1.33) | 0.48 |
TYMS | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.55(0.31–1.01) | 0.05 | 0.60(0.37–1.14) | 0.12 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.78(0.52–1.18) | 0.25 | 0.90(0.57–1.42) | 0.65 |
TYMP | | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.60(0.40–0.89) | 0.01 | 0.67(0.44–1.03) | 0.07 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.84(0.45–1.57) | 0.6 | 0.97(0.47–1.97) | 0.92 |
Chemo-signature | | | | |
low | Surgery alone | Reference | | Reference | |
| Surgery + AC | 1.37(0.65–2.97) | 0.41 | 2.11(0.84–5.56) | 0.11 |
medium | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.61(0.37-1.00) | 0.05 | 0.64(0.38–1.10) | 0.1 |
high | Surgery alone | Reference | | Reference | |
| Surgery + AC | 0.39(0.21–0.72) | 0.003 | 0.44(0.23–0.83) | 0.01 |
* Indicate p-value < 0.05, AC: Adjuvant Chemotherapy †: Adjusted by age and sex. | |
The result was validated in the TCGA cohort. Univariate COX model showed high RRM1 (HR, 0.47, 95%CI 0.24–0.90), low RRM2 (HR, 0.42, 95%CI 0.23–0.76), low RRM2B (HR,0.38, 95%CI 0.20–0.69), low POLH (HR,0.34, 95%CI 0.19–0.60), low MKI67 (HR, 0.40, 95%CI 0.21–0.72), low TYMS (HR, 0.36, 95%CI 0.19–0.66) or low TYMP (HR, 0.35, 95%CI 0.16–0.69) LAGC patients, high DUT was marginally associated with chemotherapy efficacy (P = 0.05), whereas low DUT demonstrated high responsiveness to chemotherapy (P = 2.0E-04). The multivariate Cox model also showed consistent results with the univariate model (Table S4).
C-Score is an ideal predictor for adjuvant therapy efficacy in LAGC
Cumulative analyses were conducted based on the risk scores (C-Score) generated from the identified candidate biomarkers that were significantly associated with chemotherapy efficacy. The C-Score was calculated as the sum of dichotomized biomarkers (0/1), the value of RRM1 and DUT were labeled as positive for their expressions associating with high efficacy, whereas the rest biomarkers were labeled as a minus for low efficacy. The formula was as following: C-Score = RRM1 + DUT-RRM2-RRM2B-POLH-MKI67-TYMS-TYMP. The sum was further categorized as three groups, score − 4 and − 5 were group low, score − 2 and − 3 were group medium, score − 1,0,1 were group high. The univariate COX model indicated that high group response to chemotherapy better than group medium (HR, 0.39, 95%CI 0.21–0.72 vs. HR, 0.61, 95%CI 0.37-1.00), but the group low had an inadequate response to chemotherapy (HR, 1.37, 95%CI 0.65–2.97). The multivariate Cox model indicated similar results (Table 2). Kaplan-Meier analysis showed that patients who received chemotherapy had significantly better OS than patients who received surgery only in group high and group medium (log-rank P = 0.003, and 0.05, respectively, Fig. 2). The validation in the TCGA cohort also showed similar results that patients in group high and medium response to chemotherapy better than patients who only received surgery (HR, 0.36, 95%CI 0.15–0.75, HR, 0.47, 95%CI 0.25–0.84, respectively). The patients in group low demonstrated no difference in death risk and OS in surgery plus chemotherapy and surgery alone GC patients.
To further explore the value of C-Score, we developed a predictive model by including interaction terms between chemotherapy and each of the other covariates. The concordance test indicated that including the interaction effect has significantly improved the prediction accuracy. After performing model selection logic, the final model included age, gender, chemo, tumor grade, RRM1, POLH, TYMD, TYMP, and interaction terms between chemo and each of the other covariates. The model accuracy based on the SRRSH cohort (Harrel’s C = 0.7212) was slightly higher compared to that of the TCGA cohort (Harrel’s C 0.6823). Also, due to the limited number of observations with survival time longer than three years in the TCGA cohort, here we only reported the model parameters estimated based on the SRRSH cohort. The ROC curves and AUC at specific time points have been presented in Fig. 3A. The AUC fluctuated between 0.75 to 0.81 across different survival tie points. Due to the limited number of observations with survival time longer than 84 months, the ROC value remained the same after the 84th month measurement. The time-dependent AUC and corresponding 95% confidence limits were presented in Fig. 3B, which indicated that the model prediction accuracy was comparatively stable before the 84th month. Therefore, we suggest applying this model for estimation of no-longer than 5 to 6-year survival outcomes.
Based on this model, we developed a calculator to predict the hazard ratio of receiving chemotherapy against not receiving chemotherapy for individual patients with different characteristics and biomarker profiles. A sample calculator is available in the supplementary materials in the excel spreadsheet format. The information needed to calculate the HR(chemo/non-chemo) includes gender (0-male/1-female), age in year, grade (1 to 4), RRM1, POLH, TYMS, and TYMP (0-no/1-yes) (Attached file1). One limitation of this calculator is that the prediction for low-grade patients may not be applicable. Since chemotherapy has a very significant protective effect for most low-grade patients, which could have potentially masked the modification effect from biomarker profile differences, our predictive model cannot detect the difference due to the biomarker profile variance.
The C-Score is associated with mutation type and burden in LAGC
To identify the possible mechanism underlying the predictive efficacy of C-Score, we analyzed the mutation profiles of the LAGC from the TCGA cohort. The mutation profiles in group low demonstrated a significantly higher rate of functional mutations than group median and high group (Fig. 4). Notably, the mutation type in low group cases was mainly frame shift ins/del, non-sense, and splice site, which could distinctly alter the protein translating, whereas the primary mutation type in group median and high was missense mutation. The highest frequency of the mutated gene in group low was ARID1A, and in group median and high was TP53.
To further validate our findings, we analyzed the association between C-Score biomarkers’ expression and IC50 value to 5-fluorouracil in GC cell lines from a published study (22). The result demonstrated that the IC50 value was significantly correlated with chemo-signature in GC cell lines (P = 0.044, Table 3). Moreover, the mutation profile showed a similar distribution pattern to the data from the TCGA cohort, which also showed that high frequency of frameshift mutations in group low and high frequency of missense mutation in group median and group high. The highest mutated gene was EP300 in group low and TP53 in group median and high (Fig. 4).
Table 3
Correlation analysis between C-Score and IC50 in 28 GC cell lines
IC50/C-Score | 0 | 1 | 2 | |
Resistant | 2(100%) | 10(58.8%) | 2(22.2%) | |
Sensitive | 0(0%) | 7(41.2%) | 7(77.8%) | *P = 0.044 |
*Chi-square test | | | | |