Development of a protein signature to enable clinical positioning of IAP inhibitors in colorectal cancer

Resistance to chemotherapy‐induced cell death is a major barrier to effective treatment of solid tumours such as colorectal cancer, CRC. Herein, we present a study aimed at developing a proteomics‐based predictor of response to standard‐of‐care (SoC) chemotherapy in combination with antagonists of IAPs (inhibitors of apoptosis proteins), which have been implicated as mediators of drug resistance in CRC. We quantified the absolute expression of 19 key apoptotic proteins at baseline in a panel of 12 CRC cell lines representative of the genetic diversity seen in this disease to identify which proteins promote resistance or sensitivity to a model IAP antagonist [birinapant (Bir)] alone and in combination with SoC chemotherapy (5FU plus oxaliplatin). Quantitative western blotting demonstrated heterogeneous expression of IAP interactome proteins across the CRC cell line panel, and cell death analyses revealed a widely varied response to Bir/chemotherapy combinations. Baseline protein expression of cIAP1, caspase‐8 and RIPK1 expression robustly correlated with response to Bir/chemotherapy combinations. Classifying cell lines into ‘responsive’, ‘intermediate’ and ‘resistant’ groups and using linear discriminant analysis (LDA) enabled the identification of a 12‐protein signature that separated responders to Bir/chemotherapy combinations in the CRC cell line panel with 100% accuracy. Moreover, the LDA model was able to predict response accurately when cells were cocultured with Tumour necrosis factor‐alpha to mimic a pro‐inflammatory tumour microenvironment. Thus, our study provides the starting point for a proteomics‐based companion diagnostic that predicts response to IAP antagonist/SoC chemotherapy combinations in CRC.

Resistance to chemotherapy-induced cell death is a major barrier to effective treatment of solid tumours such as colorectal cancer, CRC. Herein, we present a study aimed at developing a proteomics-based predictor of response to standard-of-care (SoC) chemotherapy in combination with antagonists of IAPs (inhibitors of apoptosis proteins), which have been implicated as mediators of drug resistance in CRC. We quantified the absolute expression of 19 key apoptotic proteins at baseline in a panel of 12 CRC cell lines representative of the genetic diversity seen in this disease to identify which proteins promote resistance or sensitivity to a model IAP antagonist [birinapant (Bir)] alone and in combination with SoC chemotherapy (5FU plus oxaliplatin). Quantitative western blotting demonstrated heterogeneous expression of IAP interactome proteins across the CRC cell line panel, and cell death analyses revealed a widely varied response to Bir/chemotherapy combinations. Baseline protein expression of cIAP1, caspase-8 and RIPK1 expression robustly correlated with response to Bir/chemotherapy combinations. Classifying cell lines into 'responsive', 'intermediate' and 'resistant' groups and using linear discriminant analysis (LDA) enabled the identification of a 12-protein signature that separated responders to Bir/chemotherapy combinations in the CRC cell line panel with 100% accuracy. Moreover, the LDA model was able to predict response accurately when cells were cocultured with Tumour necrosis factor-alpha to mimic a pro-inflammatory tumour microenvironment. Thus, our study provides the starting point for a proteomics-based companion diagnostic that predicts response to IAP antagonist/SoC chemotherapy combinations in CRC.

Introduction
Resistance to apoptosis is a classical hallmark of cancer, which can manifest clinically as lack of response to chemotherapeutic agents such as 5-Fluorouracil (5FU) and oxaliplatin [1]. Cancer cells can acquire such resistance through increasing their expression of (and therefore dependence on) anti-apoptotic proteins such as inhibitor of apoptosis proteins (IAPs), making such proteins attractive therapeutic targets [2]. IAPs exert their anti-apoptotic function via two distinct mechanisms [3]. Firstly, cIAP1 and cIAP2 can divert typically death-inducing Tumour necrosis factor-alpha (TNF-α) signals to prosurvival signalling through the ubiquitination of RIPK1, leading to the activation of Nuclear factor kappa-B transcriptional programmes that enhance cell proliferation, propagate further inflammatory signalling via transcriptional upregulation of cytokines including TNF-α itself and drive the transcription of anti-apoptotic proteins, including the caspase-8 regulator FADD-like IL-1β-converting enzyme inhibitory protein (FLIP) and the IAPs themselves [4,5]. Secondly, IAPs [specifically X-Linked inhibitor of apoptosis protein (XIAP)] can directly inhibit the complete processing and subsequent activation of caspase-3/7 and caspase-9, thereby abrogating the cell's ability to execute apoptosis [6,7].
IAP antagonists interact with IAPs through binding to their BIR domains and promote apoptosis induced by extracellular death ligands such as TNF-α by inhibiting cIAP-induced RIPK1 ubiquitination [3,8]; this leads to the formation of an intracellular death signalling complex termed complex II in which RIPK1 recruits Fas-associated death domain (FADD), which in turn recruits procaspase-8 and FLIP. Depending on the relative amounts of procaspase-8 and FLIP in this complex, procaspase-8 can form homodimers, which drive apoptosis [9,10]. In the absence of procaspase-8, this complex can instead activate necroptosis [11], an alternative form of programmed cell death in which RIPK1 recruits and activates RIPK3 eventually leading to phosphorylation of Mixed lineage kinase domain-like pseudokinase (MLKL); phosphorylated MLKL oligomerises and forms pores in the plasma membrane that result in pro-inflammatory cell lysis [12]. IAP antagonists also inhibit XIAP facilitating apoptosis mediated by the intrinsic mitochondrial apoptotic pathway and the execution phase of apoptosis mediated by caspase-3/7 [13,14].
The clinical realisation of the potential of IAP antagonists has been hampered by the lack of predictive biomarkers of response to enable patient stratification. Using a large panel of colorectal cancer (CRC) cell lines representative of the clinical diversity seen in CRC, herein we have developed a strategy to predict responses to IAP antagonist-based therapy by quantifying the expression of key apoptotic proteins.

Results and Discussion
Response of CRC cell lines to IAP antagonistbased therapy We assembled a panel of 12 CRC cell lines with a range of genotypes and microsatellite status ( Table 1) that were co-treated with a model IAP antagonist birinapant (Bir, TL37211) alone and in combination with 5FU and oxaliplatin ('OX5FU', to mimic the clinical FOLFOX regimen). In addition, we included treatment arms in which a pro-inflammatory tumour microenvironment was modelled using recombinant TNF-α. Cell death was quantified after treatment for 24, 48 or 72 h; the average cell death was also calculated over the three time points (Fig. 1).
Overall, the cell lines were resistant to Bir alone, with GP5D demonstrating the greatest apoptotic response of just 25% averaged across the three time points (Fig. 1). Although TNF-α alone induced nonsignificant levels of cell death, its addition to Bir significantly increased apoptosis; again, GP5D cells were most sensitive to Bir/ TNF-α with~60% apoptosis at 24 h and~80% at 48 h. A range of responses were observed in the models, with only the COLO320 model completely resistant to Bir/ TNF-α at all time points.
Compared to Bir/TNF-α, OX5FU-induced apoptosis had slower kinetics, with the greatest degree of cell death observed at 72 h; however, over half of the models exhibited a high level of resistance to this treatment (which mimics the clinical standard-of-care regimen), with < 10% cell death even at 72 h. It was notable therefore that the apoptotic response to chemotherapy of a number of models was significantly enhanced by the co-treatment with Bir, including LoVo, RKO and HCT116 (Fig. 1). In the context of TNF-α, there were further increases in cell death observed in cells co-treated with Bir and OX5FU in several models, including SW620, p53 null HCT116, HT29 and LIM1215.      Fig. 2A. *RIPK3 is absent from HeLa; therefore, CRC cell line absolute RIPK3 abundance was calculated using HT29 as a reference. Western blot images and densitometry are representative of three independent experiments, and quantification represents the mean AE SEM of three independent experiments.

Apoptotic protein expression profiling
In order to better understand the heterogeneity of response to IAP inhibitors in CRC, the absolute expression of a panel of key IAP interactome and other major apoptotic regulators ( Fig. 2A) was determined by quantitative western blotting. Absolute protein levels were quantified by initially determining the concentrations of these proteins within a reference cell line (HeLa or HT29) by comparison to recombinant proteins as described previously [16,18]. We generated a concentration range of recombinant proteins against which could be used to calculate the concentration of target proteins in the reference cell line (Fig. 2B). Protein concentrations in each CRC model were then determined by western blotting, with comparison of each protein in each model against the protein levels in the reference cell line (Fig. 2C). This allowed us to obtain absolute concentrations for each protein in each CRC model (Table 2). Not surprisingly given the heterogeneity of CRC and the heterogeneity of responses to Bir, the levels of key proteins of the extrinsic apoptotic pathway (procaspase-8, FLIP(L), FLIP(S), FADD, RIP1, RIP3, Cellular inhibitor of apoptosis protein 1/2, XIAP) were variable across the panel of cell lines (Fig. 2B, Table 2). cIAP1 was present in all models and expressed at a consistently higher level than cIAP2 and XIAP. RIPK1 was the most highly expressed protein in the panel, whereas RIPK3 was expressed at a much lower level overall and was absent or almost completely absent in five models.
The key adaptor protein for complex II, FADD, was similarly expressed across the cell line panel. Notably, the key apoptotic effector of the extrinsic pathway caspase-8 was expressed at a relatively high level in all models. The long splice form (FLIP(L)) of the canonical regulator of caspase-8, FLIP, was expressed at a relatively low level in comparison with procaspase-8, and the short splice form FLIP(S) was absent in several models; this is consistent with the findings of us and others that FLIP expression can be > 100-fold lower than procaspase-8, although it can still compete effectively with procaspase-8 for recruitment to effector complexes such as the DISC [26].
We also used previously established [16,18] concentrations in HeLa cells for B-cell lymphoma 2 (Bcl-2) family proteins, procaspase-3 and procaspase-9, APAF1 and Second mitochondria-derived activator of caspases (SMAC) to determine the levels of these key regulators of mitochondrial apoptosis in the CRC panel ( Fig. 2A,B, Table 2). The Bcl-2 family regulate mitochondrial outer membrane permeabilisation (MOMP), a critical commitment point in apoptosis induction [27]. For the Bcl-2 family of proteins, it was notable that BID, which mediates cross-talk between the extrinsic and intrinsic mitochondrial pathways [28], was expressed at a consistent, relatively high level across the panel. The key anti-apoptotic proteins Bcl-XL and Myeloid cell leukaemia differentiation protein 1 (Mcl-1) were expressed in all models albeit to varying degrees, whereas Bcl-2 itself was undetectable in half the models. While one of the requisite effectors for initiating MOMP, BCL2 antagonist/killer (BAK), was expressed in all models, BCL2-associated X-protein (BAX) was very low or absent in several models. A key component of the apoptosome, Apoptosis-associated factor-1 (APAF-1), was expressed variably across the panel, whilst procaspase-9, with which it forms the apoptosis-inducing apoptosome downstream of cytochrome c release from the mitochondria, was expressed in all models, albeit at a relatively low level compared with APAF-1, suggesting that procaspase-9 levels are likely the limiting factor for apoptosome formation. SMAC, like cytochrome c, is a key regulator of apoptosis downstream of MOMP that once released from mitochondria binds to and inhibits or promotes the degradation of IAPs [29]. Indeed, first-generation IAP antagonists such as Bir are based on the critical N-terminal domain of SMAC and are sometimes called SMAC mimetics [14,30]. SMAC was expressed in every model in the panel. Finally, the proform of the key executioner caspase, caspase-3, was expressed at a similar level in all models, suggestive of comparable potential to engage the executioner phase of apoptosis across the panel. p53 is a key regulator of apoptosis signalling genes [31]. We therefore compared the expression of each protein in the isogenic p53 wild-type (WT) and Null HCT116 models (Fig. 3A). Overall, the majority of proteins were constitutively expressed at a similar level in the presence and absence of p53. The canonical p53 target BAX was indeed higher in the p53 WT model. FADD and RIPK1 were also higher in the p53 WT setting, and consistent with our recent findings, FLIP was expressed more highly at the protein level in p53 WT HCT116 cells [32] as was XIAP. Moreover, hierarchical clustering did not reveal any differential patterns of protein expression between mutant and WT p53 models (Fig. 3B), suggesting that p53 status is not a major determinant of baseline apoptotic potential in CRC.

Correlation between IAP antagonist response and apoptotic protein expression
We next assessed whether the level of any of the individual proteins correlated with apoptosis induction. Linear correlation was determined using Pearson's correlation. In addition, Spearman's correlation was considered for the discovery of other monotonic dependencies and for the exclusion of outlier-driven linear correlation cases (Fig. 4). None of the IAP concentrations were correlated with Bir+TNF-α treatment; therefore, IAP levels alone were not separating responders to this combination. However, absolute levels of procaspase-8 positively correlated with response to Euclidean distance clustering of protein abundance Relative protein abundance Bir+TNF-α (Fig. 4A), and this effect was robust across all three time points (Fig. 4B). cIAP1 expression was found to be significantly correlated with response to Bir+OX5FU, as was response to OX5FU+TNF-α; this suggests that sensitivity to chemotherapy is related to cIAP1 expression, and indeed, there was a nonsignificant trend for cIAP1 expression and response to OX5FU (Pearson's r = 0.52, P = 0.08). Moreover, the increase in cell death between Bir+OX5FU and OX5FU treatments (Fig. 5A) or between Bir+OX5FU and Bir treatments (Fig. 5C) also showed a linear dependence on cIAP1 levels; however, this relationship was not found to be significant in Spearman's correlation test (Fig. 5A,C) and was lost following the addition of TNF-α (Fig. 5B).
Notably, RIPK1 was found to be a more prominent indicator of the Bir+OX5FU response. Its level was significantly correlated with Bir+OX5FU response using both Pearson's and Spearman's tests (Fig. 4A), especially after 24 h (Fig. 4B). In addition, RIPK1 was also found to be correlated with the increase in cell death response between Bir+OX5FU and OX5FU treatments (Fig. 5A).

Minimal apoptotic protein profile that separates treatment responses by means of linear discriminant models
From our protein quantification experiments, 19 proteins that characterise the extrinsic and intrinsic apoptotic pathways showed a wide range of variance and amplitude over the 12 cell lines (Fig. 2B, Table 2). To identify proteins and minimal characteristic protein profiles that separated responses, we performed a principal component analysis (PCA) (Fig. 6A-C). PCA discovered 11 component axes (principal components, PCs) that transformed the 19 proteins set into a system of coordinates that optimally captured the protein expression variance of our cell line profiles (Fig. 6A). Based on PC eigenvalues, we have found that first seven PCs cumulatively captured 91% of the variation we measured in the protein profiles. Thus, these seven PCs were kept as essential components for the subsequent analysis. Interestingly, we noted that the first PC which explained 20% of the entire variance in the protein profiles showed a very strong dependence on the core components of the Ripoptosome complex: FADD, caspase-8 and RIPK1 (Fig. 6B). It is known that downstream of IAP inhibitor treatment, the cytosolic Ripoptosome is the signalling platform required for the activation of initiator caspase, procaspase-8, thus triggering the extrinsic apoptosis pathway response [33][34][35]. Moreover, Ripoptosome assembly on its own contributes to a high variability in the cell death response as shown in experimental and mathematical modelling studies [36,37]. Thus, our results suggested that the capacity to initiate extrinsic apoptosis through the Ripoptosome might be a prominent process that is affected by protein expression differences between our cell lines. To identify a minimal protein set, the contribution of each protein was averaged over the first seven essential PCs (Fig. 6C). From this, we could down select 12 proteins that are involved in the extrinsic (FADD, caspase-8, RIPK1, RIPK3, FLIP(S), BID) and in the intrinsic (XIAP, caspase-9, caspase-3, BCL-XL, BAK, SMAC) apoptosis pathways. To find the relationship between these proteins and cell death responses, we categorised cell lines into 3 response groups according to treatment-specific cell death averaged over 3 time points (Table 3). Groups 'resistant', 'intermediate' and 'responsive' were assigned for cell death response < 10%, 10-30% and > 30%, respectively. Thus, the specific cell death responses were discretised for further comparison with the minimal protein set by means of linear discriminant analysis (LDA) [38][39][40][41]. LDA provides a linear classification model that transforms the minimal protein set into a new subspace of component axes that maximises the separation of the response groups. These axes are called linear discriminant functions (LDs) and were generated for each treatment separately (Fig. 6D). An important property of LDA is that its performance is independent of the protein data standardisation because the algorithm relies on variation between response groups only. Therefore, LDs are the linear functions of protein concentrations directly and, therefore, can be used for response classification in any data set expressed in concentration terms. The scatter plots showed clear separation between all response groups in the optimal LD spaces calculated for OX5FU, Bir+OX5FU and Bir+TNF-α treatments ( Fig. 6D) with 100% accuracy in response classification for all 12 cell lines (Fig. 6E).
Descriptive accuracy of the LDA models was achieved for most of our treatments with the minimal set of 12 proteins. Only one classification error occurred for the Bir+OX5FU+TNF-α treatment in the classification of HT29 cells. In this case, the LDA model underestimated the HT29 response, suggesting for it an intermediate rather than a responsive group (Fig. 6D,  E). This inaccuracy was found to be associated with the overall weaker separation between response groups for the triple-treatment combination LDA model. This was likely due to the underrepresentation of intermediate and resistant response groups (2 and 1 cell lines are representing these two groups, respectively) or an influence of other key signalling pathways not covered in the present study. Nevertheless, cell lines that are resistant to Bir+OX5FU and, at the same time, responsive to Bir+TNF-α all show a high level of response to Bir+-OX5FU+TNF-α. Additionally, all cell lines responsive to Bir+OX5FU were also responsive to Bir+OX5-FU+TNF-α. Therefore, the Bir+OX5FU+TNF-α LDA model (Fig. 6D), which did not separate well, is dispensable, and the triple-treatment responsive cell lines can be identified from the use of the Bir+OX5FU and Bir+TNF-α LDA models.

Clinical Positioning of IAP inhibitors in CRC
Overall, our analyses suggest that benefit from the treatment with IAP inhibitors in CRC can be calculated from the use of a panel of 12 proteins plus knowledge of whether the tumour is inflamed (indicating the presence of TNF-α). As depicted in Fig. 7, if the protein profile indicates that a tumour is in the 'Bir+OX5FU responsive' group, the patient would be predicted to benefit from treatment with an IAP inhibitor alongside FOLFOX. If not, consideration would be given to whether the patient's tumour expresses TNF-α. If inflammation is present and the patient's tumour falls into the 'Bir+TNF-α responsive' group, the patient would be predicted to benefit from treatment with an IAP inhibitor. If not, the benefits of the combinational treatment with an IAP antagonist would be predicted to be minimal. To make the above approach clinically feasible, simple protein quantification approaches are needed. In this regard, huge advances in multiplex immunohistochemistry have been made. For example, a multiplexed fluorescence microscopy method for the quantification of multiple proteins from formalin-fixed paraffin-embedded tissues was recently described that can quantify up to 60 proteins in single 5-µm sections [42,43]. The ability to do this analysis in single sections means problems of changing cellularity associated with sequential sectioning needed for traditional IHC (one section per protein) are overcome. This and similar methodologies can also identify the presence of specific immune cell populations in tumours and hence whether or not a tissue is inflamed. Conversion of protein intensities into absolute concentrations would require the incorporation of standards into the analysis; this could be achieved using formalin-fixed cell lines with known protein concentrations embedded alongside the tumour sample. Thus, these emerging approaches could be employed to convert experimentally derived proteomics-based predictive algorithms such as we have developed here into realworld companion diagnostics.

CRC cell line panel
Cell lines were selected as a representation of the molecular genetic diversity seen in CRC based on the mutation status of known driver genes (including TP53, KRAS, BRAF and PIK3CA), transcriptional profiling and microsatellite stability status, as published in Medico et al. [15]. HeLa, DLD1, LS174T, LoVo, HT29, GP5D, LIM1215, HCT116, RKO, SW620, COLO320 and LS513 cell lines were purchased as authenticated stocks from ATCC (Teddington, UK). HCT116 p53 −/− was gifted from B. Vogelstein (Johns Hopkins University, USA). Immediately after purchase, early passage stocks of each cell line were frozen down. After thawing, cells were kept for a limited number of passages and were regularly screened for the presence of mycoplasma using the MycoAlert Mycoplasma Detection Kit (Lonza, Basel, Switzerland).

Cell death assay
Sensitivity of CRC cell lines to each treatment was determined by Annexin V/propidium iodide high-content microscopy. Cells were seeded at appropriate densities and treated for 24, 48 or 72 h with TL32711 (Bir; Selleck Chemicals, Munich, Germany) alone, chemotherapy alone ['OX5FU':

Patient (A) 12 proteins profile
Bir + OX5FU will work Variable Response  μM 5FU (Medac, Chobham, Surrey, UK) plus 2 μM oxaliplatin (Accord, Barnstaple, Devon, UK)], or a combination of the two in the presence or absence of 10 ngÁmL −1 TNF-α (Alomone Labs, Jerusalem, Israel). Cells were stained with propidium iodide (Sigma, UK), FITC-labelled Annexin V (Thermo Fisher, Waltham, MA, USA) and Hoechst 33342 to visualise nuclei (Thermo Fisher). Cell death was quantified (AV+and/or PI+) on an Array Scan XTI high-content microscope (Thermo Scientific). To control for variabilities in basal cell death across the CRC panel, treatment-specific cell death was calculated as follows: {cell death in treatment group} -{cell death in control group}.

Statistics
Pairwise comparisons of the mean treatment responses with the Wilcoxon test were performed using ggpubr v0.3.0 package [19] together with the ggplot2 v3.3.0 package [20] in the R environment [21]. Spearman's and Pearson's correlation coefficients were calculated to identify significant correlations between cell line sensitivities (treatment-specific cell death) and absolute concentrations of each IAP interactome protein. Correlation and hierarchical clustering heat maps were produced with the R package clusterProfiler v3.12.0 [22]. PCA was applied for dimensionality reduction in the characteristic protein profile in the cohort of 12 colon cancer cell lines (Fig. 6A). For the PCA, we standardised protein concentrations (z-scores of 19 proteins). We have calculated the contribution of each protein on the given PC (Fig. 6B) from the correlation coefficients between a protein and PC termed loadings of variables as per the following: Contribution i, j ð Þ¼ loading i, j ð Þ ð Þ 2 Â 100% ∑ 19 j¼1 loading i, j ð Þ ð Þ 2 , where 1 ≤ i ≤ 11 is the indexing of PCs, and 1 ≤ j ≤ 19 is the indexing of proteins [23].
Further, the contribution of each protein over first seven PCs (Fig. 6C) was calculated as follows: Contr ðjÞ ¼ , where λðiÞ is the eigenvalue of PC i . With the assumption of average uniform expected contribution for all 19 proteins, the significance threshold was set to 5.26%. PCA algorithm was implemented using MATLAB 2017b [24]. LDA was then performed on the minimised protein profile (12 proteins), using absolute protein concentrations, and categorical response groups assigned based on specific cell death averaged over three time points (24, 48 and 72 h). LDA runs were performed for each treatment independently with R package MASS v7.3-51.5 [25].