Individualized therapy of HHT driven by network analysis of metabolomic profiles

Complex diseases with multi-factorial etiologies often have multiple alternative pathways leading to a particular pathophysiological state, with a wide range of resulting phenotypes. Such diseases provide a significant diagnostic and treatment challenge, and will require individual-specific, personalized treatments. Hereditary Hemorrhagic Telangiectasia (HHT) is an example of a Mendelian genetic disease with broad variability in presentation and involvement of different organs. Arteriovenous malformations (AVM) can occur in multiple beds, including the brain, liver, lungs, and nose [1]. A challenge for treatment of this disease is selecting an appropriate treatment for a given clinical setting. Since many diseases affect metabolism directly or indirectly, the field of metabolomics has rich potential for biomarker applications. A challenge however is that a metabolomic profile alone may not provide any direction for treatment, since there is not biologically coherent integration of these data. Metabolic network reconstructions can provide the framework for integration and analysis of these data [2–16]. We considered the treatment of an individual patient and through comparative analysis of her metabolomic profile with non-HHT individuals, differences were identified using constraint-based analysis of a global human metabolic network reconstruction, Recon 1 [17].

The patient in this study is a female who began experiencing syncopal episodes at rest and during exertion at 21 years of age. She was found to have a 28 mm atrial septal defect (ASD), for which percutaneous closure was attempted, but failed, ultimately requiring open heart surgery. The ASD repair resolved the syncopal episodes, however her dyspnea continued unabated to the point where she could not walk her dog without getting chest pain and expectorating small amounts of blood streaked sputum. Further investigation with contrast echocardiography showed she had a large right to left shunt at the pulmonary capillary level but had no treatable AVMs on computed tomography (CT) scan of the chest. Her resting oxygen saturations were 97-99% but with effort could decrease to the high 80's. Her pulse was regular but varied between 90 and 120 beats per minute and variations in blood pressure were also noted and treated medically when elevated. Because of a strong family history cancer of the colon and some occasional blood streaks in her stool, a colonic endoscopy was done revealing several small polyps in her small intestine and telengiectasias in her large intestine that were ablated with laser therapy. Her Endoglin and ALK-1 gene analysis were negative including SMAD-4. Based upon these findings as well as a family history of familial adenomatous polyposis and a sibling with similar diffuse AVMs, a presumed diagnosis of HHT was given based on the Curacao Criteria [18].

In this individual case study, untargeted, quantitative plasma metabolomic profiling is carried out in a patient and healthy controls. We sought to use constraint-based modeling of metabolism on an organism scale to identify potential differences in metabolism between the non-HHT and HHT patient through identification of biomarker signature profiles (as opposed to single biomarkers) that can be linked to different functional states. These differences were then used to support the use of particular drug treatments; post-treatment profiling was carried out to assess whether the treatment revised the biochemical pathways accordingly.

Fasting plasma samples drawn from 5 healthy individuals (no chronic medical conditions, no current daily medications or herbal supplements; ages 21-37) and 1 patient with HHT (age 24) (UCLA Institutional Review Board retrospective case report exemption was obtained). The blood from the healthy volunteers was taken as part of a general protocol for internal standardization/normalization at UCSD. A set of samples from individuals in the study were used and are currently part of an ongoing study that allows acquisition and analysis of blood from healthy individuals as well as HHT patients (UCLA Institutional Review Board Protocol Number 11-000843-AM-00003). Blood draws from two different time points had been previously obtained from two of the healthy individuals and for the patient, resulting in a total of 9 samples. Additionally one individual provided a fasting as well as non-fasting blood sample, which served as a non-fasting positive control. Written informed consent was obtained from the patient for publication of this report. A copy of the written consent is available for review by the Editor-in-Chief. All of the experimental research with volunteers and the patient were performed with the approval of the ethics committees of the institutions where the studies were done.

Blood samples were obtained by a licensed health care practitioner and spun down for 10 minutes at approximately 3000 rpm at 20 C. The supernatant was stored in a freezer, and after all samples were obtained, were sent for processing and profiling (Chenomx Inc, Edmonton, Alberta). The samples were filtered using 3 kDa molecular weight cut-off filters (Nanosep 3K Omega microcentrifuge filter tubes). Internal standard solution was added to each sample solution, and the resulting mixture was vortexed for 30s. 600 μL of the mixed solution was transferred to an NMR tube for data acquisition. Spectra were acquired on a 600 MHz Varian INOVA spectrometer with 32 scans per sample at 298 K. Spectra were processed and CNX files were generated using the Processor module in Chenomx NMR Suite 6.0. Metabolites were identified and quantified using the Profiler and Library Manager modules in Chenomx NMR Suite 6.0, using a metabolite reference library of 297 metabolites.

Sample analysis

Unsupervised hierarchical clustering showed clear separation between the healthy individuals and the patient. There were not enough replicates to apply conventional statistical tests (seven measurements in one group and two in another), however two-way hierarchical clustering with unweighted average distance linkage of the metabolite profiles demonstrated a clear separation between the patient and healthy individuals, which also drove the separation in the clustergram (Figure 1). The distance measure in Figure 1 is the infinity norm, but similar separations were observed with other norms (e.g. 1- and 2-norms). Separation of the pre-treatment and post-treatment of the HHT patient as well as re-grouping of the post-treatment HHT patient with the non-HHT individuals was observed with Principal Component Analysis as well (see Additional File 1).

Differences in metabolism between non-HHT individuals and the HHT patient were assessed using the plasma metabolomic profiles. Specific metabolites that were quantitatively and qualitatively different between the non-HHT and HHT patient were determined by identifying those metabolites whose maximum concentration in one group was less than the minimum of the other group or vice versa. This resulted in two sets of metabolites that were different in the two conditions; a high set and a low set. The two general classes of metabolites were labeled the as 'ketone group' and the 'amino acid/carbon rich group' and used to define qualitative metabolic pseudo-reactions.

As a means to evaluate possible related metabolic derangements, fasting plasma metabolomic profiling was carried out (see METHODS). Fasting blood samples from 4 non-HHT individuals, was obtained and metabolomic profiling carried out as well. Two of the non-HHT individuals provided multiple fasting samples (on two different occasions) and another individual provided a non-fasting sample. The serum samples were analyzed using NMR spectroscopy with subsequent quantification and identification of 30 metabolites.

Clustering of the samples demonstrated clear separation of the non-HHT individuals fasting sample, the HTT patient (pre-treatment), and the single non-HHT individual non-fasting sample (Figure 1). The trends in the non-fasting sample were in opposite direction in comparison to all of the other samples, consistent with different feeding states. The HHT patient's two fasting blood samples were in a different branch of the clustergram as well (Figure 1). Since only two measurements of the patient (pre-treatment) were available, it was not possible to carry out t-tests or ANOVA analyses and a different criterion was required to identify the most distinctive qualitative differences in the profiles. Comparisons between the non-HHT and HHT patient (pre-treatment) were carried out in the following manner: metabolites whose minimum value in the non-HHTs was greater than the maximum value in the HHT patient's measurements or whose maximum value in the non-HHTs was less than the minimum value in the HHT patient's were used to define qualitative transport constraints for a non-HHT and HHT network (Figure 2).

A global metabolic human network, Recon 1 [17], was used to analyze the metabolomic data in a biologically germane context [19]. Since the intravascular compartment interacts with all organs of the body as both a 'source' and 'sink' for nutrients and waste, respectively, the plasma metabolite measurements were used as constraints on the global network, to grossly approximate whole organism metabolism. The two different conditions were used to specify constraints, resulting in two different networks. The two different networks were then globally assessed using FVA and compared for differences in individual reactions and pathways.

Comparison of the non-HHT and HHT patient networks using flux span ratios (see METHODS) demonstrated decreased energy production in the HHT patient, reflective of a 'starvation-like' state. Interestingly, there were noted increased flux potentials in nitrogen handling and disposition pathways in the HHT patient, notably with nitric oxide synthase (NOS) (Figure 3). This observation brought to light a potential link between vascular endothelial function and the changes in vascularity found in HHT, with connections to metabolism [1, 28, 29]. There has been evidence suggesting that VEGF can decrease blood pressure through increased nitric oxide production and conversely, that inhibition of VEGF can increase blood pressure, at least in part, through the same mechanism [30–32]. The potential increase in NOS activity based on the network analysis of the metabolic profile supported the use of bevacizumab (Avastin), an anti-VEGF drug. The HHT patient underwent treatment with bevacizumab at a dose of 5 mg/kg every 2 weeks for 6 total infusions, at which point another fasting blood draw was obtained. The patient had a mild response to therapy which lasted for 2 months. Surprisingly, the patient's profile had changed to become more similar to the control individuals. When clustered with the other samples, the post-treatment profile clustered with the rest of the non-HHT individuals (Figures 1 and 2).

An individualized case-study is described for a rare disease in which we demonstrate how it is possible to progress from untargeted metabolomic profiling to identify metabolic profiles that can be then used to constrain a mechanistic network model which is then used to direct therapy decisions. Metabolomic data has been recognized as an important 'omic' data type, as it represents a quantitative biochemical phenotype and has potential to serve as a source of diagnostic and therapeutic biomarkers. As with other high-throughput data types however, a challenge with data interpretation remains. In particular, in order to achieve any practical realization of individual specific, personalized therapies, it will be important to move away from strictly statistically driven models towards more mechanistic models.

We employed the use of a global human metabolic network to interpret plasma metabolite profiling and help direct the therapy of a patient with a complex disease. Following treatment, the patient's metabolomic profile became more similar to a non-HHT individual's, suggesting the treatment was effective, at least in addressing the metabolic derangements associated with HHT. While the disease stabilized during this treatment, the patient did not undergo further cycles of treatment due to gastrointestinal intolerance of bevacizumab and long-term assessment of disease status could not be carried out. While deeper analysis with more patients are needed to further elucidate the directionality of causation, this 'individual clinical trial' illustrates how metabolomic data analyzed in the context of a network reconstruction can be potentially used to help direct therapy for complex disease states, using complicated metabolomic profiles as opposed to individual biomarkers.

One goal of personalized medicine is to analyze the data from a particular individual and determine the appropriate medication or therapeutic intervention. In this individualized case study of personalized medicine, we carry out untargeted metabolomic profiling of a patient and healthy controls. Rather than filtering the data down to a single value, these data are analyzed in the context of a network model of metabolism, in order to simulate the biochemical phenotypic differences between healthy and disease states; the results then guide the therapy. Thus, we have found that one approach to achieving the goals of individualized medicine is to use Systems Biology and causal models to drive the analysis and interpretation of data.

FJM, MDK, and NJ conceived of the study. FJM, JM, TE, and MDK provided patient care and treatment management. NJ carried out the data analysis, simulations, and drafted the manuscript. NJ and MDK analyze the results. FJM, MDK, and NJ carried out extensive revisions to the manuscript. All authors read and approved the final content.