A Stepwise Integrated Approach to Personalized Risk Predictions in Stage III Colorectal Cancer

This article is featured in Highlights of This Issue, p. 1121

Colorectal cancer is a leading contributor to morbidity and mortality worldwide (1). Colorectal cancer incidence is expected to increase due to longer life expectancy and lifestyle changes, such as poor diet and decreased physical activity (1). The current standard-of-care for patients with stage III colorectal cancer prescribes surgical resection followed by 5-fluorouracil (5-FU)-based adjuvant chemotherapy (2). Response to treatment varies greatly, and numerous studies have attempted to reconcile differential response to chemotherapy with the underlying molecular characteristics of the primary tumor (3). The mechanism of action of chemotherapeutic agents, such as 5-FU, is centered on their DNA-damaging effects and ability to induce cell-cycle arrest and apoptosis (4). Alterations in apoptotic signaling pathways foster tumor progression and contribute to insufficient response to treatment (5, 6). The major pathway of apoptosis activation by 5-FU and other genotoxic drugs is mitochondrial outer membrane permeabilization (MOMP), which triggers a complex signaling network of pro- and anti-apoptotic proteins controlling the activation of effector caspases (7). Cytochrome c and SMAC are released into the cytosol following MOMP and activate Apaf-1 and Procaspase-9 or antagonize the major caspase inhibitor XIAP. Caspase-9 can activate Caspase-3, which drives multiple positive feedback loops that ensure an efficient execution of apoptotic cell death (7).

Systems analysis of apoptosis signaling has the potential to deliver superior prognostic markers by accounting for the relative abundance of key proteins, network connectivity, and regulations of the pathways involved (8, 9). In previous work, our group developed a mathematical model, APOPTO-CELL, on the basis of ordinary differential equations that reliably simulates signaling kinetics within the caspase activation network downstream of MOMP (10, 11). As input, APOPTO-CELL requires protein concentrations of the key downstream apoptosis regulators Apaf-1, Procaspase-9, XIAP, SMAC, and Procaspase-3 to calculate cellular apoptosis competency. Apoptosis competency is calculated as the ability of Caspase-3 to cleave sufficient amounts of its substrates to ensure efficient execution of apoptosis. The mathematical model has been extensively validated in-house using single cell- and population-based approaches in cervical, colorectal, and glioblastoma cell line models (10, 12, 13). In initial proof-of-concept work, high apoptosis competency was associated with better outcome among chemotherapy-treated patients with colorectal cancer and glioblastoma (13–15).

Here, we performed a discovery and validation study aimed at developing stratification signatures to predict personalized risk in stage III colorectal cancer (Fig. 1). We investigated the potential of APOPTO-CELL as a stand-alone tool. Moreover, we synthetized a systems-biology signature by linking APOPTO-CELL with protein expression data and harnessed a machine learning approach to examine the prognostic relevance of ordinary differential equation (ODE)-based modeling, proteomics, and clinicopathologic data. We validated our signatures in an independent external cohort and assessed their prognostic value for distinct molecular subtypes of colorectal cancer.

We developed prognostic signatures for patients with stage III colorectal cancer from 3 distinct and independent collections: discovery, expansion, and validation cohorts. The discovery cohort was an in-house multicenter study of patients with stage III colorectal cancer (n = 120). The expansion dataset included stage III colon cancer participants (n = 157) with known microsatellite instability (MSI) and BRAF status (GEO repository under accession 39582; ref. 16). The validation cohort was composed of patients with stage I–IV colon cancer (n = 136; COAD) from The Cancer Genome Atlas (TCGA) project.

A detailed description of the patients data handling and inclusion criteria for downstream analyses is presented in Supplementary Methods SM1.

We developed 3 prognostic signatures: APOPTO-CELL, APOPTO-CELL-PC3, and RF.

APOPTO-CELL is a mathematical model of caspase activation resulting in apoptosis (10, 11). The model is composed of 53 ODEs, 19 state variables, and 75 kinetic parameters. It calculates apoptosis execution kinetics from the input proteins Apaf-1, Procaspase-9, XIAP, SMAC, and Procaspase-3. Previous quantitative studies found that Apaf-1 protein levels were not rate-limiting for apoptosome formation in colon cancer cells (14). Thus, Apaf-1 patient-specific protein levels were replaced by the median expression (0.123 μmol/L) previously determined in stage II/III colorectal cancer tumors (14). An amount of more than 25% substrate cleavage (SC) by active Caspase-3 served as the APOPTO-CELL prediction for apoptosis competency in line with previous single-cell imaging findings (10). While APOPTO-CELL was designed to use protein concentrations as inputs, surrogates such as RNA transcripts or gene expression can be used when protein expression is not available.

Molar protein concentrations were estimated from the normalized signal intensities via a novel pipeline illustrated in Fig. 2A.

First, a kernel distribution object was constructed from the normalized protein intensities (or surrogate transcript amounts) with the MATLAB function fitdist (Statistics Toolbox) with a normal kernel smoothing, unbounded kernel support, and default bandwidth (steps 1–2). This step resulted in a smooth nonparametric representation of the distribution of the protein (or surrogate) concentrations (step 3). Second, a kernel probability distribution object was constructed from the reference distribution with the same method as for the protein (or surrogate) distribution, except with positive kernel support (right-hand side, steps 1–3). Third, the kernel distribution of the protein (or surrogate) intensities was used to calculate the smoothed cumulative distribution corresponding to the protein (or surrogate) intensity of each patient (left-hand side, step 4). Finally, the protein (or surrogate) cumulative distribution for each patient was converted into a corresponding absolute concentration using the inverse cumulative distribution of the reference distribution kernel (MATLAB function icdf from the Statistics Toolbox, step 5). Step 6 shows the transformation function from protein (or surrogate) signal intensities to estimated protein concentration.

We estimated APOPTO-CELL inputs from reverse-phase protein arrays (RPPAs), gene expression, and RNA transcripts for the discovery, expansion, and validation cohorts, respectively (Supplementary Methods SM2).

Procaspase-3 exerts pivotal roles in multiple aspects of apoptosis both via apoptosome-dependent and -independent pathways. Thus, we generated a new signature (APOPTO-CELL-PC3) by combining the APOPTO-CELL model output with the expression of Procaspase-3 (Results and Fig. 4A). As with the APOPTO-CELL signature, Procaspase-3 concentration was estimated from RPPA, gene expression, and RNA transcripts for the discovery, expansion, and validation cohorts, respectively.

The Random Forest (RF) machine learning approach was conducted in MATLAB [built-in TreeBagger class (Statistics Toolbox) with 100.001 trees, surrogate decision split flags “on”; other options were set to “default”] and was used to calculate the RF classifier as a predictor for disease recurrence (17). We selected 36 months as a cutoff, as most recurrence events (87.9%) occurred within this time. For training purposes, we only considered patients known to have a recurrence or to be recurrence-free at 36 months (only 12.5% of patients were lost to follow-up). The training set consisted of two thirds of the patients, randomly sampled for each decision tree in the RF. For validation, recurrence predictions averaged from RF decision trees were analyzed for accuracy on patients that did not contribute to defining these trees (“out of bag” predictions). This allowed an unbiased validation without a separate validation cohort. The importance of each variable as a predictor was assessed by calculating the Permuted Variable Delta Error, a measurement for the increase in the prediction error upon random perturbation of the predictor values. The predictors that contributed positively to the accuracy of the RF (permuted variable delta > 0.05) were used to define the reduced RF classifier. Classifier performance was assessed by the area under the receiver operating characteristic (ROC) curve computed from the validation results.

We generated a synthetic cohort to further understand how the RF classifier makes predictions. This synthetic cohort had one entry for each unique hypothetical patient that each of the decision trees included in the RF ensemble was learned from. The complete set of hypothetical patients was compiled from all permutations of the unique values of each predictor. For each categorical predictor, values were given by each level (e.g., low-, medium-, and high-risk for the APOPTO-CELL-PC3 signature). For discrete numerical predictors (age and nodal count), the full set of values was taken from the complete list of unique cutoff points extracted from all the trees included in the RF ensemble.

We used the distribution of the patient-specific recurrence probabilities determined by the RF classifier in the discovery cohort to define risk groups (Fig. 5E). Survival analysis was performed using all patients (n = 120), including patients lost to follow-up who did not contribute to the RF classifier identification. The probability of recurrence, and thus the risk group, was computed on the basis of the out-of-bag predictions for patients included in the training set and de novo predictions for patients lost to follow-up within the first 36 months.

Kaplan–Meier estimates were used to compare disease-free (DFS) and overall survival (OS) curves between groups and statistically significant differences were determined by log-rank tests. We used univariate and multivariate Cox proportional hazard models to estimate the relative risks of outcome associated with the signatures (APOPTO-CELL, APOPTO-CELL-PC3, and RF) and clinical factors. We computed HRs and 95% confidence intervals (CIs) and we evaluated statistical significance with likelihood ratio tests. An in-depth account of the survival analyses is provided in Supplementary Methods SM3. Association between APOPTO-CELL and APOPTO-CELL-PC3 signatures with T stage and lymphovascular invasion was assessed by χ² test.

Data processing and statistical analysis were performed in MATLAB (MATLAB and Statistics Toolbox Release 2014b, The MathWorks, Inc.) unless stated otherwise. Median follow-up time among censored patients, log-rank tests, Cox regression modeling, and assessment of proportional hazards assumptions were conducted in R (version 3.3.0) employing the functions Surv, coxph, and cox.zph from the package “survival” (version 2.39-2) and Anova from the package “car” (version 2.1-2). All tests were 2-sided and P < 0.05 was considered statistically significant.

We explored whether mathematical modeling of effector caspase activation in the mitochondrial apoptosis pathway could identify high-risk patients in a multicenter cohort of patients with stage III colorectal cancer (n = 120; Fig. 1). Supplementary Table ST1 presents clinicopathologic and demographic characteristics.

Protein amounts of Procaspase-9, XIAP, SMAC, and Procaspase-3 were quantified in primary tumor samples by RPPAs (Supplementary Methods SM2).

We established a workflow to determine the molar concentrations required by APOPTO-CELL as input from normalized RPPA data (Fig. 2A and Materials and Methods). For each protein, absolute concentrations (Fig. 2B, iv) were estimated from normalized RPPA signal intensities (Fig. 2B, i) via a transformation function (Fig. 2B, iii) determined from a reference distribution (Fig. 2B, ii). We previously determined these reference distributions in fresh tumor tissues from a colorectal cancer cohort, with analogous demographic and clinicopathologic characteristics, using quantitative Western blotting (14). Figure 2B (i–iv) presents the normalized RPPA signal intensities, reference distributions, results from the alignments, and final estimated concentrations for all proteins from patient tumor samples included in the discovery cohort.

Tumor concentrations of Procaspase-9, XIAP, SMAC, and Procaspase-3 were input into the APOPTO-CELL mathematical model. We calculated apoptosis execution kinetics for a duration of 300 minutes (14) and computed substrate cleavage (SC) over time as a surrogate for the efficiency of apoptosis execution. Substrate cleavage profiles demonstrated that tumor apoptosis competency was variable between patients (Fig. 2C). We observed 2 major subpopulations displaying ≥ 90% or ≤ 10% substrate cleavage (58% and 32% of patient tumors, Fig. 2D), reflecting very high and low apoptosis competency, respectively. Thus, apoptosis execution functions as a biologic all-or-nothing switch between apoptosis competency and resistance, emanating from the systems-level interplay of key proteins (18, 19).

We investigated whether substrate cleavage was associated with clinical outcome among patients with stage III colorectal cancer treated with adjuvant chemotherapy. As 25% substrate cleavage serves as an optimal decision threshold for apoptotic cell death in cancer cell lines (10, 12) and colorectal cancer patient tumors (14), patients were dichotomized as those with SC > 25% or SC ≤ 25% at 300 minutes (Fig. 2C and D). We observed statistically significant differences between OS and DFS curves when comparing these groups (log-rank P < 0.05; Fig. 2E and F). Patients with SC ≤ 25% had increased risks of relapse (HR, 2.05; 95% CI, 1.03–4.05; P = 0.04) and death (HR, 3.78; 95% CI, 1.42–10.08; P = 0.006) compared with those with SC > 25%.

We examined whether APOPTO-CELL predictions of apoptosis susceptibility would notably change when accounting for noise in the protein concentrations (Fig. 3A). For each patient, we built a normal distribution centered on the reference (unperturbed) value with a given SD for each of the input proteins (including Apaf-1). Next, for each patient, we ran 1,000 simulations with values randomly drawn from these distributions (bootstrapping; ref. 20) and computed the predictions for apoptosis competency. We then computed a robustness index (RI) as the fraction of simulations matching the observed substrate cleavage (SC > 25% vs. SC ≤ 25%) in the absence of noise. We investigated the RI for up to 30% variation (upper limit for intercellular heterogeneity observed in refs. 21, 22) and defined patient-specific predictions as robust when RI ≥ 90% (gray line in Fig. 3B). When increasing the SD from 10% to 30%, we observed a marginal decrease (87% to 71%) in the number of patients consistently classified as low- versus high-risk. We previously found the relative standard variation in intratumor protein amounts in a set of patients with colorectal cancer to be approximately 11% (14). Thus, in this study, we selected 10% as the coefficient of variation when assessing robustness in downstream analyses. Visualizing the distribution of substrate cleavage simulations revealed that despite adding variation, apoptosis competency remained stable in regions of very low or high substrate cleavage for most patients (Fig. 3C). Next, we sought to examine the clinical implications of the robustness analyses. We categorized patients into 3 subpopulations (“bootstrap robustness group”): (i) robust low-risk if at least 90% of the simulations classified the patient as low-risk, (ii) robust high-risk if at least 90% of the simulations classified the patients as high-risk, and (iii) non-robust if otherwise. We observed statistically significant differences in DFS (P = 0.01) and OS (P = 0.0004) curves when comparing the robust low- versus high-risk groups (Fig. 3D and E). Only 16 patients (13%) were not classified as low- or high-risk for at least 90% of the simulations (Fig. 3F). These results suggest that while a degree of uncertainty may be present in the protein measurements (in terms of both technical measurement error and true biological variability), APOPTO-CELL still retained its prognostic value. This robustness is a distinctive advantage over classical threshold-based stratification by protein biomarkers, for which continuous readouts and cutoff values are strongly influenced by measurement noise.

Apart from their defined roles in the mitochondrial apoptosis pathway, Procaspase-9, XIAP, SMAC, and Procaspase-3 may play roles in MOMP- and apoptosome-independent cell death pathways. They may also regulate other cellular processes such as proliferation, autophagy, immune response, and differentiation (23–29). As these processes could affect patient response to therapy, we examined whether these proteins were independently associated with clinical outcome. Only Procaspase-3 (PC3) showed potential as an independent biomarker (Supplementary Fig. S1). We therefore included Procaspase-3 levels as an additional prognostic factor in our analysis. This approach compensates for apoptosome-independent Procaspase-3 activation. To explore whether PC3 could further improve prognostic accuracy, we combined PC3 quantifications with APOPTO-CELL model outputs (APOPTO-CELL-PC3; Fig. 4A). We categorized patients into 3 groups: (i) high-risk, patients with low PC3 expression (≤ median) and apoptosis resistance (SC ≤ 25%); (ii) low-risk, patients with high PC3 expression (> median) and apoptosis competency (SC > 25%), and (iii) medium-risk, all other patients. We observed statistically significant differences between OS and DFS curves when comparing these groups (log-rank P < 0.01; Fig. 4B and C), which was driven by differences between the high- and low-risk groups (pairwise log-rank P < 0.01). Patients categorized as high-risk by the APOPTO-CELL-PC3 signature had significantly increased risks of relapse (HR, 3.90; 95% CI, 1.59–9.57; P = 0.008) and death (HR, 9.30; 95% CI, 2.06–41.98; P = 0.002) compared with patients categorized as low-risk. Harrell's concordance index suggested that the APOPTO-CELL-PC3 signature had superior prognostic discrimination compared with the APOPTO-CELL signature alone for both OS (0.73 vs. 0.67) and DFS (0.64 vs. 0.59).

We used Cox proportional hazards regression models to examine the prognostic value of APOPTO-CELL-PC3 and established clinical risk factors (Fig. 4D). T stage and lymphovascular invasion were significantly associated with the risks of relapse or death in univariate analyses. Patients categorized as high-risk by the APOPTO-CELL or APOPTO-CELL-PC3 signature were significantly overrepresented in stage T4 (χ²P = 0.03 and χ²P = 0.006, respectively). No association was detected between lymphovascular invasion and the APOPTO-CELL (χ²P = 0.58) or APOPTO-CELL-PC3 signatures (χ²P = 0.65). When adjusting for T stage and lymphovascular invasion, patients categorized as high-risk by the APOPTO-CELL-PC3 signature had increased risks of relapse (HR, 3.26; 95% CI, 1.27–8.35; P = 0.04) and death (HR, 11.10; 95% CI, 2.35–52.33; P = 0.001) compared with patients categorized as low-risk.

We explored whether staging, primary tumor location, and lymphovascular invasion further aided in patient stratification (Supplementary Fig. S2). We observed significant differences between OS and DFS curves by lymphovascular invasion (log-rank P = 0.04 for both, Supplementary Fig. S2C, i–ii), although not for staging or tumor location (Supplementary Fig. S2A and S2B, i–ii). When stratifying by stage and location, we observed significant differences between OS and DFS curves (log-rank P < 0.05) by the APOPTO-CELL-PC3 signature among stage III-A/B and proximal tumors (Supplementary Fig. S2A and S2B, iii–iv). There were differences between DFS curves by APOPTO-CELL-PC3 among patients without lymphovascular invasion (Supplementary Fig. S2C, iii) and between OS curves among patients with lymphovascular invasion (Supplementary Fig. S2C, iv).

The prognostic relevance of BRAF and MSI status has been comprehensively examined, and testing of MSI status has been introduced in clinical practice (2). We used an expansion cohort (GSE39582 dataset; ref. 16) to explore whether MSI and BRAF status potentially confound associations between our prognostic signatures and recurrence risk (Fig. 1). We analyzed stage III patients with known status for microsatellites and BRAF mutation (n = 157; Supplementary Table ST1).

APOPTO-CELL and APOPTO-CELL-PC3 signatures, estimated from gene expression (Supplementary Methods SM2), were associated with DFS in univariate and multivariate analyses (Supplementary Table ST2). In unadjusted analyses, patients categorized as high- versus low-risk by APOPTO-CELL (SC ≤ 25% vs. SC > 25%) had an increased risk of relapse (HR, 1.94; 95% CI, 1.17–3.24; P = 0.01). Patients categorized as high- and medium-risk by APOPTO-CELL-PC3 had an increased risk of relapse compared with those classified as low-risk (HR, 2.39; 95% CI, 1.20–4.78 and HR, 2.66; 95% CI, 1.32–5.34; P = 0.008). We fit 2 multivariate models (i) stratifying by MSI (dMMR vs. pMMR) status and adjusting for BRAF (wild-type vs. mutant) as well as treatment received (chemotherapy vs. no chemotherapy) and (ii) additionally adjusting for age (continuous linear), sex (male vs. female), tumor location (distal vs. proximal), and KRAS mutational status (wild-type vs. mutant). Relative risk estimates were similar between unadjusted and adjusted models, suggesting that these factors did not likely confound associations between our signatures and DFS. Furthermore, both signatures remained independent prognostic markers for DFS (Supplementary Table ST2).

We built a RF signature using the APOPTO-CELL-PC3 signature, the remaining single proteins (dichotomized as > or ≤ median), and the clinical variables analyzed in Fig. 4D (“RF-all predictors”). A reduced RF classifier was constructed using the most influential predictors of relapse (“RF-reduced predictors set”). The reduced classifier predicted disease recurrence without any loss in accuracy compared to the full RF classifier (McNemara P > 0.05). Thus, we performed our downstream analysis with the reduced classifier, which included the APOPTO-CELL-PC3 signature, T stage, sex, N stage, age, and nodal count (Fig. 5B). The AUC (see Materials and Methods) suggested this model satisfactorily predicted relapse (0.73; 95% CI, 0.62–0.74; Fig. 5C).

RF classifiers are considered “black box” machine-learning algorithms that have limited interpretability. To investigate how the recurrence probability predicted by the RF classifier depended on each feature and their interactions (Supplementary Fig. S3), we generated a synthetic cohort where each patient was characterized by a unique combination of input features (see Materials and Methods). The RF classifier predicted a reduced likelihood of recurrence for patients categorized as low-risk by the APOPTO-CELL-PC3 signature, aged > 40, with a nodal count between 10 and 50, males, and those with less advanced T and N stage (Supplementary Fig. S3A–S3F).

We examined how the RF features interact to deploy the final recurrence predictions. For each combination of predictors, we visualized recurrence risk relative to the overall average recurrence risk (the reference baseline; Fig. 5D). The relative risks associated with age and nodal count were aggregated into 3 levels by averaging over all risks in the corresponding group. Supplementary Figure S3G presents this visualization without aggregation of these predictors. Overall, patients categorized as low-risk by the APOPTO-CELL-PC3 signature had a reduced recurrence risk. This effect was lost in advanced cancers (T4 and N2), which exhibited an overall higher probability of recurrence. Females also had a higher risk of relapse than males.

We categorized patients into 3 risk groups according to the probability of recurrence predicted by the reduced RF classifier: < 20%, 20%–50%, and > 50%. There were significant differences between DFS curves by these groups (log-rank P < 0.001; Fig. 5F). Patients with > 50% risk had an approximately 5-fold increased recurrence risk (HR, 5.13; 95% CI, 2.12–12.41; P = 0.002) compared with patients with < 20% risk. Results were similar after adjusting for T stage and lymphovascular invasion. Harrell's concordance index suggested that the RF signature was better at discriminating recurrence than the APOPTO-CELL-PC3 signature (0.67 vs. 0.64).

We evaluated the prognostic value of our signatures using an independent publically available cohort: TCGA COAD (Fig. 1). Supplementary Table ST1 presents clinicopathologic and demographic characteristics of patients who met our inclusion criteria (Supplementary Methods SM1).

The 3 signatures identified in the discovery cohort were tested on stage III patients in the validation cohort. In this dataset, protein expression of SMAC and XIAP were available. For the amounts of Procaspase-3 and Procaspase-9, transcript abundance was used as surrogate for protein expression (Supplementary Methods SM2). Despite the limited sample size and follow-up time, we nevertheless observed statistically significant differences between DFS curves when comparing risk groups for the APOPTO-CELL, APOPTO-CELL-PC3, and RF signatures (log-rank P < 0.05; Fig. 6A–C).

A recent study identified 4 distinct consensus molecular subtypes (CMS) for colorectal cancer (30), with CMS4 (“mesenchymal”) versus CMS1–3 subtypes being associated with poorer outcome. Among our validation cohort, there were no differences between DFS curves by CMS4 versus 1–3 subtypes among stage III (log-rank P = 0.90) or stage I–IV patients (log-rank P = 0.54). Nevertheless, substrate cleavage retained its prognostic value among CMS1–3 (log-rank P = 0.02), although not among CMS4 (log-rank P = 0.11; Fig. 6D). Analogously, the APOPTO-CELL-PC3 and RF signatures had prognostic value among CMS1–3, but not CMS4 (Fig. 6E and F).

We evaluated APOPTO-CELL, a mathematical model that calculates apoptosis competency of patient tumors, as a signature of stage III colorectal cancer patient outcome to 5-FU–based chemotherapy. Moreover, we present a novel systems medicine workflow to optimize model predictions. This study is the first large-scale demonstration that determining apoptosis competency by systems modeling adds independent prognostic value to current clinicopathologic markers.

Previous studies demonstrated that systems biological approaches toward apoptosis signaling may have great potential as novel avenues toward discovery of biomarkers (31) and therapeutic viable targets (e.g., SMAC-mimetics; ref. 32) resulting in an educated stratification of patients for clinical trials (33). Further development and exploitation of such new strategies is warranted, as despite extensive efforts to identify prognostic and predictive biomarkers for colorectal cancer TNM staging remains the most reliable prognostic marker that still drives treatment decisions. Other clinicopathologic tumor characteristics, such as tumor location, differentiation, and lymphovascular invasion, have been associated with prognosis and response to treatment (2). MSI and wild-type status for KRAS, BRAF, PIK3CA, and TP53 have been associated with better outcome (2, 34, 35). Conversely, patients with wild-type APC showed adverse survival (35). Recent studies have highlighted the intertwined relationship of MSI and somatic mutations status (35–39). MSI/BRAF^wild-type and MSS/BRAF^mutant patients exhibited the most favorable and unfavorable outcome, respectively (36–39). Remarkably, MSI/BRAF^mutant showed superior prognosis compared with MSS/BRAF^wild-type, suggesting that microsatellite status prevails over BRAF mutation in steering patients survival in these subgroups (36–39). Mutations in TP53, KRAS, and APC have been more frequently observed in microsatellite-stable (MSS) patients with the most adverse outcome being observed in the MSS triple-mutant subpopulation (35). Remarkable advances in “-omics” technologies have brought about a shift in prognostic marker discovery, moving from conventional clinicopathologic risk factors toward their integration with molecular characteristics. Multigene-based classification algorithms have been shown to aid in stratifying patient outcome, but these classifiers so far provide only small improvements in stratification compared with the use of clinical markers alone (40). Reviewing and re-assessing promising multigene and multiprotein panels using systems biological approaches may allow us to further improve the prognostic value of such signatures.

We demonstrated that patient-specific predictions by the APOPTO-CELL signature were highly robust against noise in measurements and parameterization. This robustness emanates from the multiprotein interplay during the apoptosis execution phase, which includes several positive feedback loops. If sufficiently stimulated, these bring about an irreversible cell death decision with a pronounced binary/switch-like character (9, 10, 18). As our discovery cohort was composed of patients recruited at 3 different centers in three European countries, this high robustness may be fundamental for successful model-based patient stratification. Our approach of apoptosis modeling was successful in the expansion and validation cohort, which made use of FF materials and transcript abundance in colorectal cancer tumor samples collected at various centers across the France and United States, respectively.

We introduced the APOPTO-CELL-PC3 signature which capitalizes on both the prognostic information embedded in the apoptosome-dependent cell death pathways modeled by APOPTO-CELL and Procaspase-3 paracrine signaling to further improve prognostic accuracy. Procaspase-3 exerts important roles outside of MOMP-dependent apoptosis execution including mediating tissue regeneration and immune responses (23–29). Upon Caspase-3 activation, apoptotic cells release growth signals such as prostaglandin E2 that ultimately induce tissue regeneration and wound healing in the neighbor cells via the “Phoenix Rising” pathway (25). Furthermore, antitumor immune responses trigger apoptosis cascades through Procaspase-8 activation or granzyme B activation, that can drive Procaspase-3 activation independent of the mitochondrial pathway (23–25, 29).

There is extensive genomic, epigenomic, and molecular interpatient heterogeneity in colorectal cancer (41). A key step toward personalized precision oncology is to develop panels of combinatorial biomarkers, as opposed to single biomarkers, that can better capture disease complexity and provide individualized predictions on cancer progression and treatment response. To develop optimal stratification tools, it is paramount to integrate clinical patient-specific characteristics with molecular phenotypes or, as in our case, patient-specific systems-level characteristics. By applying RF machine learning, we investigated how the APOPTO-CELL-PC3 signature could be combined with conventional clinical variables currently used for colorectal cancer prognosis. The RF approach identified APOPTO-CELL-PC3 as the most impactful predictor, and a reduced RF classifier allowed us to optimally stratify patients into risk groups from a limited number of variables. Reducing the number of clinicopathologic variables and molecular markers required for optimal and individualized prognostication minimizes the economic burden of otherwise possibly excessively complex assays and data acquisition. We generated a valuable visualization of how the predictors contribute to shifting recurrence risk, which could serve as a nomogram to aid medical practitioners in planning disease management.

Extensive RNA profiling work (16, 42–45) recently resulted in the definition of 4 major subtypes for colorectal cancer (CMS1–4; ref. 30). Compared with patients with CMS1–3, CMS4 patients had higher risks of relapse and mortality. Analysis of the interplay between apoptosis competency and CMS revealed a potential differential role of apoptosis signaling in CMS1–3 compared with CMS4, with apoptosis competency associated with good prognosis in CMS1–3. The salient features of the CMS4 subtype (“mesenchymal”) include CIN/MSI alterations, NOTCH3 overexpression, TGFβ, and VEGF activation (30). TGFβ and VEGF signaling has been reported to suppress Caspase-3-dependent apoptosis execution in various cell types (46, 47). Conversely, caspases have been shown to be enriched in the CMS1 (“immune”) subtype.

Previous proof-of-concept studies on apoptosis modeling in colorectal cancer have been limited to the use of FF materials for the generation of tissue protein extracts (14). Our study shows that formalin-fixed, paraffin-embedded (FFPE)-derived protein extracts are suitable for calculating apoptosis competency, representing a key advancement toward aligning with routine surgical and pathologic workflows. It also supports independent retrospective validation in larger biobanked collections of stage III colorectal cancer cohorts. While RPPA-based protein quantification pipelines may not be amenable for integration into standard clinical workflows, the development of calibrated protein quantification assays, such as multiplex ELISAs or novel quantitative digital histopathology techniques, may address current limitations in routine quantitative clinical proteomics (48).

Recent research has pinpointed the pivotal role played by the tumor microenvironment in regulating response to chemotherapeutics (49). Factoring in relative abundance of key apoptotic proteins in both the tumor and the surrounding tissue (stroma, immune cells, etc.) will further fine-tune risk scores. Emerging technologies, such as multiplexed fluorescence microscopy (MxIF; ref. 50) now allow simultaneous co-staining of multiple proteins and reliable quantitative measurements within different compartments which will be instrumental in future studies to tackle these challenges.

P.G. Johnston is a consultant/advisory board member for Chugai. P. Laurent-Puig is a consultant/advisory board member for Amgen, Boerhinger-Ingelheim, Integragen, Merck Serono, Pfizer, Roche, and Sanofi. No potential conflicts of interest were disclosed by the other authors.

Conception and design: M. Salvucci, M. Lawler, P.G. Johnston, D.A. McNamara, P. Laurent-Puig, M. Rehm, J.H.M. Prehn

Development of methodology: M. Salvucci, M.L. Würstle, P. Laurent-Puig, M. Rehm, J.H.M. Prehn

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M. Salvucci, M.L. Würstle, C. Morgan, M. Cremona, O. Bacon, R. O'Bryne, L. Flanagan, S. Dasgupta, N. Rice, C. Pilati, S. Toomey, R. Wilson, S. Camilleri-Broët, M. Salto-Tellez, D.A. McNamara, E.W. Kay, P. Laurent-Puig, D.B. Longley

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M. Salvucci, M. Cremona, A.J. Resler, N. Rice, E. Zink, L.M. Schöller, P.G. Johnston, E.W. Kay, P. Laurent-Puig, B.T. Hennessy, M. Rehm, J.H.M. Prehn

Writing, review, and/or revision of the manuscript: M. Salvucci, M.L. Würstle, A.U. Lindner, A.J. Resler, M. Lawler, P.G. Johnston, R. Wilson, M. Salto-Tellez, D.A. McNamara, E.W. Kay, P. Laurent-Puig, S. Van Schaeybroeck, B.T. Hennessy, D.B. Longley, M. Rehm, J.H.M. Prehn

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M. Salvucci, M.L. Würstle, S. Curry, O. Bacon, Á.C. Murphy, L. Flanagan, S. Dasgupta, N. Rice, D.A. McNamara, S. Van Schaeybroeck

Study supervision: M. Rehm, J.H.M. Prehn

The authors acknowledge support for their work by the European Union Framework Programme 7 (FP7 APO-DECIDE) under contract no. 306021. J.H.M. Prehn is supported by a Science Foundation Ireland Investigator Award (13/IA/1881) and by the Irish Cancer Society BreastPredict Collaborative Research Centre (CCRC13GAL). J.H.M. Prehn, M. Rehm, and D.B. Longley are supported by a Science Foundation Ireland/Department of Enterprise and Learning Partnership Award (14/IA/2582). E. Zink was supported by a Health Research Board Translational Research Supplementary Award (TRA/2007/26). L.M. Schöller was supported by the European Community Action Scheme for the Mobility of University Students (ERASMUS).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.