Conceived and designed the experiments: Roger L. Chang, Philip E. Bourne. Performed the experiments: Roger L. Chang. Analyzed the data: Roger L. Chang, Li Xie. Contributed reagents/materials/analysis tools: Philip E. Bourne, Bernhard Ø. Pallson. Wrote the paper: Roger L. Chang, Li Xie, Lei Xie, Philip E. Bourne, Bernhard Ø. Pallson. Provided conceptual content to the study: Bernhard Ø. Pallson, Li Xie.
A patent application has been filed for the modeling and simulation approach used in this study for predicting drug responses, metabolic disorders, and risk factors for treatment.
Recent advances in structural bioinformatics have enabled the prediction of protein-drug off-targets based on their ligand binding sites. Concurrent developments in systems biology allow for prediction of the functional effects of system perturbations using large-scale network models. Integration of these two capabilities provides a framework for evaluating metabolic drug response phenotypes
Pharmaceutical science is only beginning to scratch the surface on the exact mechanisms of drug action that lead to a drug's breadth of patient responses, both intended and side effects. Decades of clinical trials, molecular studies, and more recent computational analysis have sought to characterize the interactions between a drug and the cell's molecular machinery. We have devised an integrated computational approach to assess how a drug may affect a particular system, in our study the metabolism of the human kidney, and its capacity for filtration of the contents of the blood. We applied this approach to retrospectively investigate potential causal drug targets leading to increased blood pressure in participants of clinical trials for the drug torcetrapib in an effort to display how our approach could be directly useful in the drug development process. Our results suggest specific metabolic enzymes that may be directly responsible for the side effect. The drug screening framework we have developed could be used to link adverse side effects to particular drug targets, discover new uses for old drugs, identify biomarkers for metabolic disease and drug response, and suggest genetic or dietary risk factors to help guide personalized patient care.
Despite the advantages gained from drug therapy in medicine, drug development has historically presented an expensive and frequently perplexing challenge for researchers. Identifying useful drug targets for treating disease and matching them to chemical compounds that can elicit the desired effect through drug-target interaction has been the paradigm for the drug development process in the era of molecular medicine. However, this approach has yielded many failed drug treatments and an incomplete understanding of the consequences of treatments for human health, even with drugs that have made it to market and been prescribed for decades. Two major contributing factors that confound individual molecular target-based drug discovery are drug off-target binding and the lack of systems-level understanding of drug response
The growing wealth of omics data offers a valuable opportunity for novel approaches in systems medicine but also presents significant challenges for data integration
Biological systems exhibit redundant pathways and synergistic effects conferring a robustness of phenotype when confronted with external stimuli. As a result, multi-target drugs are generally more clinically efficacious than single-target drugs. These facts highlight the critical importance of studying polypharmacology in a systems level context
CETP inhibitors are intended to treat patients at risk for atherosclerosis and other cardiovascular diseases by raising high-density lipoprotein cholesterol (HDL-C) and lowering low-density lipoprotein cholesterol (LDL-C)
The predicted torcetrapib off-targets include many metabolic enzymes and metabolite transport proteins. Although there are several mechanisms involved in regulating blood pressure that may be responsible for the hypertensive side effect, one possible mode is the renal regulation of blood pressure via metabolite reabsorption and secretion. The kidneys are the primary organs that filter the blood and therefore are strong contributors to maintaining a normotensive state even independent of RAAS function. Thus a model of renal metabolism was developed as the system context in which to analyze torcetrapib off-targets and predict drug response phenotypes. The two best-supported causal off-targets predicted in this study are prostaglandin I2 synthase (PTGIS), due to decreased capacity for renal prostaglandin I2 (PGI2) secretion, and acyl-CoA oxidase 1 (ACOX1), due to decreased capacity for renal citrate and amino acid reabsorption. Four other predicted off-targets are also predicted to impact amino acid, glucose, citrate, or bicarbonate reabsorption. As well, the model predicts no effect on renal reabsorption or secretion for a number of other predicted off-target metabolic proteins.
The goal of this study is not only to provide new insight into the torcetrapib problem but also to reveal the theoretical implications that this computational systems medicine platform has for drug development and personalized medicine. Characterizing the influence that genetic variation has in determining drug response phenotypes has been recognized as a crucial goal for the future of drug development
Although many of the predictions generated by this approach are supported by clinical and other experimental evidence that describe the impact of loss of function for predicted causal off-targets and genetic deficiencies, the full set of exact metabolic mechanisms of drug action predicted by our model remain to be completely validated. While this is seen as a limitation of this study, it also offers a number of opportunities to experimentally evaluate promising hypotheses that, if validated, will lead to significant advancements in developing CETP inhibitors for treatment and novel insight into certain renal disorders.
The approach for context-specific organ modeling proposed in this study (see
Preliminary constraints were imposed upon metabolite exchange fluxes of the full metabolic network based on coordinated experimental detection of transportable metabolites both in the organ tissue and the biofluids processed by the organ. Metabolites detected in both biofluid and organ were assumed freely exchangeable in the model, and the remainder of the metabolite exchanges were tentatively constrained to zero. Organ physiology literature was reviewed to compile an objective function consisting of the metabolic functions of the organ. Each function was tested for compatibility with the preliminary model. Metabolite exchange, transport, and demand reactions required to achieve some functions were added to the network, and exchange fluxes for objective metabolites were directionally constrained in accordance with the literature. Functions not compatible with the model were removed from the overall objective function. The objective function was then integrated with gene expression data obtained from an organ tissue sample to derive a net, context-specific metabolic organ model representing the metabolic exchange between the organ and the rest of the body and the metabolic reactions that take place within the organ to achieve this exchange.
Exchange | Class | Metabolite | Abbreviation | Relation to Blood Pressure | Reference |
secretion | hormones | Prostaglandin I2 | PGI2 | vasodilation | |
Prostaglandin D2 | PGD2 | vasodilation | |||
Calcitriol | - | lowers blood pressure, Ca2+ reabsorption | |||
urea | Urea | - | water/ion counter current system regulating osmolality | ||
cyclic amp | Cyclic AMP | cAMP | important for vaso-dilation/constriction | ||
urate | Urate | - | uknown, but secreted | ||
tryptamine | Tryptamine | - | uknown, but secreted | ||
absorption | water | Water | H2O | determinant of blood pressure, ion absorption | |
ions/electrolytes | Phosphate | - | determinant of blood pressure | ||
Sodium | Na+ | determinant of blood pressure | |||
Calcium | Ca2+ | determinant of blood pressure | |||
Chloride | Cl- | determinant of blood pressure | |||
Protium | H+ | determinant of blood pressure | |||
Potassium | K+ | determinant of blood pressure | |||
Bicarbonate | HCO3- | determinant of blood pressure | |||
carboxylates | Acetate | - | unknown, but reabsorbed | ||
Citrate | - | effects sodium reabsorption | |||
Oxalate | - | effects sodium reabsorption | |||
glucose | D-Glucose | - | effects sodium reabsorption | ||
amino acids | L-Alanine | Ala | associated reduction of hypertension/vasodilation | ||
L-Arginine | Arg | associated reduction of hypertension/vasodilation | |||
L-Asparagine | Asn | associated reduction of hypertension/vasodilation | |||
L-Aspartate | Asp | associated reduction of hypertension/vasodilation | |||
L-Cysteine | Cys | associated reduction of hypertension/vasodilation | |||
L-Glutamine | Gln | associated reduction of hypertension/vasodilation | |||
L-Glutamate | Glu | associated reduction of hypertension/vasodilation | |||
Glycine | Gly | associated reduction of hypertension/vasodilation | |||
L-Histidine | His | associated reduction of hypertension/vasodilation | |||
L-Isoleucine | Ile | associated reduction of hypertension/vasodilation | |||
L-Leucine | Leu | associated reduction of hypertension/vasodilation | |||
L-Lysine | Lys | associated reduction of hypertension/vasodilation | |||
L-Methionine | Met | associated reduction of hypertension/vasodilation | |||
L-Phenylalanine | Phe | associated reduction of hypertension/vasodilation | |||
L-Proline | Pro | associated reduction of hypertension/vasodilation | |||
L-Serine | Ser | associated reduction of hypertension/vasodilation | |||
L-Threonine | Thr | associated reduction of hypertension/vasodilation | |||
L-Tryptophan | Trp | associated reduction of hypertension/vasodilation | |||
L-Tyrosine | Tyr | associated reduction of hypertension/vasodilation | |||
L-Valine | Val | associated reduction of hypertension/vasodilation | |||
oligopeptides | L-Carnosine | - | unknown, but reabsorbed | ||
Glutathione | GSH | unknown, but reabsorbed |
The kidney model included 336 explicitly predicted active metabolic genes (
The pie chart at bottom represents the Recon1 gene activity predictions resulting from deriving the kidney model. Genes predicted inactive are those genes with no associated active reaction fluxes in the kidney model. Genes for which no activity prediction was made are those associated with active reaction fluxes in the kidney model but either are not represented in the gene expression data or were not determined as the gene whose expression level is most limiting for any associated reaction through evaluation of GPR Boolean rules with respect to gene expression data. The slice at top represents genes predicted active in the kidney model.
The active reactions in the model reflect both the possible pathways by which the kidney can achieve the specified renal objectives as well as other functions supported by the gene expression data. The model included 1587 active reactions (
The distribution of metabolic reactions predicted to be active in the reduced kidney model with respect to broad metabolic subsystem categories is shown. The distribution excludes objective function, exchange, and demand reactions used to perform simulations in the model.
The integrative framework adopted for predicting causal drug targets associated with response phenotypes employed both structural bioinformatics tools as well as modeling techniques of systems biology (see
First, the human proteome was screened to identify off-target drug-binding sites. The resulting list of putative off-targets was filtered to focus on just metabolic proteins. Then, for each predicted metabolic off-target, the endogenous functional sites were compared to the predicted drug-binding site to identify overlap. Off-target proteins for which overlapping binding sites were identified were considered to be competitively inhibitable by the drug at the overlapping endogenous functional sites. The functional consequences of such inhibitions were then tested in an appropriate context-specific metabolic model. All possible individual gene knockouts were also simulated to predict genetic disorders that lead to functional deficiencies either alone or in combination with drug treatment. Those off-targets whose inhibition impacted model function represent causal off-targets predicted to be associated with the drug response phenotype, and the gene knockouts that impacted model function represent genetic risk factors for metabolic disorders, which may lead to amplification of the drug response phenotype.
Elements of the color matrix represent the percent of the maximum normal, untreated renal objective flux achievable by the CETP-inhibitor-treated normal kidney model. The x-axis corresponds to individual renal objective functions, and the y-axis corresponds to the predicted drug off-targets. Metabolite abbreviations are defined in
Official Symbol | PDB ID | Gene ID | SMAP Prediction | Functional Site Overlap | Reduced Model Reactions Limited by Expression | Impacts Renal Function in Simulation | Stronger Drug Binding Affinity | Cryptic Genetic Risk Factors | References Supporting Causal Drug Target Prediction |
PTGIS | 2IAG | 5740 | × | × | × | × | × | ||
ACOX1 | 1IS2 |
51 | × | × | × | × | × | ||
AK3L1 | 2BBW | 205 | × | × | × | × | |||
HAO2 | 1LTD |
51179 | × | × | × | × | SLC3A1;SLC7A9;SLC7A10;ABCC1 | ||
MT-COI | 1V54 |
4512 | × | × | × | × | CYP27B1;ABCC1 | ||
UQCRC1 | 1PP9 |
7384 | × | × | × | × | CYP27B1;ABCC1 |
Non-human protein structures were mapped to human genes via bi-directional BLAST against the human proteome and choosing the top hit only if it had E-value <10−50.
The renal response phenotypes for inhibition of two of the predicted drug off-targets were supported by existing scientific literature. Simulated PTGIS inhibition completely precluded PGI2 secretion. Based on the relation of renal PGI2 secretion to blood pressure (see
Two predicted causal CETP inhibitor off-targets, PTGIS and ACOX1, exhibited notable binding affinity differences when comparing docking results for their endogenous substrates to those for the three CETP inhibitors (
(
Similar to the use of the model to test inhibitory effects on drug targets, the model was also used to predict genetic deficiencies that lead to renal disorders and drug off-targets that act synergistically with genetic deficiencies. Simulated gene knockouts predicted to impact renal objective functions are displayed in
Some renal disorders were only predicted in the gene-deficient models in combination with drug treatment, not in the untreated gene-deficient models or in the normal drug-treated model, and are referred to in this study as cryptic genetic risk factors. Five such gene deficiencies were predicted (see
Multiple evaluations were performed to analyze and validate the content of the reduced kidney model. The reduced kidney model effectively predicted activity of significantly expressed metabolic genes. The ability of our modeling approach to correctly and robustly predict activity of highly expressing genes was evaluated by a five-fold cross validation (see
We compared the metabolic gene activity predictions from the reduced kidney model to the set of significantly expressed genes as well as to a proteomic dataset derived from normal, healthy human kidney glomerulus tissue
(
The literature-curated renal functions achievable by the kidney model were also compared to those achievable by a model derived from the predictions of Shlomi
Next, the model was functionally validated by comparing the gene deficiencies predicted to cause renal disorder to disease phenotypes in the OMIM database collected from clinical studies. Twenty known gene deficiencies leading to specific disease phenotypes were accurately predicted using the model (see
To more rigorously quantify the predictive ability of our model simulation approach, we performed area under receiver operating characteristic (AROC) analysis based on not only the abovementioned clinical validations of our gene-deficient phenotype predictions but based on the entire set of such known clinical phenotypes that could potentially have been investigated using our model (see
In order to assess the effects of some of the critical assumptions made in the model development and simulation procedures, we performed sensitivity analysis with respect to the predicted renal disorder phenotypes.
First, we compared the predictive capability of our reduced kidney model to that of the original, unconstrained human Recon1 metabolic network. The same approach to simulating renal disorder states was employed using both models (see
The dotted black line is the line y = x for ease of visual comparison. Red marks represent predictions resulting from inhibition of a predicted CETP inhibitor off-target. Blue marks represent predictions resulting from non-drug-target gene inhibition. Pluses represent predictions validated in the OMIM database. There are 608 marks in total plotted and exact and partial overlap of some marks precludes complete visual resolution.
Second, we investigated the sensitivity of our drug off-target response phenotype predictions to the variability of two important parameters used in our simulations, the system boundary flux constraint, set as equal fractions of the upper bound on renal objective fluxes (see
The system boundary flux constraint was imposed upon demand and exchange reactions other than those optimized during a given simulation. By default we set this constraint assuming that all allowed boundary fluxes can carry an equal fraction of the potential maximum renal objective flux. This assumption was made to allow all pathways that could possibly contribute to the objective to be used simultaneously in the optimal flux state, providing the most flexible state while maintaining maximum sensitivity of our model to additional system perturbations such as gene deficiencies or drug effects. This approach was unbiased in that it did not favor any possible pathway over another in achieving a set objective without imposing additional constraints, which may not always reflect biological reality but was the most conservative assumption in the absence of additional experimental data required to more precisely set these flux constraints. In our sensitivity analysis, we varied this parameter between 0 and 1000 flux units, the absolute lower and upper magnitudes possible in our model, and repeated the simulations of drug off-target effects. The result of this analysis (
It is clear from
We similarly analyzed the sensitivity of our predictions to changes in the degree of enzyme inhibition assumed to follow from drug treatment (
A novel approach for making functional predictions of drug response phenotypes has been introduced that integrates techniques of both structural bioinformatics and systems biology. Although the current study focused on a specific metabolic system, the general methodology excluding techniques particular to metabolic modeling are extensible to other systems such as signaling or transcriptional regulation. Non-metabolic protein drug off-targets are predictable using the same structural analysis tools, and many such off-targets have indeed been predicted as well for CETP inhibitors
The context-specific organ metabolic modeling strategy employed in this study represents an improvement upon previous efforts in this realm. Model development algorithms such as GIMME
As an early effort at modeling such a context-specific metabolic system it is important to discuss the limitations of our model. Although the functional validations presented here are compelling, currently available clinical data only permits the assessment of a subset of the predictions possible in the model. Also, the functional portion of the model, the reduced kidney model, does not and is not intended to represent a global model of kidney metabolism but only the specific renal functions studied in this work. As such, our model does not fully resolve of complexity of the human kidney. The human kidney fulfills a number of functions not studied here and is a spatially distributed system across multiple distinct tissue types. Here we have summarily replaced the various kidney sub-tissues with a single, net system model. Because we integrated expression data with curated renal functions that operate across multiple kidney tissues, it is likely that our model approximates a superset of the metabolic pathways supporting these functions. Although we have made several simplifying assumptions in the model development process, even the current level of model validation suggests that the gene and reaction content of the model is fairly accurate and that simulations in this model indeed hold predictive capability.
The simulation approach taken, optimization of a linear objective function, does not fully capture the full physiological role of the kidney. The goal of these simulations was to determine drug-target effects that may limit the capacity of the kidney to move towards a homeostatic nominal state from a state of high blood pressure, thereby decreasing the capacity of the kidney to lower blood pressure. This strategy is appropriate for the goals of the current study but would not be appropriate to simulate all physiological states of interest in the kidney. On a related note, the choice to define a disorder state based on the ratio of perturbed to unperturbed maximum achievable renal objective flux demonstrates a difference in the capacity of the renal function and not necessarily a precise flux state. Therefore this strategy too is not appropriate for modeling all physiological states.
The predictions made for CETP inhibitors in this study serve as illustrative examples of many important implications that this approach has for drug development and personalized medicine. Predicted causal off-targets for renal metabolic disorders related to blood pressure may be responsible in part or full for the clinically observed hypertensive side effect of torcetrapib. The evidence resulting from this study suggests that PTGIS and ACOX1 are both potential causal torcetrapib off-targets, the inhibition of which may explain the side effect of hypertension. In addition, AK3L1, HAO2, MT-COI, and UQCRC1 may also play a role in this side effect as we have predicted, although our docking trials did not suggest that they are bound as strongly by torcetrapib. The specific predicted deficiencies in renal function associated with the drug off-targets can serve as biomarkers to be measured in patients participating in clinical trials. A positive correlation of these biomarkers with side effects would lend support to the predictions of this study and confirm these biomarkers as risk indicators in future patient treatment. It is important to note that although these predictions comprise the basis for a renal filtration and secretion-based hypothesis explaining the hypertensive side effect of torcetrapib, these results do not refute the hypothesis based on a RAAS-mediated mechanism. These two hypotheses are not mutually exclusive and could potentially contribute alternatively or synergistically to the clinically observed side effects. This possibility illustrates the major tenet for systems biology: studying a single protein or even a single pathway is not necessarily sufficient to explain complex biological phenomena.
Aside from the confirmation that some of our predicted off-targets are known to be involved in renal disorders, we do not currently present direct experimental verification that torcetrapib binds and inhibits the predicted targets and that this inhibition leads to the predicted response phenotypes. Although this would be the obvious next step, a retrospective validation is currently hampered by the availability of the drug and the nature of the phenotypes both predicted and known. Ideally, relevant physiological studies would be carried out during actual clinical trials, when a method such as ours would be most useful, in preclinical and clinical phases of drug development.
The extended structural analysis of causal drug off-targets to identify differential binding affinities for endogenous substrates and drug molecules suggests possible differences in drug response phenotypes across the CETP inhibitors tested. The results suggest that anacetrapib may potentially lead to a similar response phenotype to that of torcetrapib, while JTT-705 may not carry the same adverse effect, at least with respect to the off-targets detailed in this study. This particular type of analysis may aid in differentiating between likely response phenotypes expected for chemically and functionally similar drugs. Results of the computational pipeline for interaction prediction between proteins and CETP inhibitors employed in this study, SMAP and docking, have yet to be confirmed experimentally. Although we are currently unable to provide direct experimental evidence for the off-target interaction predictions for this class of drugs, multiple recent studies have shown experimental support for the general efficacy of this approach for interaction prediction
The predicted renal metabolic disorders with a genetic basis suggest classes of individuals in which treatment with CETP inhibitors may pose a higher risk for adverse side effects. These predictions suggest a likely relationship between participants in torcetrapib clinical trials exhibiting symptoms of these disorders and the observed adverse side effects. The concept of cryptic genetic risk factors for drug treatment introduced in this study suggests a novel approach to personalized medicine. Should polymorphisms within these genes be clinically linked to side effects of drug treatment, the result would comprise a basis for genetic screening to assess the risk of drug treatment for future patients. Given that these cryptic risk factors are not expected to elicit the predicted abnormal phenotypes in the absence of drug treatment, identification of causal polymorphisms through association studies could only occur during clinical phase when a sufficient number of patients could be observed to gain the statistical power needed to draw significant correlations.
As illustrated above, this approach for
The binding site for CETP inhibitors on the CETP structure and the predicted off-target binding sites for this class of drug across the proteome were assumed to be as previously predicted using the SMAP program
In order to integrate the result of drug off-target predictions with the metabolic network, it was necessary to first map all PDB structures (
The metabolic enzymes predicted as CETP inhibitor off-targets using SMAP were evaluated to determine potential enzymatic inhibition by the drug. The predicted drug-binding sites of the putative off-targets were compared to endogenous ligand-binding sites from existing PDB protein-ligand complex structures (
Enzyme substrates were identified from Recon1 reaction formulas. Certain molecules (H+, H2O, O2, phosphate, ferricytochrome C, and ferrocytochrome C) were excluded from docking trials due to size or structural challenges prohibiting a useful docking result for the purposes of binding affinity predictions. All protein structures used in this study were downloaded from the PDB (
Protein structures were pre-processed for docking using AutoDockTools (ADT) version 1.5.2
Docking was performed using AutoDock Vina
As the preliminary step in modeling human renal function, the scientific literature was reviewed to compile a list of specific metabolic functions of the kidney, with a focus on functions implicated as determinants of blood pressure. This list includes a number of renal reabsorptions and secretions. Each function in this list was tested for compatibility with Recon1, downloaded from the BiGG database (
Metabolite exchange and transport reactions needed to achieve some of the renal functions were also added to the network. It was observed that Recon1 as a base model could not achieve flux through certain key renal metabolite reabsorptions: sodium, calcium, chloride, potassium, and oxalate. These deficiencies were corrected for by simply adding demand fluxes for these metabolites in the cytosol model compartment. Demand fluxes were also added for the remaining kidney reabsorptions and secretions as well to enable an array of simulations involving individual components of the renal objective function to be tested. In the case of reabsorption, this allows for direct reabsorption of metabolites in addition to indirect reabsorption in which the absorbed metabolite is first metabolized into other compounds and then reabsorbed into the blood, as is the primary mechanism of reabsorption for some metabolites, such as reduced glutathione (GSH)
A preliminary model was created by imposing kidney-specific exchange flux constraints representing the metabolic exchanges the kidney carries out with the blood and urine. The preliminary model was initialized by loading Recon1 into the COBRA Toolbox
The Human Metabolomics Database (HMDB) (
The resulting preliminary model was again tested for the ability to achieve all kidney-specific metabolic functions. It was found that this model could not absorb and metabolize GSH, without also absorbing oxidized glutathione, the exchange of which was subsequently unconstrained. Also, L-threonine and L-methionine could not be absorbed and metabolized in this model without exchange of 2-hydroxybutyrate and 2-methylcitrate, the exchanges of which were similarly unconstrained as a corrective measure. The resulting preliminary model could still achieve all the same renal objectives as the fully unconstrained model. As a final preliminary constraining measure, all system effluxes were bound to equal fractions of the default upper bound on influxes of 1000 flux units; we term this parameter the system boundary flux constraint. This was done so that any available direct or indirect reabsorption pathways could possibly be used to achieve metabolite reabsorption without biasing the model towards use of any particular pathways without further evidence. This represents the state of the model just prior to final processing using the GIMME algorithm. The fitting of the allowable fluxes to the gene expression data by GIMME ultimately determined the usable reabsorption and secretion pathways in accordance with gene expression.
Two gene expression microarray dataset for normal, healthy kidney tissue
The gene-protein-reaction associations (GPRs) in Recon1 use Entrez Gene IDs to annotate reactions in the network. To map the data from the AffyHG-U95 chips to Recon1, all genes included in Recon1 were mapped to corresponding AffyHG-U95 probesets using Bioconductor
Finally, a significant expression threshold was established for subsequent use in the GIMME algorithm. This was done by fitting the normalized gene expression data to a Gaussian distribution, estimating the mean and standard deviation of this distribution, and calculating p-values associated with each data point by subtracting the cumulative distribution function from one. The normalized data value corresponding to the p-value closest to but not exceeding 0.05 was chosen as the significance threshold; this resulted in a threshold of 991.3698 for the normalized expression data.
To integrate the renal objective function and kidney gene expression data with the preliminary model to derive a functional kidney model, the GIMME algorithm
Gene activity predictions made when deriving the metabolic kidney model were compared to the set of expressed genes with normalized expression values above the significance threshold described above. Activity predictions were also validated against a comprehensive proteomics dataset from normal human kidney glomerulus tissue
To evaluate the modeling approach used in this study, a five-fold cross validation was performed in which the data corresponding to the most highly expressed 20% of network-associated genes were excluded before deriving the kidney model. The ability of each approach to correctly predict the activity of these most highly expressed 20% of genes was measured from the overlap of predictions with the highly expressed gene set assuming a hypergeometric distribution, and the resulting probability was Bonferroni-adjusted.
All predicted metabolic protein drug off-targets were tested in the kidney model to assess the drug response phenotype caused by inhibitory effects in this system. Inhibition of metabolic proteins by the drug was modeled by constraining corresponding reactions catalyzed by drug targets to 0 flux units. Simulations of the consequences of these drug effects were performed using FBA as implemented in the COBRA Toolbox
Single gene deficiencies were also simulated in the kidney model to assess their effects on renal function and their potential as risk factors for treatment with CETP inhibitors. Each of the genes annotated to reactions in Recon1 was knocked-out of the kidney model and simulations were run using the gene-deficient kidney model both with and without drug treatment to assess effects on each individual renal reabsorption and secretion.
Drug response and metabolic disorder phenotypes were assessed by taking the ratio of maximum gene-deficient, untreated renal function flux to maximum normal, untreated renal function flux. A ratio of less than unity indicates a deleterious phenotype. Predicted metabolic disorder phenotypes were validated against previous clinical studies as represented in the Online Mendelian Inheritance in Man (OMIM) database (
Cryptic genetic risk factors for drug treatment were also predicted for which the maximum gene-deficient, untreated renal objective flux equals the maximum normal, untreated renal objective flux but the ratio of maximum gene-deficient, drug-treated renal objective flux to maximum normal, drug-treated renal objective flux is less than unity.
Sensitivity of our prediction approach to variability in parameters was performed through repeated simulation in which we varied the parameter value across the full range of possible values. We investigated sensitivity with respect to each parameter independently. A normalized sensitivity coefficient was calculated as the result of each of these simulations. This coefficient was calculated by first taking the percent difference in the predicted outcome relative to a base case, our primary results, and then dividing it by the maximum possible percent difference.
Benchmark data was collected from the OMIM database (
Reduced kidney model.
(0.50 MB XML)
Genetic deficiencies causing renal metabolic disorders. Elements of the color matrix represent the percent of the maximum normal, untreated renal objective flux achievable by the drug-treated gene-deficient kidney model. The x-axis corresponds to individual renal objective functions, and the y-axis corresponds to the individual gene deficiencies represented by their official gene symbols. Metabolite abbreviations are defined in
(0.53 MB TIF)
Genetic deficiencies causing renal metabolic disorders (continued).
(0.48 MB TIF)
ROC curves for gene-deficient phenotype prediction. The blue line represents the analysis of the predictions of the model simulations presented in this study. The red lines represent the analysis of 100 different permutation trials. The dashed black line is the line y = x.
(0.11 MB TIF)
System boundary flux constraint sensitivity. Only those drug targets and renal functions are shown for which a deficient phenotype was predicted. The x-axis is in units of flux. The black star represents the base case which is presented as our primary result.
(0.77 MB TIF)
Degree of drug-induced inhibition sensitivity. Only those drug targets and renal functions are shown for which a deficient phenotype was predicted. The x-axis values correspond to the fraction of maximal enzymatic flux achievable in the untreated simulation, which represents the constraint placed on associated reactions for each simulation. The black star represents the base case which is presented as our primary result.
(0.93 MB TIF)
Kidney gene activity predictions. A one indicates that the given gene satisfies the criterion corresponding to that column.
(0.15 MB XLS)
Kidney reaction activity predictions. Reactions with at least one non-zero flux bound are predicted active in the kidney model. All other reactions are predicted inactive in the kidney model.
(0.74 MB XLS)
Metabolic protein drug target predictions. All predicted metabolic protein targets for CETP inhibitors were subject to multiple levels of analysis to determine their possible causal role in adverse drug response phenotypes. An x indicates a positive result with respect to column labels representing each level of analysis. The proteins are sorted by the number of analyses suggesting a causal role.
(0.03 MB XLS)
Summary and validation of gene deficiencies causing renal metabolic disorders. The complete list of gene deficiencies predicted to cause renal metabolic disorders is presented. Specific impacted renal secretions and reabsorptions are also displayed using metabolite abbreviations as represented in Recon1. Clinically observed phenotypes validating the association of these genetic deficiencies with predicted metabolic renal disorders found in the OMIM database and literature are noted. These validated predictions are listed first in the table, and the remaining predictions are sorted by the type of renal disorders caused by the gene deficiencies in simulation.
(0.05 MB XLS)
Clinical gene-deficient renal disorder benchmark data. The set of clinically-observed gene-deficient renal metabolic disorder phenotypes collected from the OMIM database are presented that were used to perform AROC analysis. Clinical positives are observed disorder phenotypes with respect to specific metabolite reabsorptions or secretions associated with a gene deficiency, and clinical negatives are disorder phenotypes that were clinically confirmed not to occur in association with a gene deficiency.
(0.03 MB XLS)