PI3K/Akt/mTOR Signaling and Plasma Membrane Proteins Are Implicated in Responsiveness to Adjuvant Dendritic Cell Vaccination for Metastatic Colorectal Cancer

Colorectal cancer is the fourth leading cause of cancer death in the world (1). About one in three patients diagnosed with colorectal cancer dies from metastatic disease (2). The majority of metastases develop in the liver and lung. Although metastases can be completely resected in many patients, 60% to 80% of these patients eventually die due to the growth of small metastases that were undetectable at the time of surgery (3, 4). Currently, the addition of adjuvant chemotherapy has limited impact on patient survival (5, 6). The pressing need for better treatments and growing recognition of the immune system's role in cancer progression have prompted tremendous efforts to develop immunotherapies for colorectal cancer (7).

Immunotherapies aim to either enhance the antitumor immune response or prevent immunosurveillance suppression by cancer cells. Whole tumor vaccines for colorectal cancer have not been shown to confer survival benefit. Since only a tiny fraction of proteins in an autologous tumor vaccine is specific to cancer cells, dilution by self-antigens is believed to make the vaccine poorly immunogenic (8). Specific colorectal cancer tumor-associated antigens have subsequently been identified and used as peptide vaccines (9–14). However, the extent of antigen presentation varies widely depending on patients' HLA type (15). Dendritic cell (DC) vaccines overcome the limitations of these two earlier vaccine approaches. Following ex vivo stimulation by tumor lysate, DCs are not confined to process any specific antigen chosen in advance and can also activate a tumor-specific T-lymphocyte response (16).

We have previously demonstrated that patients with metastatic colorectal cancer who exhibit immune responses to a DC vaccine have superior 5-year recurrence-free survival rate following surgical resection of liver metastases compared with counterparts in whom immune responses do not occur (63% vs. 18%, P = 0.037; ref. 17). Since T-cell densities in the tumor microenvironment are well-known to correlate with survival of patients who have either primary or metastatic colorectal cancer (18, 19), we now investigated potential relationships between DC vaccine response and densities of cytotoxic, memory, and regulatory T cells in the context of metastatic colorectal cancer. Furthermore, we pursued genomic analyses. A recent phase II study of pembrolizumab, an anti-programmed death 1 immune checkpoint inhibitor, for colorectal cancer showed that clinical benefit is more likely to be observed in patients with tumors carrying higher mutation burdens due to deficiency in DNA mismatch repair (MMR; ref. 20); however, the prevalence of MMR deficiency among metastatic colorectal cancer cases is only 3.5% (21). Therefore, DC vaccination may be an effective therapy for a larger portion (17) of such patients. To evaluate the feasibility of precision immunotherapy using this approach, we examined mutation patterns and constructed a vaccine response prediction model on the basis of colorectal cancer metastasis sequencing.

Details on the preparation of DC vaccines, detection of immune response, and approval for the study of human subjects have been previously reported (17). Briefly, 22 patients with colorectal cancer underwent surgical resection of liver metastases (Table 1) within 1 to 2 months of diagnosis at the Norris Cotton Cancer Center. Tumor lysates were derived from resected samples. These lysates were then added to immature DC cultures differentiated from autologous leukopheresed monocytes that had been supplemented with recombinant human granulocyte macrophage colony-stimulating factor (GM-CSF) and recombinant human IL4. At 4, 7, and 10 weeks after surgery, tumor lysate-pulsed DCs (5 × 10⁶ cells in 0.5 mL) were injected into two inguinal lymph nodes of each patient. All patients were confirmed to have no detectable residual disease by computed tomography (CT) scan prior to the initial vaccine injection. Development of immune response was defined as observing significantly greater secretion of IFNγ (22) or induced proliferation of T lymphocytes (23, 24) by peripheral blood mononuclear cells 1 week following the third autologous tumor lysate-pulsed DC vaccination, compared with unpulsed DC stimulation.

The resected specimens were formalin-fixed, paraffin-preserved, serially sectioned, and confirmed to contain metastatic tumor by hematoxylin and eosin staining. We performed immunohistochemical analysis on tumor-containing slides using the Leica BOND-MAX automated IHC system (Leica Microsystems) along with primary antibodies to human cytotoxic T cells (CD8, Novacastra #NCL-CD8-4B11), human memory T cells (CD45RO, Novacastra #NCL-L-UCHL1), and human regulatory T cells (FoxP3, BioLegend #623802). These subpopulations were specifically chosen to examine both immune-promoting (cytotoxic and memory) and immunosuppressive (regulatory) T-cell roles in the tumor microenvironment. Blinded to the clinical characteristics of patients, an experienced gastrointestinal pathologist (A.A. Suriawinata) enumerated the positively stained CD8, CD45RO, and FoxP3 T cells within the tumor, at the tumor periphery, and in surrounding normal liver tissue. Five high-power fields with a field area of 0.24 mm² were counted at each location, and average counts were recorded as cells/mm². We tested for associations between response to the DC vaccine and densities of T-cell subpopulations as well as other clinical metrics using logistic regression (Table 1).

The 22 metastatic tumor samples were submitted to the Genomics & Molecular Biology Shared Resource at Dartmouth to sequence 541 genes (Supplementary Table S1) from the HaloPlex Cancer Research Panel (Agilent Technologies). DNA was isolated from samples using the QiaAmp DNA FFPE Tissue Kit (Qiagen) and digested with restriction enzymes to create a library of DNA fragments. Digests were hybridized to HaloPlex probes for target enrichment and sample barcoding. Targeted DNA fragments were then captured by magnetic beads, amplified, and purified using AMPure XP beads (Beckman Coulter). The final library sizes were 150 to 550 bp, as determined by Qubit (Thermo Fisher Scientific) after quality control by Fragment Analyzer (Advanced Analytical). Sequencing was performed using the Ion Proton System (Thermo Fisher Scientific) with the Ion PI Template OT2 200 v2 Kit, the Ion PI Sequencing 200 v2 Kit, and a PI v2 chip. Sequenced reads in FASTQ file format (25) were aligned to the hg19 reference genome using TMAP (https://github.com/iontorrent/TMAP). The Genome Atlas Toolkit (GATK; ref. 26) was used to call variants in VCF format (27), filtering for quality score > 100 and read depth > 30 in all 22 patients. Likely germline variants were removed from consideration by matching their position, reference allele, and alternative allele with records in the National Center for Biotechnology Information (NCBI) Short Genetic Variations database (28).

Sequencing genotype data facilitate the calling of both common and rare variants. To identify associations between genes containing these variants and a given outcome, collapsing approaches that test for the collective presence of variants within genes have been developed to optimize the detection of rare variant effects (29, 30). We identified associated genes using the Combined Multivariate and Collapsing (CMC) method, which has been shown to be more powerful than previous single-marker tests and more robust to misclassification of relevant variants than previous multiple-marker tests (30). The method ranks genes by computing Hotelling's t² statistics and corresponding P values (Supplementary Table S2). Given the small number of subjects in this study, we set a liberal cutoff for choosing genes (41 genes with P < 0.1) to evaluate statistical enrichment of biologic pathways. For every curated BioCarta, KEGG, PID, and Reactome pathway (set C2) in the Molecular Signatures Database (MSigDB; ref. 31), the hypergeometric test was used to assess the probability that 41 random genes from 8,384 genes in 1,291 pathways (in MSigDB at the time of this analysis) would better represent the pathway than our top 41 genes. These probabilities can be interpreted as P values and have been adjusted for Benjamini–Hochberg false discovery rate in multiple hypothesis testing (32). In addition, we accounted for spurious enrichment due to pathway size bias and the fact that most subsets of genes on the HaloPlex Cancer Research Panel should by default be overrepresented in cancer-related pathways. Null distributions of pathway enrichment were generated by randomly sampling 41 genes from the original panel gene list and applying the hypergeometric test over 1,000 permutations. We retained only pathways (Table 2) whose observed enrichment values surpassed the 95th percentile of their respective null distributions (33).

Although the CMC method is superior to single-marker tests at detecting gene-level associations, the former masks association direction by combining variants that both positively and negatively influence the outcome of interest (34). Therefore, examining individual markers is more informative for predicting vaccine response. We derived a response prediction score for each patient using his/her tumor genotype data and variant association effect sizes computed from the other 21 patients:

S_i is the prediction score for individual i, t_k is the t statistic for difference in allele frequency of variant k between vaccine responders and nonresponders excluding individual i, all n variants chosen for consideration have t_k with corresponding P < 0.005, w_k is an assigned weight for variant k, and x_k,i is the allele count (0, 1, or 2) of variant k carried by individual i. Under this construction, S_i is more positive for individuals with tumors that have an abundance of variants associated with vaccine response; conversely, S_i is more negative for individuals with tumors that have an abundance of variants associated with lack of vaccine response. We then performed 22 rounds of leave-one-out cross-validation. Inference of response to the DC vaccine for each individual i was determined by whether S_i > C, a consistent cutoff applied to every round of cross-validation. Vaccine response prediction accuracy (percentage correct out of 22 predictions) was recorded as the highest achieved accuracy through incrementing C along the range of 22 S_i values. We repeated this procedure for various weighting schemes of w_k. First, we set w_k = 1 for all variants. With respect to each curated MSigDB pathway, we then differentially weighted variants according to pathway membership of their spanning genes (w_k = 8 for inclusion, w_k = 1 otherwise). The choice of 8 was arbitrary, as results do not change for integer weights between 4 and 15. Finally, pathways were ranked on the basis of the prediction accuracy of their annotation-guided S_i calculations, with top findings outlined in Table 3. Since leave-one-out cross-validation requires determining variant associations using only 21 of the 22 total individuals, effect sizes for some variants fluctuate between rounds of cross-validation. The upweighted components of each pathway listed in Table 3 have association effect sizes that are significant (P < 0.005) and carry the same sign in all 22 rounds of cross-validation.

Patient characteristics of DC vaccine responders and nonresponders are presented in Table 1. Of the 11 patients who had immune responses, eight patients displayed greater IFNγ secretion, five patients displayed greater T-cell proliferation, and two patients displayed both. None of the measured attributes was associated with vaccine response. The Fong clinical risk score (3), a popular prognostic tool for patients with colorectal cancer undergoing liver metastasis resection (constructed from indicators for preoperative carcinoembryonic antigen blood level > 200 ng/mL, size of largest metastasis ≥ 5 cm, >1 metastasis, lymph node–positive primary tumor, and < 12 months between primary tumor resection and liver metastasis diagnosis), was not useful for vaccine response inference. Recruitment of cytotoxic and memory T cells to metastases was significantly elevated relative to surrounding liver abundances (P < 0.001). However, vaccine response among patients with metastatic colorectal cancer did not correlate with these tumor T-cell densities. The mutation burden within targeted sequencing regions of tumor samples used to form the DC vaccine also did not differ between vaccine responders and nonresponders. Nevertheless, sequencing of the metastatic tumors revealed other important insights.

Genes identified to be associated with DC vaccine response by the CMC method demonstrate statistical enrichment of pathways that primarily involve the PI3K/Akt/mTOR signaling axis (pathway Nos. 1, 2, 4, and 7 in Table 2). In brief, PI3K is induced by extracellular signals through receptor tyrosine kinases to phosphorylate and activate Akt, which then promotes mTOR to localize to the nucleus and alter gene transcription (35). A diversity of transcription repertoire can then be affected, influencing many mechanisms such as cell survival, proliferation, motility, and metabolism. These also intersect with functions of the other highlighted transcription factor pathways (pathway Nos. 5 and 6 in Table 2). Although it is intriguing that disturbance to B-lymphocyte receptor signaling is implicated as well (pathway No. 3 in Table 2), this finding may be a statistical artifact due to the pathway's considerable overlap with the actions of Akt and mTOR, including another shared downstream effector of Akt, tuberous sclerosis complex 2 (TSC2; ref. 35). The antitumor response provoked by antigen presentation has been characterized to be predominantly T-cell–mediated, whereas the roles of B cells are still being clarified (16).

Table 3 presents the pathways that optimally guide variant weights for prediction of DC vaccine response using Eq. (A). Prediction scores of the 21 training set individuals and leave-one-out test set individual across all rounds of cross-validation for these top pathways are plotted in Supplementary. S1. Not surprisingly, training set scores cluster well into distinguishable groups of vaccine responders and nonresponders, as their differences in allele frequencies were used to determine standardized variant association effects t_k (P < 0.005) in Eq. (A). Among test set individuals, applying equal weights to associated variants correctly predicted 73% of vaccine responses (16 of 22). Differential weighting guided by top pathways increased accuracy up to 86% (19 of 22). In the absence of restrictions for pathway definitions, placing greater weight on variants within genes that comprise any top pathway further improved accuracy to 95% (21 of 22; Fig. 1); the only misclassified DC vaccine response was that of Individual 10. Even so, the corresponding prediction score became more positive with pathway-based variant weighting compared with equal weighting, which indicates adjustment in the correct direction as Individual 10 was a vaccine responder.

There is a conspicuous theme underlying upweighted variants found to improve the accuracy of vaccine response prediction (Table 3). They tend to reside within genes that encode plasma membrane–related proteins. Many are proteins that interact with the extracellular matrix (integrins, collagens, fibronectin, and cluster of differentiation; ref. 36). Some are receptor tyrosine kinases that relay signals from outside of the cell to inside (VEGF receptor, platelet-derived growth factor receptor, TGFβ receptor, and IGF receptor). Other proteins are downstream effectors (PI3K and mitogen-activated protein kinases) of cascades initiated by the aforementioned classes of transmembrane proteins. Two important contributors to prediction using Eq. (A) and distinct outliers to this trend are variants within genes for glutaminase (GLS2) and methionine synthase (MTR). Normally, enzymes involved in amino acid metabolism (pathway No. 10 in Table 3), GLS2 and MTR are suspected to promote tumorigenesis when their underexpression or impaired function compromises regulation of oxidative stress (36, 37). Furthermore, the majority of association directions across highlighted variants is negative (Table 3 and asymmetry of prediction scores with respect to zero seen in Fig. 1). In other words, response to adjuvant DC vaccination is more likely to be observed in a patient with colorectal cancer who has liver metastases that carry the reference genome allele instead of the alternative allele, for most somatic mutations with association P < 0.005 identified by our tumor sequencing.

Through histologic enumeration of T-lymphocyte subpopulations, targeted gene sequencing of colorectal cancer liver metastases, and subsequent pathway-based genomic analyses, we explored the potential biologic mechanisms that confer greater likelihood of response to an adjuvant DC vaccine among patients with colorectal cancer following metastasis resection. Identification of biomarkers that predict response to immunotherapies is critical for optimizing their use in treating patients (38). Previous investigations have shown that cytotoxic and memory T-cell infiltration of primary and metastatic colorectal cancer tumors is positively associated with patient survival after surgical resection (18, 19). Conversely, depletion of regulatory T cells enhances the tumor antigen–specific immune response (39). However, in the present study, none of these microenvironment T-cell parameters that characterize the tumor interior, periphery, and surrounding of colorectal cancer liver metastases trended with DC vaccine response or survival.

Our pathway analyses implicate signaling by PI3K/Akt/mTOR and a variety of plasma membrane–related proteins as factors that may influence vaccine response among patients with metastatic colorectal cancer. Although these inferences seemingly differ on the basis of whether we performed multiple-variant joint association testing (the CMC method) or single-variant association testing, respectively, the findings are not contradictory but rather complementary. The CMC method is more statistically robust and conservative in identifying gene associations, but the generated results suffer from absence of association direction. And while single-variant association testing is more liable to identify false positives, it is able to produce association magnitudes and directions that can be directly used in prediction modeling. Biologically, the PI3K/Akt/mTOR axis (Table 2) is in fact a convergence point for upstream signaling from integrins, receptor tyrosine kinases, and apoptosis regulators Shp2 and the caspase family (Table 3; refs. 35, 40).

These cellular mechanisms can plausibly account for tumor attributes that promote DC vaccine resistance and poorer survival. First, impaired sensitivity to molecular cues for cell death maintains tumor survival (pathway Nos. 4, 5, and 10 in Table 3). Second, tumor surface proteins interact with the extracellular matrix to repurpose the microenvironment for optimal growth (pathway Nos. 1, 3, 6, 8, and 9 in Table 3). Reciprocal signaling from the extracellular matrix to transmembrane proteins that is improperly relayed to the rest of the cell can stimulate unwarranted growth as well (10, 41). Specifically, integrins on the cell membrane regulate migration, invasion, and proliferation, especially through endothelium interactions during metastasis and tumor-initiated angiogenesis (pathway No. 7 in Table 3; refs. 40, 42). Integrin α_Vβ₃ expression in colorectal cancer liver metastases (pathway No. 3 in Table 3) has been shown to be nearly double the level in nonmetastatic primary colorectal cancer tumors (43). Tumors can also hijack the vascular migration apparatus of innate immune cells (pathway No. 2 in Table 3) by inducing them to release membrane vesicles that contain integrin α_Mβ₂ (pathway No. 8 in Table 3) for fusion with tumor cells, ultimately facilitating metastasis (44). Moreover, integrin α_Mβ₂ is involved in monocyte development and differentiation (45). Substantial loss of integrin α_Mβ₂ from monocytes may compromise their ability to secrete IFNγ or prime T lymphocytes following antigen presentation, the principal antitumor action of DC vaccines.

Third, the directions of variant association effects in Table 3 offer additional clues. As mentioned earlier, the abundance of negative effects suggest that collective loss of function among some pathway components leads to pathway dysregulation and may give rise to a more aggressive cancer phenotype. Every other variant with a positive effect is contained within a gene that encodes a plasma membrane–related or secreted protein. It has been shown that although an intracellular tumor-specific protein can trigger both cell-mediated and humoral immune responses, no appreciable tumor eradication is observed, likely a consequence of the protein's intracellular expression (46). In contrast, considerable exposed changes in peptide sequence would be expected for membrane proteins linked to the vaccine response-predisposing variants in Table 3; most are frameshift indels. Even the remaining variants are still missense mutations that cause amino acid substitutions. Despite their potential deleterious effects on pathway functions, the positive associations between these variants and DC vaccine response may be due to their gene products' cell surface immunogenicity, especially in the face of proactive tumor efforts to hide foreign antigens through underexpressing major histocompability complex class I (MHC-I) molecules or displaying MHC-I surrogates (47).

In constructing an accurate prediction model for DC vaccine response among patients with metastatic colorectal cancer following surgery (95% correct, 21 of 22 patients), we implicated several mechanisms of cancer proliferation and immune system evasion that are likely to impact immunotherapy success in this disease context. A weakness of our study is its small sample size. Substantially larger sample sizes were used to identify the survival T lymphocyte relationships reported in previous studies (18, 19) that could not be replicated here for DC vaccine response. Having more than 22 patients with colorectal cancer with liver metastases can allow the derivation of stronger and more reliable variant association effects. It would be interesting to assess whether the prediction score cutoff shown in Fig. 1 can maintain the current prediction accuracy when applied to somatic genotype data from an independent cohort of patients.

Given the discovery nature of this study, targeted gene sequencing was performed in favor of whole-genome and whole-exome sequencing. With the latter two approaches in future studies, new pathway suspicions may be raised. Furthermore, whole-genome sequencing would be able to reveal outliers in total mutation burden which limited gene panel sequencing cannot (Table 1). Perhaps uncommon patients with metastatic colorectal cancer with orders of magnitude more tumor mutations due to deficient MMR (21) will have greater tendency to benefit from DC vaccination, as they do for immune checkpoint inhibition (20). Interestingly, this within-cancer phenomenon has been shown to be untrue for DC vaccination across cancers, as clinical benefit relative to placebo is lower among patients with melanoma patients (more mutations) than patients with glioma and renal cell carcinoma (fewer mutations; ref. 48). Another important difference compared with checkpoint inhibitors is that no toxicities were noted (17), so it is conceivable that even higher doses of DC vaccine may be indicated. Sequenced variants were also called relative to the publically available hg19 reference genome instead of paired normal colorectal tissue genomes. Although we removed germline variants on the basis of annotations in the NCBI database (28), we may have still captured noise signals that are not relevant to colorectal cancer given their undiscerned presence in normal tissue.

In addition to more comprehensive bioinformatics studies, companion experimental pursuits are warranted. Our genomic analyses serve to not only construct a meaningful DC vaccine response prediction model but also suggest novel directions for molecular investigations that may lead to an increase in the response prevalence or absolute efficacy of this immunotherapy. If metastatic colorectal cancer resistance to DC vaccination is truly influenced by the display of tumor surface neoantigens and aberrant transmembrane signaling through integrins and the PI3K/Akt/mTOR axis, combination immunotherapy with pharmacologic therapy may synergistically enhance tumor-specific immune responses and also lower the required dose, plus corresponding adverse effects, of each regimen. Among patients with colorectal cancer, PI3K inhibitors and integrin inhibitors have already been separately shown in clinical trials to be well-tolerated anticancer agents, in particular with superior effects for metastatic disease (49, 50). Findings in the present study merit further work to evaluate the potential benefits of DC vaccine co-administration.

No potential conflicts of interest were disclosed.

Conception and design: D.C. Qian, R.J. Barth Jr.

Development of methodology: D.C. Qian, C.I. Amos, R.J. Barth Jr.

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A.A. Suriawinata, R.J. Barth Jr.

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D.C. Qian, X. Xiao, J. Byun, S. Her, C.I. Amos, R.J. Barth Jr.

Writing, review, and/or revision of the manuscript: D.C. Qian, A.A. Suriawinata, S. Her, C.I. Amos, R.J. Barth Jr.

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C.I. Amos, R.J. Barth Jr.

Study supervision: R.J. Barth Jr.

We would like to thank Heidi Trask and Dr. Craig Tomlinson in the Genomics & Molecular Biology Shared Resource at Dartmouth for sequencing the tumor samples.

This research was financially supported by the NIH (grants P30CA0123108 and P20GM103534) and a generous contribution from Ted and Rae Bachelder to R.J. Barth Jr.

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.