Hyperacute Monocyte Gene Response Patterns are Associated with Lower Extremity Vein Bypass Graft Failure

Abstract

Background

Despite being the definitive treatment for lower extremity peripheral arterial disease, vein bypass grafts fail in half of all cases. Early repair mechanisms following implantation, governed largely by the immune environment, contribute significantly to long-term outcomes. The current study investigates the early response patterns of circulating monocytes as a determinant of graft outcome.

Methods and Results

In 48 patients undergoing infrainguinal vein bypass grafting, the transcriptomes of circulating monocytes were analyzed pre-operatively and at 1, 7 and 28 days post-operation. Dynamic clustering algorithms identified 50 independent gene response patterns. Three clusters (64 genes) were differentially expressed, with a hyperacute response pattern defining those patients with failed versus patent grafts twelve months post-operation. A second independent data set, comprised of 96 patients subjected to major trauma, confirmed the value of these 64 genes in predicting an uncomplicated versus complicated recovery. Causal network analysis identified eight upstream elements that regulate these mediator genes, and Bayesian analysis with a priori knowledge of the biological interactions were integrated to create a functional network describing the relationships among the regulatory elements and downstream mediator genes. Linear models predicted the removal of either Signal Transducer and Activator of Transcription 3 (STAT3) or Myeloid Differentiation Primary Response gene 88 (MYD88) to shift mediator gene expression levels towards those seen in successful grafts.

Conclusions

A novel combination of dynamic gene clustering, linear models and Bayesian network analysis has identified a core set of regulatory genes whose manipulations could migrate vein grafts towards a more favorable remodeling phenotype.

Introduction

Surgical bypass grafting is the essential treatment for lower extremity arterial occlusive disease, with approximately 80,000 vein graft bypass operations occurring each year¹. Despite its widespread use, lower extremity revascularization still suffers from fairly modest outcomes, with 5-year graft patency rates of 50–70%². Aspirin and other anti-platelets have reduced the incidence of acute graft thrombosis³, but the vast majority of graft failures are due to the chronic processes of intimal hyperplasia and subsequent graft atherosclerosis^{4, 5}. Both exposure of the newly-implanted vein to arterial hemodynamics and the vascular trauma sustained during the operation are thought to induce a critical response from the host’s immune system which, in large part, drives the process of vein graft remodeling^{5, 6}.

While the association between circulating markers of systemic inflammation, such as C-reactive protein (CRP), and post-implantation graft events in humans has been described in the literature⁷, less is known about the influence of cell-mediated immunity following vein bypass operations. Animal models have demonstrated that monocytes, a cell type central to the innate immune system, are known to play a particularly critical role in the vessel’s response to injury^{8, 9}. In particular, early monocyte recruitment into the vein graft wall, mediated by inflammasome activation within injured vascular smooth muscle cells, has been identified as a critical driver of the hyperplastic response^10–12. Despite this mounting evidence in small animal models, the potential contribution of monocyte biology to human vein bypass graft failure has not been explored.

In the current study, we hypothesized that the early response pattern of circulating monocytes would be an important determinant of long-term vein graft success or failure. Specifically, we sought to identify key elements within the monocyte genome that could be selectively targeted to manipulate monocyte activity and, consequently, graft outcome. By exploring the evolution of the human monocyte transcriptome throughout the first month directly following infrainguinal vein graft implantation, we identify a core set of genes that are associated with graft failure. Employing Bayesian network analysis and a series of regression models to describe interactions among these key monocyte regulatory genes, a predictive model describing the downstream effects of their manipulations upon graft outcome is developed. Ultimately, through a combination of high-throughput genomics, gene pathway analysis and probabilistic network modeling, we identify specific monocyte genes whose alterations may result in improved vein graft patency in human patients.

Methods

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Patient Selection and Outcome Measures

Patients undergoing autogenous infrainguinal vein bypass grafting for critical limb ischemia or disabling claudication between 2007 and 2012 at either the UF Health-Shands Hospital or the Malcom Randall Veterans Affairs Medical Center were candidates for enrollment in this prospective study. All patients provided informed consent prior to participation. The project was approved by the Institutional Review Board at the University of Florida and the Research and Development Committee at Malcom Randall VAMC.

Infrainguinal vein bypass grafting was performed in a standard fashion, using non-reversed ipsilateral long saphenous vein whenever possible. All patients were placed on aspirin; those with composite vein conduit or disadvantaged arterial outflow were placed on warfarin. Post-operatively, the subjects underwent duplex ultrasound, ankle-brachial indices, and physical exam at 1 week and 1, 3, 6, and 12 months to monitor graft patency. Grafts that occluded, required re-intervention, or showed evidence of stenosis (>3.5-fold peak systolic velocity increase across the segment and a >70% stenosis on subtraction angiography) within the first post-operative year were considered failures. Grafts demonstrating stenosis or requiring re-intervention within 1 week of surgery were considered technical failures and excluded from the analysis.

Means comparisons of demographic and procedural variables were performed with SPSS software (version 22, Chicago, IL). Either Pearson’s chi-squared or Fisher’s exact test was used for categorical variables, and Student’s t-test was employed for continuous variables. P < 0.05 was the criterion for null hypothesis rejection.

Sample Collection, Monocyte RNA Isolation, and Microarray Hybridization

Fifteen milliliters of whole blood were collected in ethylenediaminetetraacetic acid (EDTA) directly prior to the operation and then at 1, 7 and 28 days post-operation. Monocytes were isolated by negative selection, evaluated for surface marker phenotype using flow cytometry, and prepared for RNA isolation as described in the Expanded Methods. cDNA was generated using Ovation Pico WT kit (NuGEN, San Carlos, CA) and labeled using GeneChip WT Terminal Labeling (Affymetrix, Santa Clara, CA). Samples were hybridized to either a proprietary Glue Grant Human Transcriptome Array (GGH2, Affymetrix) or standard Human GeneChip (GGH3, Affymetrix)¹³. The raw expression data were normalized and transformed with Partek Genomics Suite (Partek, St. Louis, MO).

Microarray Gene Expression and Genomic Analysis

BRB-Array Tools (NIH, Biometric Research Branch) linear mixed effects modeling was used to evaluate the interrelationship of time and outcome upon regulator gene expression, with time and outcome modeled as fixed factors and patient treated as a random factor such that

$$\mathit{gene}{\mathrm{expression}}_{j(i)k}={\mathit{outcome}}_{j(i)+}{\mathit{time}}_{\mathit{ik}}+{\mathit{patient}}_i+e_{j(i)k},$$

where i = patient index, j = graft outcome, k = time index and e is the residual error term.

Hierarchical clustering and unsupervised heat map creation were performed using dChip software (Harvard, Boston, MA). Principal component analysis (PCA), which illustrates cumulative gene expression relationships among time points and between outcome groups, was performed with Partek Software and Matlab. Distance from reference (DFR) scores were used to quantitatively compare the gene expression deflection from the entire baseline group between successful and failed grafts, using the equation

$$\text{DFR}=\ln {\displaystyle \sum }_{\mathit{probe}\mathit{sets}}\frac{{(e_i-M_i)}^2}{V_i}$$

where e_i is the patient’s gene expression level and M_i and V_i are the whole group mean and variance for the i^th probe set¹⁴. Split-plot ANOVA tested the interaction between time and outcome (or severity) group for DFR scores.

A previously-described custom clustering algorithm grouped together genes with similar expression profiles over time in order identify those clusters most likely to be associated with vein graft outcome¹⁵. Briefly, gene expression levels over time were modeled using a non-parametric approach based on Legendre orthogonal polynomials (LOP). The optimal number of gene clusters and LOP order were determined by plotting Bayesian information criterion (BIC) values for various numbers of gene clusters at several LOP orders. The combination of gene cluster number and LOP order that offer the best curve fitting was selected for modeling. By calculating the posterior probabilities of each gene belonging to each cluster, the most likely cluster to which each gene belongs was determined.

Two-way ANOVA, minimum normalized effect size, and expression fold change criteria were used to identify those pathways most highly associated with the outcome variable. Pathway and gene ontology analyses, as well as identification of important upstream regulatory genes¹⁶, were performed with Ingenuity Pathway Analysis (IPA) software (QIAGEN Silicon Valley, Redwood, CA).

Time-Dependent Gene Expression Kinetics

Expression data for the 64 genes contained within the three final clusters were fit to the following exponential function in order to model the maximum expression and decay time constants (Matlab, Version R2014b, MathWorks, Natick, MA).

$$E(t)=\frac{1}{E_{\mathit{max}}^{\frac{t-t_{\mathit{max}}}{\mathrm{\Delta }T}}},\mathit{where}$$

E(t) = gene expression at time point, t; E_max = the patient’s maximum value of post-operative gene expression; t_max = the time point at which the patient’s maximum value of gene expression occurs; and ΔT is the time constant of decay.

Interrelationship among Upstream Regulator Genes and a Gene-based Model to Predict Graft Phenotype

Gene expression data for the eight upstream regulator genes were discretized as 1 (downregulation) or 2 (upregulation) by comparing with the baseline expression level. A Bayesian network was created to describe the quantitative relationships among these genes, as describe in the Expanded Methods.

A model to predict the directional change in mediator gene expression that would result from silencing of a specific upstream regulator gene was created. Using those interactions among regulator genes that were identified by the Bayesian network and the known biologic interconnectedness between downstream mediator and upstream regulator genes, two linear models were developed to describe the relationship among the regulator genes and between the regulator and mediator genes.

Comparative Gene Expression Data from Lower Extremity Angioplasty and Trauma Patient Cohorts

Gene expression data from two of our group’s previously published parallel studies on lower extremity angioplasty and blunt trauma patients were compared with results of the current study. These data were collected from circulating monocytes and processed and analyzed using technique analogous to those used in this study. Data from 14 patients undergoing lower extremity angioplasty/stenting (8 success, 6 failure)¹⁹ and 126 patients suffering blunt traumatic injury (63 complicated, 63 uncomplicated recovery course)²⁰ were included. Criteria for angioplasty success and failure were similar to those used in this study. Trauma patients who recovered in <5 days were considered “uncomplicated”; a recovery time >14 days, no recovery by 28 days, or death was categorized as “complicated”.

Results

Patient Demographics and Procedural Details

Data collected from 48 patients undergoing autogenous infrainguinal vein bypass grafting for critical limb ischemia or disabling claudication were included in the final analysis. Thirteen grafts (27%) failed within the first post-operative year (Figure S1). Mean age at operation was 64.3 years, and all but two patients were men. A history of coronary artery disease, diabetes mellitus, hypertension, hyperlipidemia and current or prior smoking was common, as was usage of aspirin and lipid-lower statin agents (Table S1). The mean pre-operative ankle-brachial index (ABI) was 0.39. In all but three patients, surgical bypass was performed to resolve critical limb ischemia. The common femoral artery was the most common inflow vessel (77%) and the tibial vessels the most common distal target (71%). The preferred conduit, a single segment of non-reversed greater saphenous vein, was utilized in the majority of cases (63%) (Table S2). With the exception of hyperlipidemia (P = 0.048), no pre-operative co-morbidity, medication or operative detail was present at a significantly greater frequency in either of the outcome groups.

Isolated Monocyte Phenotypes

Classical monocytes (CD14⁺CD16⁻) made up 82-87% of the total monocyte pool in patients whose grafts remained patent, whereas they comprised 74-82% of the pool in failed grafts (Table S3). Independent t-tests at each time point showed no differences in the percentages of classical monocytes between successful and failed grafts. Similarly, the time-dependent trajectories (time*outcome effect) of classical monocyte proportions were not significantly different between the two groups.

Identification of a Temporally-dynamic Gene Subset

Mixed effects modeling demonstrated time rather than outcome to be the dominant influence on gene expression. Expression levels of 1,870 of the 20,213 genes represented on the microarray chip changed significantly over time (FDR<0.001), yet no single gene showed a significant difference in expression between outcome groups. PCA revealed that the greatest perturbation in gene expression levels occurred within the first post-operative day and corroborated the mixed model’s results by showing no separation between the outcome groups (Figure 1A). Similarly, an unsupervised heatmap of the 1,870 genes at each patient*time point showed rough clustering by time, notably at Day 1, but no clustering by outcome (Figure S2). A more quantitative evaluation of these relationships demonstrated minor separation of successful and failed graft DFR expression profiles at Day 7 and Day 28 which failed to reach statistical significance (P = 0.98, Figure 1B).

Figure 1. The effects of time and outcome on expression levels of the 1,870 gene set.

(A) PCA analysis of patients’ gene expression profiles (n=48 patients) showed a time-dependent separation, with the greatest perturbation from baseline seen at day 1. Examination by outcome yielded no clear spatial difference between the two groups. (B) Consistent with the PCA analysis, DFR calculations showed insignificant (P = 0.98) separation between outcome groups.

Gene Cluster Generation and Selection

While mixed effects modeling identified no single gene with outcome-dependent differential expression, we suspected that there may be groups of genes that, when considered in aggregate, might show an expression difference between the outcome groups¹⁷. To test this hypothesis, a custom clustering algorithm was applied to the 1,870 time-dependent genes and divided the genes into 50 unique clusters, each with its own time-dependent signature pattern¹⁵. Each gene was assigned to only one cluster (Figure 2).

Figure 2. Clustering genes with similar time-dependent expression profiles.

A custom clustering algorithm partitioned the 1,870 time-dependent gene set into 50 independent clusters. Within a particular cluster, expression profiles for patient with a successful (S) or failed (F) graft outcome displayed a similar expression profile over time. Each gene belonged to only one cluster.

These clusters were then subjected to a stringent selection process intended to identify those clusters that are most likely to be associated with vein graft outcome. Only three of the 50 clusters satisfied all four of the following selection criteria: 1) significant outcome effect identified by the clustering algorithm, P < 0.001; 2) significant outcome effect identified by traditional ANOVA, P < 0.05; 3) normalized Cohen’s d effect size at any time point between outcome groups greater than 0.5; and 4) greater than 1.4-fold or less than 0.7-fold change in mean expression between any two time points. These four criteria were used to ensure that gene expression within each selected cluster was substantially dynamic over time and met a generally-accepted threshold for differential expression between outcome groups (Figure 3, Figure S3). Common to all 64 genes within the three clusters was early augmentation of expression at Day 1 with eventual return to baseline toward the end of the post-operative month. This heightened early response was significantly more pronounced in those grafts that progressed to failure (Mean fold change from baseline, 2.17 vs 1.76, P < 0.001, Figure 4A). Furthermore, failed grafts demonstrated a more rapid decrease in gene expression levels over the post-operative month (decay time constant 178 s. vs. 207 s., P < 0.001). Enriched canonical pathways for this 64-gene set, identified with the IPA software, included Interleukin-10 (IL-10) and Interleukin-6 (IL-6) signaling, glycogen degradation, Signal transducer and activator of transcription 3 (STAT3) signaling and granulocyte adhesion and diapedesis (Figure 4B).

Figure 3. Identifying genes most likely to be associated with graft outcome.

The 50 gene clusters were required to meet four criteria in order to be included in the pathway analysis. 1) A significant outcome effect identified by the clustering algorithm, P < 0.001; 2) A significant outcome difference identified using a two-way (time/outcome) ANOVA, with P < 0.05 for outcome; 3) A normalized Cohen’s d effect size at any time point between outcome groups greater than 0.5; and 4) greater than 1.4-fold or less than 0.7-fold change in mean expression between any two time points.

Figure 4. Expression profiles of the final 64 genes contained in 3 clusters.

(A) All genes showed early augmentation of expression at Day 1 followed by eventual return toward baseline over the course of the month. This heightened early response was significantly more pronounced in those grafts that progressed to failure (Mean fold change from baseline, 2.17 vs 1.76, P < 0.001). Failed grafts also demonstrated a more rapid decrease in gene expression levels over the post-operative month (decay time constant 178 s. vs. 207 s., P < 0.001). (B) Enriched canonical pathways for this 64-gene set identified with IPA software.

Comparison to the Monocyte Response following Peripheral Angioplasty and Major Trauma

Using datasets available through two previous clinical studies performed by our group, a comparison of the aggregate monocyte response pattern for this set of 64 genes was performed for vein bypass grafting, peripheral angioplasty, and major trauma. Significant time-dependent variation in gene expression was observed in all three clinical scenarios, with major trauma demonstrating the most robust Day 1-deviation from baseline among these groups (Figure 5A). Consistent with the less invasive nature of the procedure, a more modest response following angioplasty was observed.

Figure 5.

Comparison of the monocyte response pattern following vein bypass grafting, angioplasty, or major trauma. (A) PCA of the 64-mediator gene set demonstrated spatial separation as a function of time for both the bypass and trauma groups. Patients undergoing angioplasty showed a limited deviation from pre-procedural levels. (B) Although not apparent in the less severe scenarios of bypass grafting and angioplasty, significant differences in gene expression at 1 and 7 days were observed for individuals experiencing a complicated versus and uncomplicated trauma recovery (P<0.001). (C) Among the patient groups, the largest perturbation in gene expression was observed in the trauma group, with both bypass and angioplasty inducing a more tempered response. Consistent with the PCA, clear separation between the trauma complicated and uncomplicated outcome groups was observed. P-values test the time*outcome or time*severity interaction within the three patient groups.

Examining the relationship of these 64 genes to patient outcome (i.e. vascular intervention patency or trauma recovery course) demonstrated significant differences in both Day 1 and Day 7 expression for complicated versus uncomplicated clinical recovery following trauma (Figure 5B). In contrast, limited separation of aggregate gene expression as a function of angioplasty or bypass outcomes was observed. Examination of these data sets using a DFR plot corroborated these findings (Figure 5C).

Comparison to Weighted Gene Co-expression Network Analysis

To compare the robustness of our clustering approach to other networking algorithms, the monocyte mRNA dataset was analyzed using the Weighted Gene Co-expression Network Analysis (WGCNA) methodology. Technical aspects of the analysis and results are in the Expanded Results section. Of the three post-operative time points, common expression patterns associated with vein graft outcome were identified only at Day 7, where four of the 19 clusters reached a significance threshold of P<0.05. Unlike our current data mining technique, analysis of these genes within IPA failed to identify any upstream regulators, implying limited biologic commonality among these genes. No gene was common between the set of 64 genes derived from our clustering algorithm and that identified by WGNCA. One key difference between the two methods is the importance of gene dynamics. While our clustering algorithm implicitly integrates changes in gene expression over time, WGCNA employs a time-static approach.

Identification of Upstream Regulatory Elements

IPA Pathway analysis of this gene set identified eight upstream elements as important regulators of the larger set of 64 genes (Figure 6). Seven of these genes [Interleukin 6 (IL-6), STAT3, Angiotensinogen (AGT), Colony stimulating factor 3 (CSF3), Endoribonuclease dicer (DICER1), Hepatocyte growth factor (HGF), Vascular endothelial growth factor A (VEGF-A)] were common among both success and failure groups, at Days 1 and 7. However, by Day 28, only IL-6 and STAT3 were relevant regulators for those grafts that failed, and AGT, DICER1 and an eighth gene [Myeloid differentiation primary response gene 88 (MYD88)] were dominant regulators for successful grafts. Ontology analysis revealed a set of highly interconnected regulatory genes with central roles in inflammation, cell growth, angiogenesis and neutrophil proliferation.

Figure 6. Constructing a regulatory gene network.

Eight genes were identified as the dominant upstream regulators of 64-mediator gene set that was identified through cluster analysis. Dotted gray lines indicate known biological relationships between genes. Genes are colorized based on their relative change from baseline expression for successful versus failed grafts. Positive values (i.e. the scenario in which successful grafts had a greater increase from baseline compared with failed grafts) are colored red. Negative values (i.e. the reverse scenario) are colored blue.

While the expression levels of these upstream regulatory genes within the monocyte were dynamic over time and varied between outcome groups (Figure 7A), mixed effects modeling demonstrated only moderately significant time-dependent differences in several regulators (DICER, HGF, STAT3; Table S4). We hypothesize that these regulators form an integrated network that acts in concert to modulate the expression of downstream meditator genes.

Figure 7.

Regulator gene expression profiles and Baysian network configuration. (A) Mixed effects modeling identified DICER1, HGF and STAT3 expression as changing significantly over time (P = 0.01, P = <0.01, P = <0.01). CSF3 expression demonstrated significantly different time-dependent trajectories between the outcome groups (P = 0.05). The remaining genes showed temporally dynamic expression that failed to reach statistical significance. (B) Gene expression data for the eight upstream regulator genes were discretized as 1 (downregulation) or 2 (upregulation) by comparing with the baseline expression level. A Bayesian network was created to describe the relationships among the genes. This analysis identified four interactions to which the linear regression models for predicting regulator gene expression changes were constrained.

Alteration of Predicted Graft Outcome through Manipulation of Regulator Genes

Bayesian network analysis among the eight upstream regulator genes identified four interactions to which the linear regression model for predicting regulator gene expression changes were constrained (Figure 7B). Silencing regulator genes individually within our model yielded a wide range of predicted aggregate effects upon the downstream mediator genes. In general, early regulator gene inhibition tended to favor movement towards a more successful phenotype.

Interestingly, the silencing of some regulators had varying effects at different time points, likely reflecting the time-dependent nature of the processes that govern graft remodeling. Silencing of either the transcription factor STAT3 or the signal transduction molecule MYD88 were predicted to have the most beneficial influence, modulating downstream mediator expression strongly towards success throughout the post-operative month. In contrast, removal of HGF was associated with a negative effect on graft phenotype, which escalated throughout the postoperative month. Other regulators, such as AGT or IL-6, demonstrated a mixed picture, where inhibition at Day 1 shifted gene expression strongly towards success but silencing at subsequent time points had minimal impact on the predicted phenotype (Figure 8A, Figures S4 and S5).

Figure 8. Aggregate predicted phenotypic effects of regulator gene silencing.

(A) Two integrated, linear models were used to predict the change in expression for each mediator gene following removal of an individual regulator from the network. Summation of these changes in individual mediator gene expression yields an aggregate predicted effect for the removal of each regulator from the network, at each post-operative time point. Note the wide range of predicted phenotypic effects between genes and even within genes at varying time points. (B) Simultaneous removal of pairs of regulator genes from the network failed to demonstrate any synergistic predicted effects.

Pairs of regulators were inhibited simultaneously to explore the possibility of synergistic effects. However, in every case the aggregate effect of simultaneous knockout was either very similar or identical to the effect achieved of simply summing the individual effects (Figure 8B).

Discussion

The linkage between chronic activation of the immune system and the severity of vascular disease has been well established^{18, 19}. General markers of inflammation, such as CRP and IL-6 have been correlated with advanced atherosclerosis and an elevated risk of morbidity and mortality²⁰. While humoral immunity has a long recognized role in the initiation and progression of atherosclerosis²¹, the importance of cell-mediated immunity in physiologic and pathophysiologic vascular adaptation has been increasingly recognized. Specifically, monocyte infiltration and macrophage activation are a critical and necessary component to homeostatic arterial remodeling triggered by hemodynamic stress^{22, 23}.

These concepts have been extended to vein graft biology and the potential mechanisms of failure. Enhanced systemic inflammation, as assessed by CRP levels, has been associated with impaired remodeling and increased failure in human vein grafts^{24, 25}. Further mechanistic investigation in animal models has identified enhanced monocyte binding and macrophage accumulation as importance drivers of accelerated intimal growth^{9, 26}, particularly in areas exposed to reduced shear stress²⁷. Subsequent studies have provided granularity to these findings, identifying early monocyte recruitment and activation by necrotic smooth muscle cells, injured during implantation, as fundamental components of pathologic vein graft remodeling^{10–12, 28}. The potential importance of this mechanism in human vein graft disease is supported by the observation that enhanced monocyte adhesion immediately following surgical manipulation is associated with subsequent graft failure^29–31. However, a comprehensive understanding of the early monocyte response to vein graft implantation and the genomic modulation that drive grafts toward either patency or stenosis and failure is lacking. Hence, we have taken a “top-down” approach by using high throughput genomics to examine the dynamic signatures within the monocyte and then, in a systematic and logical fashion, culling these approximately 20,000 genes to a set of eight, therapeutically-modifiable genes that appear integral to determining a monocyte response phenotype associated with graft outcome.

This relationship between the systemic inflammatory environment and local graft remodeling begs the question of causality. We believe that the circulating monocyte acts at the site of implantation to influence vein graft biology and, ultimately, clinical phenotype. The reverse scenario, in which the monocyte is simply a marker of graft health is less plausible, given the existing data derived from animals showing monocytes and macrophages as drivers of intimal proliferation ^{9, 26}.

Our finding that surgical bypass grafting elicits a significant perturbation in monocyte gene expression, as evidenced by the identification of 1,870 genes whose expression changed significantly throughout the first post-operative month, is consistent with previously published studies. In cells isolated from the outer wall of human saphenous vein grafts, Kenagy recently showed that 5% of the genes significantly altered their expression following stimulation with either PDGF or serum³². Similar changes have been observed in other vascular pathologies, where blood leukocytes isolated from patients undergoing open thoracoabdominal aneurysm repair showed time-dependent changes of 146 genes within the first 24 hours alone³³.

The early (Day 1) augmentation of monocyte gene expression that was seen uniformly among our group of 64 genes is consistent with existing data, indicating that post-operative inflammatory processes reach a zenith within the first several post-operative days¹. Our finding that expression levels of these same 64 genes following trauma exhibit a similar pattern suggests that this gene group may be fundamentally important in the immune response to injury. Genomic analysis of endothelial and smooth muscle cells derived from canine vein bypass grafts showed a similar temporal pattern of expression, with the maximal number of differentially expressed genes seen at 12 hours and 7 days post-operation¹. The striking pattern of increased monocyte gene expression among failed grafts compared to patent grafts at Day 1 is congruent with the general observation that increased inflammation is associated with vein graft failure.³⁴

Our approach to identifying the discrete set of biologically important genes from among the larger pool relies critically upon the idea that it is likely groups of genes acting in concert, rather than lone genes in isolation, which drive the metabolic processes governing graft remodeling. The creation of our custom clustering algorithm to group together genes with similar expression profiles over time was borne from this concept¹⁵. Kenagy et al., also appreciating the importance of considering genes in groups rather than as independent entities, utilized WGCNA in an attempt to identify the core genes that may play critical roles in vein graft failure. This technique, like ours, assumes the fundamental concept that genes with similar expression likely act in a coordinated manner. As in our study, their genomic analysis of “outer wall cells” isolated from human vein grafts demonstrated that, while no single gene was differentially expressed between successful and failed grafts, two separate groups comprising a total of 42 genes were associated with graft outcome ³².

The identification of eight key regulatory genes through pathway analysis is an important contribution of the current study, proposing these upstream regulators as prime targets for possible intervention to modify vein graft outcome. Among our set of eight genes, several have known roles in vascular pathophysiology. Among them, VEGF is well-established as a principal regulator of angiogenesis³⁵.

Our Bayesian network analysis identified DICER1 as the central node among the eight regulator genes, implying its role as a governor of this small gene network. DICER1 encodes the enzyme Dicer, which plays a critical role in RNA silencing through the creation of small interfering RNA (siRNA) and microRNA (miRNA). A substantial amount of research has identified associations between Dicer and endothelial atherosclerosis by its ability to generate miRNA that regulate processes such as macrophage activation, inflammation, and vascular smooth muscle cell phenotype^36–39. Our model’s identification of Dicer as a central network regulator supports the notion that miRNA might play important roles in vascular remodeling within the context of our study.

The transcription factor STAT3 plays a role in converting vascular smooth muscle cells from a quiescent state to a synthetic state⁴⁰. When transfected into these cells, STAT3-targeted siRNA inhibition reduced their proliferation in vitro and also led to decreased neointimal thickness in rat jugular-carotid grafts⁴¹. Further work confirmed this observation, where inhibition of STAT3 activation by SOCS3 overexpression attenuated smooth muscle cell growth and migration⁴². Our model identified STAT3 silencing as one of the strongest pushes toward graft success and thus corroborates, for the first time with human data, the clinically beneficial effects of STAT3 inhibition that previously have been demonstrated solely in animal studies.

Our analysis revealed a highly interconnected set of regulator genes with variable early response patterns. Not surprisingly, the model identified several regulator genes, most notably hepatocyte growth factor (HGF), whose in silico inhibition migrated the predicted graft phenotype towards a failed outcome. HGF’s positive role in vein graft remodeling was demonstrated in rabbit vein grafts which showed reduced wall thickening among those overexpressing lentiviral-delivered HGF⁴³. In-vivo delivery of an engineered HGF fragment within an extracellular matrix-derived hydrogel improved left ventricular remodeling and increased arteriole density in a rat model of myocardial infarction⁴⁴. Its use has already expanded to humans, as a phase I clinical trial demonstrated the safety of intracoronary administration of adenoviral-mediated HGF to patients with severe coronary artery disease⁴⁵.

Limitations of our study are largely related to certain assumptions that are implicit in the analysis. Among the most critical of these assumptions are that the most biologically important genes act in concert with similar expression profiles, that the biological interactions amongst the regulatory elements are properly characterized by the Bayesian network, and that the final graft phenotype can be adequately predicted by a linear, combinatorial analysis. Our study also lacked an independent validation data set to confirm our findings. While not definitive, the observation that the most critical regulators have known roles in influencing the activation state of inflammatory cells lends support to our conclusions.

Statistical modeling is the backbone of our analysis. The collective use of mixed-effects modeling to identify temporally-dynamic genes, a gene clustering algorithm to group together those that behave similarly, a Bayesian network to identify gene regulatory relationships and linear regression modeling to predict the vein graft phenotypic effects of regulator gene manipulations is a novel approach that provides tremendous insight into the monocyte genomic influence on vein graft outcome in human patients. Importantly, this approach has given us, through statistical inference, results that would otherwise have been logistically unobtainable through conventional experimentation on humans. This same process can be applied to numerous other analyses in which one needs to distill large amounts of gene expression data down to only that tiny subset which is of greatest interest.

Clinical Perspective.

Despite its widespread use, lower extremity vein bypass grafting still suffers from fairly modest outcomes, with 5-year patency rates of 50%-70%. Both exposure of the newly implanted vein to arterial hemodynamics and the vascular trauma sustained during the operation are thought to induce a critical response from the host’s immune system that, in large part, drives the process of vein graft remodeling. Monocyte recruitment into the wall, mediated by inflammasome activation within injured vascular smooth muscle cells, has been postulated as one of the driving mechanisms for an accentuated hyperplastic response and narrowing of the lumen in the failing vein graft. Examining the genomic response of the circulating monocyte in the initial days to weeks following vein graft placement, we demonstrate that a hyperacute early response pattern of a distinct set of 64 genes is associated with clinically significant stenosis or thrombosis of the graft within the next 12 months. The clinical importance of these early response genes was also examined in an independent data set, which was acquired from monocytes isolated from patients subjected to a major traumatic injury. Analogous to the vein graft findings, augmented expression of these 64 genes within 24 hours was associated with major complications in the patients’ recovery. Pathway analysis identified a core set of eight upstream regulatory elements that control the expression of these critical response genes, offering potential targets to modify the biology and subsequent negative sequelae associated with early monocyte activation.