Personalised therapeutic management of epileptic patients guided by pathway-driven breath metabolomics.

Background: Therapeutic management of epilepsy remains a challenge, since optimal systemic antiseizure medication (ASM) concentrations do not always correlate with improved clinical outcome and minimal side effects. We tested the feasibility of noninvasive real-time breath metabolomics as an extension of traditional therapeutic drug monitoring for patient stratification by simultaneously monitoring drug-related and drug-modulated metabolites.
Methods: This proof-of-principle observational study involved 93 breath measurements of 54 paediatric patients monitored over a period of 2.5 years, along with an adult's cohort of 37 patients measured in two different hospitals. Exhaled breath metabolome of epileptic patients was measured in real time using secondary electrospray ionisation-high-resolution mass spectrometry (SESI-HRMS).
Results: We show that systemic ASM concentrations could be predicted by the breath test. Total and free valproic acid (VPA, an ASM) is predicted with concordance correlation coefficient (CCC) of 0.63 and 0.66, respectively. We also find (i) high between- and within-subject heterogeneity in VPA metabolism; (ii) several amino acid metabolic pathways are significantly enriched (p < 0.01) in patients suffering from side effects; (iii) tyrosine metabolism is significantly enriched (p < 0.001), with downregulated pathway compounds in non-responders. Conclusions: These results show that real-time breath analysis of epileptic patients provides reliable estimations of systemic drug concentrations along with risk estimates for drug response and side effects.

Introduction

The concept of personalised medicine revolves around the idea of providing the most effective treatment with the least side effects for a given patient. In this context, the purpose of therapeutic drug monitoring (TDM) is individualising the dose to achieve maximum efficacy and, at the same time, minimise toxicity, for certain drugs with a narrow therapeutic window. Standard-of-care TDM is based on the measurements of plasma/serum drug concentration. TDM has obvious clinical benefits for patients and healthcare systems. However, it also has limitations. First, it relies on blood sampling to determine drug concentrations, which can be cumbersome to perform in infants and children. Second, drug concentrations can often not correlate with improved clinical outcome and/or minimal side effects, due to highly variable, patient-specific drug metabolism[1].

Epilepsy is a complex neurological disorder affecting around 50 million people worldwide characterised by recurrent unprovoked seizures[2]. However, treatment with one or more antiseizure medications (ASMs) allows roughly 70% of patients to live seizure free, but in the long run, 40% of those patients relapse and about 25% develop pharmaco-resistance[3]. As a result, the overall therapeutic management of epilepsy (especially, in paediatric patients) remains a challenge. Such individualised responses to medication with narrow therapeutic ranges calls for a more comprehensive phenotyping approach, beyond just monitoring systemic drug concentrations. Breath analysis has made substantial progress over the last decade by emerging analytical technologies such as secondary electrospray ionisation-high-resolution mass spectrometry (SESI–HRMS). Breath-metabolome analysis by SESI–HRMS offers a number of advantages, including noninvasiveness, short analysis time, wide metabolic coverage and capabilities to perform actual compound identification of the detected molecules (as opposed to other techniques such as chemical sensors). The latter is key to provide biochemical interpretations, hence gaining insights into the pathophysiology and drug-disease interplay. Over the last decade, a number of efforts have lifted this technology to transition from an interesting analytical platform to a standardised technique with real potential in clinical settings[4-7]. Based on prior work suggesting that this technology is capable of detecting drugs as well as drug-modulated metabolites in exhaled breath[8-11], we hypothesised that this would be the case in a clinical setting, whereby it might contribute to improved phenotyping of patients with chronic epilepsy requiring TDM. Here we show that such breath-metabolomics approach has potential to reliably predict blood levels of valproic acid (VPA, an ASM) and to offer an additional patient screening layer by providing scores for side effects and response to ASMs with minimal interference into routine clinical practice and patient invasiveness.

Methods

Participants

In total, 66 paediatric epileptic patients (mean ± SD age, 10.7 ± 3.9 years; 37 males and 29 females, Supplementary Data 1) from the University Children’s Hospital Basel (UKBB), under treatment with various ASMs requiring TDM per standard care were enroled in this study. Furthermore, we also used real-time breath data of 41 adult epileptic patients (mean ± SD age, 51.6 ± 17.1 years; 29 males and 12 females, Supplementary Data 1) from the University Hospital Zurich (USZ) to predict blood concentrations of total and free VPA. All subjects were under steady state of their ASMs at the time of measurements.

In the paediatric dataset (from UKBB) out of 123 attempted breath measurements (from 66 patients), 30 failed (Supplementary Data 1), whereas in the adult dataset (from USZ), out of 41 attempted measurements (from 41 patients; in USZ, there were no multiple visits from the same patient during study duration), four failed (Supplementary Data 1). The reasons being either (i) patients suffered from severe neurological impairment, preventing them to understand the instructions of the exhalation maneuver, or that the side effects would not allow them to perform the breath test (hence they are unable), (ii) some technical issue with instrument during patient visit, or (iii) in rare cases, clinical laboratory could not return blood concentration of (any) ASMs (Supplementary Fig. 1). This means our final dataset used in the study contained 93 paediatric measurements (from 54 subjects) and 37 adult measurements (from 37 subjects).

Paediatric measurements/patients were further annotated as follows: no such annotations were made for the adult dataset as it was only used to predict blood concentrations of total and free VPA.

Following the aetiologic classification of epilepsy by the international league against epilepsy (ILAE)[12], paediatric patients were divided into three groups. Group 1 consisted of patients with epilepsies of structural origin, group 2 consisted of patients suffering from genetic epilepsies, as well as epilepsies of unknown origin and finally, in order to differentiate them from genetic epileptic encephalopathies, and group 3 consisting of developmental and epileptic encephalopathies was created. None of the enroled paediatric patients were suffering from epilepsies of metabolic, immune, or infectious origin.

In order to assess the clinical outcome of ASMs, we further classified each data point for three categories: side effects, response to medication, and electroencephalography (EEG). Patients were labelled for each of these categories as classes I, II or III (see Supplementary Table 1 for class definitions). Side-effects questionnaire PESQ[13] (see Supplementary Table 2) was used to facilitate the side-effect comparison for epilepsies of different origins.

For the downstream prediction of clinical outcome, the dataset was subdivided as follows:

Side effects: “no side effects” (class I) vs. “side effects” (class II and III combined).

Response to medication: “responders” (class I) vs. “non-responders” (class II and III combined).

EEG: “normal” (class I) vs. “abnormal” (class II and III combined).

Instrumentation

The analytical platform employed for real-time breath analysis consisted of a SESI source (SUPER-SESI, Fossil Ion Technology, Spain) coupled to a HRMS (Q Exactive Plus, Thermo Fisher Scientific, Germany; Fig. 1a and Supplementary Fig. 2). The SESI ion source was fed with 0.1% ammonium formate in water solution, flowing (solution driving pressure 1.3 bar) through a 20 µm ID noncoated TaperTip silica capillary emitter (New Objective, USA) to generate electrospray. The settings of the SESI source were as follows: sheath gas flow rate 60, auxiliary gas flow rate 2, spray voltage 3.5 kV, capillary temperature 275 °C and S-lens RF level 55.0. Under these conditions, the nano-electrospray currents were typically in the range of 130–135 nA. SESI temperatures were set at 130 °C for the sampling line and 90 °C for the ion-source core.

Fig. 1

Overview of the study pipeline.

a The procedure begun with a patient performing five-to-six simple exhalations into a SESI–HRMS analytical platform located in the hospital premises. The breath metabolomics fingerprint was acquired in positive and negative-ion mode (5, 6 exhalations per mode). Shortly before the breath test, blood was drawn to evaluate blood/serum concentrations of ASMs. b SESI–HRMS is a real-time, noninvasive, and fast breath-metabolome analysis method. The whole breath test (i.e., positive- and negative-ion mode), lasts typically 10–15 min per patient. Positive-mode TIC from two patients, one receiving VPA and another one receiving LEV, is shown as an example (TIC of patient 29 is inverted to ease visual inspection). c Comparison of the average mass spectra between the two subjects taking VPA and LEV. The inset shows an example of time-trace at m/z 143.1066 (mass spectrum and time-trace of patient 29 inverted to ease visual inspection). For each ion, area under the curve during each exhalation was computed (shaded regions) and normalised by the exhalation time (nAUC). Then, the nAUCs of 5, 6 exhalations were finally averaged to represent mean nAUC of the ion. d This resulted in a 75 × 3252 (measurements × mass spectral features present in at least 10% of total measurements and correlated with exhalations) data matrix (z-score is only used here to ease visual representation; actual downstream analysis was done on raw numbers). e Analysis workflow used to predict VPA serum concentration based on drug-related metabolites. f The workflow used to predict side effects and drug-response scores based on drug-regulated metabolites. See Methods for more detail about panels e and f. Colour key for heatmaps is shown in-between panels e and f.

All real-time breath mass spectrometry measurements were performed in full MS mode (scan range m/z 100–400, AGC target 1e6 and maximum injection time 500 ms) in both positive- (microscans 2 and resolution of 140,000 at m/z 200) and negative- (microscans 2 and resolution of 70,000 at m/z 200) ion mode. Q Exactive Tune software (version 2.9) was used to directly control MS for these measurements. The mass spectrometer was externally calibrated on a weekly basis using a commercially available calibration solution (Pierce™ Triple Quadrupole, extended mass range, Thermo Fisher Scientific, Germany) and internally calibrated by using common background mass spectrometric contaminant masses as lock masses (positive mode: m/z 149.02332, 279.15909, 355.06993, 371.10123, and 391.28429; negative mode: m/z 60.99312, 73.0295, 87.04515, 89.02442, 101.0608, 115.07645, 225.23295 and 283.26425).

Serum concentrations of ASMs were measured at the clinical chemistry laboratory of University Hospital Basel (USB) as per their standard operating protocol (Supplementary Table 3).

Procedures

All subjects and/or parents, whichever applicable, signed informed consent to participate in the study in the presence of their neurologist. This study was approved by the Ethics Committee of North–western and Central Switzerland (ID 2017-01537; see supplementary information for complete clinical protocol) and the Cantonal Ethics Committee Zurich (ID 2019-00030). The sample-size calculation included in the clinical protocol is shown in Supplementary Fig. 3. Subjects performed prolonged exhalations directly into SESI–HRMS system following blood draw for TDM (median = 21.2 min; IQR = 38.6 min). Figure 1a shows a representation of a child exhaling into the device (see Supplementary Fig. 2 for a bigger image). During each measurement, the subjects provided 5, 6 replicate exhalations, both in positive- and negative-ion mode (Fig. 1b and Fig. 1c). The total time spent on the breath test was typically in the range of 10–15 min.

For VPA compound-identification purposes, we collected exhaled breath condensate (EBC) from one patient using an in-house condensation apparatus (containing dry ice and isopropanol). Collected EBC and pure standard of suspected molecules (dissolved in water) were analysed by ultra-high-performance liquid chromatography (UHPLC) system (Vanquish, Thermo Fisher Scientific, Germany) connected to HRMS. Samples were separated on a 50 °C heated pentafluorophenyl (PFP) column (Raptor FluoroPhenyl, 1.8 µm, 150 × 2.1 mm, Restek, USA) at a flow rate of 0.240 ml/min and eluted with a gradient between solvent A (water with 0.1% FA) and solvent B (methanol with 0.1% FA). The gradient profile was 50% solvent B between 0 and 1 min, 50–54% solvent B between 1 and 5 min, 54–95% solvent B between 5 and 5.2 min and 95% solvent B between 5.2 and 8 min followed by column reequilibration to 50% solvent B in a total 10 min run. Mass spectrometer was operated in positive-polarity full MS mode (scan range m/z 100–400, AGC target 1e6, maximum injection time 200 ms and resolution of 140,000 at m/z 200) triggering MS/MS acquisition (AGC target 1e6, maximum injection time 100 ms, resolution of 70,000 at m/z 200, loop count 5, isolation window 0.4 m/z and normalised collision energy 30) if it detects signal higher than 5000.

Data analysis

Raw mass spectra data from paediatric training-set patients were converted into mzXML file format using ProteoWizard’s msConvert (version 3.0.11233) and imported into MATLAB (version 2019b, MathWorks Inc., USA) for further analysis. First, each spectrum from all files was aligned using the RAFFT algorithm implemented in MATLAB[14]. Then MATLAB’s mspeaks and ksdensity functions were used to appropriately pick and extract the final list of 3252 features. These features were present in at least 10% of all measurements (to avoid noisy features) and were correlated with exhalations (ρspearman > = 0.6 and FDR < = 0.01) in each measurement (to avoid non-breath-related features). Finally, the mean of area under the curve during each exhalation normalised by exhalation time (nAUC) was computed for each of these features in all measurements (Fig. 1c). This resulted in a data matrix of 75 × 3252 (measurements × mass spectral features; Fig. 1d and Supplementary Data 2). This data matrix was then used to develop models (i) to predict drug concentrations (Fig. 1e) and (ii) to predict side effect and drug-response scores (Fig. 1f).

For VPA-concentration prediction (Fig. 1e), first, the full training-set was reduced to 240 features, which were present in at least 80% of the measurements, whereby the patients were receiving VPA (i.e., drug-related features). Later features in the training-set were further reduced to only 11 VPA-related features (Supplementary Table 4). Afterward, time-traces for these 11 features were directly extracted from all paediatric and adult measurements using in-house C# console app based on RawFileReader (version 5.0.0.38), an open-source.Net assembly from Thermo Fisher Scientific. These time-traces were then used to generate nAUC and three different matrices (UKBB training-set, UKBB test-set and USZ test-set). The ComBat[15] function from sva (version 3.34.0)[16] was then used to remove the known batch effect from these matrices (Supplementary Fig. 4 and Supplementary Data 2). This reduced training-set was finally used to screen for the best regression model (see Supplementary Figs. 5 and 6). Finally, we found Gaussian process regression using exponential kernel (i.e., eGPR) to be best performing on the training-set and hence it was used on an independent test-set containing paediatric and adult patients for final predictions.

In order to gain further insights into the rest of the metabolic signature captured in breath (Fig. 1f), first, the full training-set was reduced to 1005 features present in at least 50% of total measurements and with a CV greater than 30% (i.e., drug-regulated features). Later, two-sample t-test was performed followed by false-discovery rate (Supplementary Fig. 7). Afterward, MetaboAnalystR (version 2.0.4)[17] was used to add more biological insights into differentially abundant ions, by translating ions to metabolic pathways. The prediction of side effects vs no side effects and non-responders vs. responders in the training-set was conducted using significant metabolites identified by the enrichment analysis using first-principle-component (PC1) score. On this score, using only training-set data, a cutoff was assigned (based on Youden’s index) to separate predicted classes (Supplementary Fig. 8). Later, we projected UKBB test-set data on the training-set PC1 score to complete this analysis.