Novel Transcriptomic Interactions Between Immune Content and Genomic Classifier Predict Lethal Outcomes in High-grade Prostate Cancer

Abstract

Grade group 4 and 5 (GG-45) prostate cancer (PCa) patients are at the highest risk of lethal outcomes, yet lack genomic risk stratification for prognosis and treatment selection. Here, we assess whether transcriptomic interactions between tumor immune content score (ICS) and the Decipher genomic classifier can identify most lethal subsets of GG-45 PCa. We utilized whole transcriptome data from 8071 tumor tissue (6071 prostatectomy and 2000 treatment-naïve biopsy samples) to derive four immunogenomic subtypes using ICS and Decipher. When compared across all grade groups, GG-45 samples had the highest proportion of most aggressive subtype—ICS^High/Decipher^High. Subsequent analyses within the GG-45 patient samples (n = 1420) revealed that the ICS^High/Decipher^High subtype was associated with increased genomic radiosensitivity. Additionally, in a multivariable model (n = 335), ICS^High/Decipher^High subtype had a significantly higher risk of distant metastasis (hazard ratio [HR] = 5.41; 95% confidence interval [CI], 2.76–10.6; p ≤ 0.0001) and PCa-specific mortality (HR = 10.6; 95% CI, 4.18–26.94; p ≤ 0.0001) as compared with ICS^Low/Decipher^Low. The novel immunogenomic subtypes establish a very strong synergistic interaction between ICS and Decipher in identifying GG-45 patients who experience the most lethal outcomes.

Patient summary:

In this analysis, we identified a novel interaction between the total immune content of prostate tumors and genomic classifier to identify the most lethal subset of patients with grade groups 4 and 5. Our results will aid in the subtyping of aggressive prostate cancer patients who may benefit from combined immune-radiotherapy modalities.

High-grade patients with grade group 4 and 5 (GG-45) prostate cancer (PCa) experience worse clinical outcomes following treatment than those with GG1–3, in whom progression-free 5-yr recurrence rates are as low as 35% [1]. While GG-45 patients have an overall inferior clinical outcome, the genomic heterogeneity manifest in poorly differentiated GG-45 tumors can be exploited to identify patient subsets that are at higher risk of lethal outcomes and that may benefit from targeted treatment strategies [2,3]. Further, in GG-45 tumors, the effect of the complex interplay between immune cells in the prostate tumor microenvironment on the genomic classifier (Decipher score) remains unexplored. This is particularly important as intratumoral immune content of PCa appears to be associated with disease progression [4]. We hypothesized that the immunogenomic interaction of immune content and Decipher score can be leveraged to inform subsets of high-grade PCa patients who are at maximum risk of lethal outcomes.

A total of 8071 PCa patients were identified from Decipher GRID, including 6071 prostatectomy and 2000 treatment-naïve validation biopsy samples (Supplementary Fig. 1). All the samples (n = 8071) underwent uniform transcript quantification using a Human Exon 1.0 ST microarray (Thermo-Fisher, Carlsbad, CA, USA) at a Clinical Laboratory Improvement Amendments (CLIA)-certified clinical laboratory (Decipher Bioscience, Inc., San Diego, CA, USA) and were used in the immunogenomic subtyping (Supplementary Fig. 1) [5]. Institutional review board approval was obtained prior to analysis. The immune content score (ICS) was derived computationally using the mean expression of 264 immune cell–specific genes for all the samples. A detailed description of gene selection and ICS calculation is provided in the work of Zhao et al [4]. Decipher score is a commercially validated 22-gene classifier that provides a robust distant metastasis (DM) risk score and is categorized as Decipher^High (Decipher score >0.60) and Decipher^Low (Decipher score ≤0.60) [6]. ICS categorization—ICS^High and ICS^Low—was based on the observed nonlinear association between ICS and the probability of high-risk Decipher in a binomial spline model using both prostatectomy (n = 6071) and biopsy (n = 2000) samples. Probabilities of high-risk Decipher along with ICS distribution are shown in Supplementary Fig. 2. In the spline model, an inflection point was noted, beyond which the probability of high genomic risk was higher, which was located near the 75th percentile in both independent datasets (1.42 for prostatectomy and 1.37 for biopsy). A correlation analysis between ICS and Decipher score was also carried out using Spearman correlation. Immunogenomic subtypes were grouped as follows: ICS^High/Decipher^High, ICS^Low/Decipher^High, ICS^High/Decipher^Low, and ICS^Low/Decipher^Low. We also utilized a validated 24-gene expression signature postoperative radiation therapy outcome score (PORTOS) and a radiation sensitivity index (RSI) score as a surrogate for response to radiotherapy and genomic radiosensitivity, respectively [7,8]. To deconvolute the absolute abundance of immune cell contents of major cell types, we performed a CIBERSORT analysis and compared the distribution of these cell types across immunogenomic categories [9]. Kaplan-Meier curves and a multivariable Cox proportional hazard model were used to estimate the median time to the events of interest and hazard ratio (HR), respectively. Lethal outcomes including DM and progression-free survival were used as primary endpoints. For both the endpoints, the date of radical prostatectomy was used as a follow-up start date. For metastasis outcomes, the censoring date was defined as the last date of metastasis-free evidence. To estimate the freedom from progression based on PCa-specific mortality (PCSM), death due to causes other than PCa was censored.

When compared across ICS quartiles using all the prostatectomy cases (n = 6071), samples with ICS in the highest quartile had the highest median Decipher score, whereas the lowest ICS quartile had lower Decipher scores (p ≤ 0.0001; Supplementary Table 1). Accordingly, ICS correlates modestly with Decipher (Spearman correlation 0.17, p ≤ 0.0001; Fig. 1A). The association of ICS and Decipher was also validated in treatment-naïve biopsy samples (no outcome data available, n = 2000; Supplementary Table 1). Overall clinicopathologic characteristics of immunogenomic subtypes are provided in Supplementary Table 2. When comparing immunogenomic subtypes across grade groups, a large proportion of ICS^High/Decipher^High subtypes were clustered within high-grade GG-45 compared with low-intermediate grade GG1–3, which had the smallest proportion of ICS^High/Decipher^High subtypes (37% vs 26%, p ≤ 0.0001; Fig. 1B and Supplementary Table 3). Consistently, 57% of the high-grade GG-45 biopsy samples had ICS^High/Decipher^High compared with only 35% of the aggressive subtypes in GG1–3 sample (p ≤ 0.0001; Fig. 1C and Supplementary Table 3). We then selected high-grade prostatectomy (GG-45 [n = 1420]) and validation biopsy (GG-45 [n = 319]) samples to evaluate the association between immunogenomic subtypes and radiation response signatures. In both samples, ICS^High/Decipher^High subtype was genomically more radiosensitive (a lower median RSI score) and had a higher median PORTOS (Supplementary Fig. 3A–D). We further performed the CIBERSORT algorithm to derive absolute abundance of immune cell types in GG-45 tumors and identified an enrichment of T cell, B cell, mast cells, and monocyte/macrophages. When compared across immunogenomic subtypes, ICS^High/Decipher^High had higher median abundance of T cells (p = 0.004) and monocyte/macrophages (p = 0.02), whereas that of B cells (p ≤ 0.0001) were lower (Fig. 2A–C and Supplementary Table 3). CIBERSORT results were also validated in biopsy GG-45 tumor samples and were largely consistent with those from prostatectomy samples (Fig. 2D–F and Supplementary Table 3).

Fig. 1 –

Immunogenomic subtypes. (A) Scatter plot between ICS and Decipher to show a correlation using Spearman coefficient in prostatectomy samples (n = 6,071). Immunogenomic subtypes were grouped as follows: ICS^High/Decipher^High (>75th percentile of ICS and Decipher >0.60), ICS^Low/Decipher^High (≤75th percentile of ICS and Decipher >0.60), ICS^High/Decipher^Low (>75th percentile of ICS and Decipher ≤0.60), and ICS^Low/Decipher^Low (≤75th percentile of ICS and Decipher ≤0.60). (B) Proportion of immunogenomic subtypes across all the grade groups in prostatectomy samples (n = 6071). Compared with other grade groups, GG-45 has the highest combined proportion of aggressive ICS^High/Decipher^High subtype (p < 0.0001). (C) Proportion of immunogenomic subtypes across all the grade groups in treatment-naïve biopsy samples (n = 2000). Compared with other grade groups, GG-45 has the highest combined proportion of aggressive ICS^High/Decipher^High subtype (p < 0.0001). GG = grade group; GG-45 = grade groups 4 and 5; ICS = immune content score.

Fig. 2 –

CIBERSORT immune cell deconvolution within GG45 prostatectomy and validation biopsy samples. Absolute immune abundance was obtained across the GG-45 prostatectomy samples. A total of 1402 GG-45 samples were included in CIBERSORT algorithm. A deconvolution p value of ≤0.1 was used to exclude samples with poor deconvolution results. Finally, 600 samples passed the criteria and were used in the figure. Proportions of immune cell types across immunogenomic subtypes using 600 GG-45 samples were compared. Eosinophils are not shown due to an extremely small proportion. Comparisons of (A) T cells, (B) B cells, and (C) monocytes/macrophages across immunogenomic subtypes using 600 prostatectomy samples. A total of 319 GG-45 validation biopsy samples were included in the CIBERSORT algorithm. A deconvolution p value of ≤0.1 was used to exclude samples with poor deconvolution results. Finally, 136 samples passed the criteria and were used in the boxplots. Comparison of (D) T cells, (E) mast cells, and (F) B cells across immunogenomic subtypes using 136 treatment-naïve biopsy samples. All reported p values are for the comparison between ICS^High/Decipher^High and ICS^Low/Decipher^Low, and are adjusted for multiple comparisons in Dunn’s test of post hoc analysis using Bonferroni methods. GG-45 = grade groups 4 and 5; ICS = immune content score.

Finally, to assess the impact of immunogenomic subtypes on lethal outcomes, we utilized a survival cohort (n = 1075; Supplementary Fig. 1), which comprised the cases with known metastasis information. GG-45 samples (n = 335) with an event-free median follow-up of 89 mo (60–130 mo) were analyzed in the Cox model. A similar analysis was performed in GG1–3samples; however, due to a lower proportion of ICS^High/Decipher^High subtypes and fewer events in low-intermediate groups, survival estimates were not significant across all the outcomes (Supplementary Fig. 4A–F). In a multivariable Cox model using GG-45 cases, ICS categories alone did not predict DM, whereas they were significantly associated with PCSM (Table 1). When introduced as a continuous variable in a Cox model, ICS association with DM remained weaker, whereas a nonlinear association was observed between higher ICS and PCSM (data not shown). The immunogenomic subtypes established a very strong and consistent synergistic interaction in the prediction of metastasis between ICS and Decipher (p^Interaction = 0.002; Table 1 [models 1–3]). However, the multiplicative interaction was not significant in the prediction of PCSM (p^Interaction = 0.1). Compared with ICS^Low/Decipher^Low patients, those with ICS^High/Decipher^High had extremely high risk of both DM (HR = 5.41; 95% confidence interval [CI], 2.76–10.6; p < 0.0001) and PCSM (HR = 10.6; 95% CI, 4.18–26.94; p < 0.0001; Table 1). Similarly, in the Kaplan-Meier analysis, ICS^High/Decipher^High subtypes had lowest metastasis and progression-free survival (Fig. 3A and 3B). Relative excess risk due to interaction (RERI) statistics derived from the estimates of Cox models (for both DM and PCSM) also revealed a positive additive interaction between immunogenomic subtypes (RERI^Metastasis = 4.08 and RERI^PCSM = 7.5; Table 1). Therefore, the immunogenomic subtypes were able to identify patients with the most lethal outcomes as compared with Decipher score and ICS alone. Taken together with the modest correlation between ICS and Decipher score, the observed association of the immunogenomic subtypes and lethal outcomes is not driven by Decipher risk classifier alone; instead, it is a synergistic effect of aggressive immunogenic interaction within GG-45.

Table 1 -.

Risk of distant metastasis prostate cancer–specific mortality (PCSM) by Decipher score, ICS, and immunogenomic subtypes among GG-45 patients in multivariable Cox model (n = 335).

	Risk of metastasisa			Risk of PCSMb
	Events/no. at risk	HR (95% CI)	p value	Events/no. at risk	HR (95% CI)	p value
Model 1
Decipher score
DecipherLow	98/229	1 (ref)		35/198	1 (ref)
DecipherHigh	71/106	2.01 (1.44–2.78)	<0.0001	34/91	2.48 (1.50–4.10)	0.0004
Model 2
ICS
ICSLow	146/288	1 (ref)		58/256	1 (ref)
ICSHigh	23/47	1.16 (0.70–1.94)	0.5	11/33	2.64 (1.36–5.11)	0.004
Model 3
Immunogenomic subtypes
ICSLow and DecipherLow	88/197	1 (ref)		31/176	1 (ref)
ICSLow and DecipherHigh	58/91	1.71 (1.20–2.43)	0.002	27/80	2.26 (1.31–3.91)	0.003
ICSHigh and DecipherLow	10/32	0.62 (0.27–1.43)	0.6	4/22	1.84 (0.64–5.26)	0.2
ICSHigh and DecipherHigh	13/15	5.41 (2.76–10.6)	<0.0001	7/11	10.6 (4.18–26.94)	<0.0001

CI = confidence interval; HR = hazard ratio; ICS = immune content score; ref = reference; RERI = relative excess risk due to interaction.

All the models are adjusted for pretreatment prostate-specific antigen, seminal vesicle invasion, surgical margins, lymph node invasion, and extraprostatic extension.

Model 1 estimates the risk of distant metastasis and PCSM for Decipher^High as compared with Decipher^Low, model 2 estimates the risk of distant metastasis and PCSM for ICS^High as compared with ICS^Low, and model 3 estimates the risk of distant metastasis and PCSM for immunogenomic subtypes.

The likelihood-ratio value of interaction p value between ICS and Decipher was 0.002 for metastasis and 0.1 for PCSM. Whereas on an additive scale, RERI statistics (HR[1,1] − HR[0,1] − HR[1,0] + 1) showed a positive additive interaction between immunogenomic subtypes: RERI^Metastasis [5.41 − 1.71 − 0.62 + 1] = 4.08 and RERI^PCSM [10.6 − 2.26 − 1.84 + 1] = 7.5.

^aIn the survival cohort, only 311 cases had complete follow-up information on time to were included in multivariable model.

^bIn the survival cohort, only 273 cases had complete follow-up information on time to PCSM and were included in multivariable model.

Fig. 3 –

Time to distant metastasis and progression-free survival in Kaplan-Meier curves. (A) Time to distant metastasis by immunogenomic subtypes in GG-45 survival cohort (n = 335). ICS^High/Decipher^High has the lowest median time to distant metastasis among all subtypes. A total of 311 at-risk GG-45 patients with complete time to metastasis information were included in the K-M curves. (B) Prostate cancer–specific survival by immunogenomic subtypes in GG-45 survival cohort (n = 335). ICS^High/Decipher^High has the lowest median survival among all subtypes. A total of 273 at-risk GG-45 patients with complete time to PCSM information were included in the K-M curves. GG-45 = grade groups 4 and 5; ICS = immune content score; K-M = Kaplan-Meier; PCSM = prostate cancer–specific mortality.

Overall, we identified a novel interaction between the immune content of prostate tumors and the genomic classifier that is highly clustered within high-grade GG-45 tumors, as compared with low- to intermediate-grade tumors. This analysis reveals significant heterogeneity in the risk of lethal outcomes among GG-45 patients with similar genomic risk, whereby not every GG-45 patient with a high Decipher score experiences a poor outcome. Rather, a subset of patients with tumors having both a high Decipher score and a high immune content has the most lethal outcomes. Further, an immunogenomic subset of patients at risk of lethal outcomes also demonstrated a consistent relation with the markers of radiation response, suggesting that this GG-45 subset may benefit from combined immunotherapy and radiotherapy. However, it is unclear whether the poorer outcomes with ICS^High/Decipher^High subtypes are driven by the differential immune abundance of T cells, B cells, and monocytes/macrophages, and therefore further research is warranted to unravel the biologic mechanisms of this association. Future direction of this work includes the validation of these novel immune subtypes in other genomic platforms.

The major caveats to this work include possible variations in computationally estimated genomic immune content that relies on mRNA expression, which may not represent unbiased estimates of actual immune content derived from directly measuring immune cell numbers within prostate tumors. However, the method used to derive immune content yielded results comparable with those presented by Zhao et al [4], which serve as an external validation of the work presented herein. Additionally, despite the large number of samples in the GRID, the long-term clinical data acquisition through follow-up is lagging behind in GRID and therefore not every sample has rigorous follow-up information available. Lastly, our findings may not be applicable to other tumor genomic platforms, as the association of ICS and Decipher score is based on the Decipher GRID platform. Nonetheless, given that Decipher is currently used routinely in clinical care, our results can be very readily validated in various ongoing clinical trials simultaneously and promises to be practice changing in the near term.