Normal variations in brain oxygen extraction fraction are partly attributed to differences in end-tidal CO₂

Abstract

Cerebral oxygen extraction fraction is an important physiological index of the brain’s oxygen consumption and supply and has been suggested to be a potential biomarker for a number of diseases such as stroke, Alzheimer’s disease, multiple sclerosis, sickle cell disease, and metabolic disorders. However, in order for oxygen extraction fraction to be a sensitive biomarker for personalized disease diagnosis, inter-subject variations in normal subjects must be minimized or accounted for, which will otherwise obscure its interpretation. Therefore, it is essential to investigate the physiological underpinnings of normal differences in oxygen extraction fraction. This work used two studies, one discovery study and one verification study, to examine the extent to which an individual’s end-tidal CO₂ can explain variations in oxygen extraction fraction. It was found that, across normal subjects, oxygen extraction fraction is inversely correlated with end-tidal CO₂. Approximately 50% of the inter-subject variations in oxygen extraction fraction can be attributed to end-tidal CO₂ differences. In addition, oxygen extraction fraction was found to be positively associated with age and systolic blood pressure. By accounting for end-tidal CO₂, age, and systolic blood pressure of the subjects, normal variations in oxygen extraction fraction can be reduced by 73%, which is expected to substantially enhance the utility of oxygen extraction fraction as a disease biomarker.

Introduction

Cerebral oxygen extraction fraction (OEF) is a key physiological index of the homeostasis of oxygen demand and supply in the human brain. Previous studies suggested that brain OEF, even just measured at a global level, may be a useful biomarker in a variety of diseases, such as stroke,^1–4 arteriostenosis,^5,6 Alzheimer’s disease,⁷ multiple sclerosis,⁸ metabolic disorder,⁹ anorexia,¹⁰ end stage renal disease,¹¹ hepatic encephalopathy,¹² and sickle cell disease.^5,13,14 Therefore, quantitative assessment of OEF is of potential clinical value for diagnosis and treatment monitoring of neurological conditions.

Multiple imaging modalities have been used to measure OEF in humans. Positron emission tomography (PET) with ¹⁵O-labeled tracers have been used extensively for OEF determination,¹⁵ although its use has diminished in recent years due to the need of on-site cyclotron, arterial sampling, and the exposure to ionizing radiation. Near-infrared optical spectroscopy has also been used to measure OEF of the brain, especially in neonates in whom the depth of light penetration is sufficient.¹⁶ More recently, with the advances of magnetic resonance imaging (MRI) methods, a number of techniques have been proposed to measure cerebral OEF without any exogenous agent,^17–31 thus potentially enabling the routine assessment of OEF in adult human patients. These techniques were based on the association between blood oxygenation level and MR properties of the blood, such as T₂,^17–22 susceptibility,^23–29 and phase.^30,31 These approaches allow the determination of brain OEF on standard MRI within just a few minutes.^{17–19,21–26}

However, one issue that remains to cloud the prospect of OEF as a useful biomarker in brain diseases is the fact that there exist considerable inter-subject variations in OEF within the healthy population.^32–35 For example, in a multi-center study involving 250 healthy subjects, an inter-subject coefficient of variation (CoV) of 16% in OEF was observed using a T₂-Relaxation-Under-Spin-Tagging (TRUST) MRI method.³³ Similarly, an inter-subject CoV of 13% in OEF has been reported by Ito et al. in a multi-center study of 70 healthy volunteers using the ¹⁵O-PET method.³⁴ These inter-subject variations are expected to obscure the interpretation of OEF data and weaken the power of OEF as a diagnostic marker. Therefore, it is important to investigate the physiological underpinnings of OEF variations among normal subjects.

End-tidal carbon dioxide (EtCO₂), the partial pressure of CO₂ in the exhaled air at end-expiration and closely related to the arterial CO₂ level (PaCO₂),^36,37 is known to have a major influence on cerebral hemodynamic parameters.^38,39 Moreover, it has been reported in a within-subject setting (e.g. before and after hypercapnia challenge) that EtCO₂ is negatively associated with cerebral OEF.^24,40–42 Therefore, we hypothesized that variations in OEF among normal subjects can be partly attributed to their differences in EtCO₂. In this work, we conducted two studies, one discovery study and one verification study, to examine the extent to which EtCO₂ can explain inter-subject variations in cerebral OEF. In addition, we investigated the potential effects of other physiological parameters, including systolic blood pressure (SBP), diastolic blood pressure (DBP), and heart rate (HR), on OEF.

Material and methods

Participants

Two studies were performed. Study 1 was a discovery study conducted on young healthy adults. Study 2 was a verification study conducted on a larger cohort of elderly subjects with normal cognitive functions. The study protocols were approved by Johns Hopkins University Institutional Review Board. Both studies were performed in accordance with the ethical guidelines of Johns Hopkins University Institutional Review Board, which were based on The Belmont Report, the Declaration of Helsinki, and The Nuremberg Code. All subjects gave written informed consents before participating in the study.

In Study 1, we recruited 10 young adult volunteers (24.0 ± 3.2 years old, ranging from 19 to 31 years old; 4 females, 6 males) via local advertisements. All subjects were in general good health according to self-report.

In Study 2, 25 cognitively normal elderly subjects (69.5 ± 6.7 years old, ranging from 59 to 83 years old; 18 females, 7 males) were recruited. All subjects enrolled in this study had a score of 0 in the Clinical Dementia Rating and a mean score of 27.6 ± 1.6 (mean ± SD) in the Montreal Cognitive Assessment, and were assigned the status of cognitively normal via a consensus between a neuropsychologist, a gerontologist, and a psychiatrist, based on their medical histories and performances on cognitive tasks. Subjects with any other medical, psychiatric, or neurological conditions that could affect cognition were also excluded.

Measurement of OEF

Whole brain OEF was evaluated using TRUST MRI. The details of TRUST have been described previously.^17,18 Briefly, TRUST MRI is based on the well-established relationship between blood oxygenation and blood T₂.⁴³ To isolate pure venous blood signal from surrounding tissues, TRUST utilizes radiofrequency (RF) pulses to label the incoming venous blood and acquires labeled and control images by alternating the RF pulses. Pair-wise subtraction of control and labeled images yields pure venous blood signal in the superior sagittal sinus (SSS), eliminating partial-volume effects. The T₂ of the blood signal was quantified by exploiting T₂-preparation with varying numbers of refocusing RF pulses, which is characterized by the effective echo time (eTE). Blood T₂ is then converted to venous oxygenation (Y_v) via known calibration plots.⁴³ Finally, OEF can be computed by

$$\mathit{OEF}=\frac{Y_a-Y_v}{Y_a}\times 100\%$$ (1)

where Y_a is the arterial oxygenation, and was assumed to be 98% in this work.

MRI experiments

In Study 1, the subjects were scanned on a 3T Siemens Prisma system (Siemens Healthcare, Erlangen, Germany). Each subject underwent three TRUST scans. A nasal cannula (Model 4000F-7, Salter Labs, Arvin, CA, USA) was used to sample the exhaled gas during the TRUST scans, and EtCO₂ was measured using a capnograph device (Capnogard Model 1265, Novametrix Medical Systems, Wallingford, CT, USA). To improve data stability, the OEF and EtCO₂ were averaged across the three TRUST scans.

In Study 2, the subjects were scanned on a 3T Philips Achieva system (Philips Healthcare, Best, The Netherlands). Each subject underwent one TRUST scan, among other MRI sequences. EtCO₂ of the subject during the TRUST scan was sampled by nasal cannula (Model 4000F-7, Salter Labs, Arvin, CA, USA) and recorded with a capnograph device (NM3 Respiratory Profile Monitor, Model 7900, Philips Healthcare, Wallingford, CT, USA). We also collected other physiological parameters, which were SBP, DBP, and HR, from each subject during the same visit as the MRI scan. SBP, DBP, and HR were measured at the brachial level in seated position using automatic devices (Patient Monitor Model 3150M, Invivo, Orlando, FL, USA; Omron Model BP742N, Omron Healthcare, Lake Forest, IL, USA), and were measured twice before the MRI scan and twice after the MRI scan. Values of SBP, DBP, and HR were averaged across the four measurements for each subject. In addition, the hematocrit level was measured from each subject through a blood draw.

In both studies, the MRI experiments were performed using a body coil for transmission and a 32-channel head coil for receiving. Foam padding was placed around the head to minimize motion during the MRI scans. TRUST was performed with the following sequence parameters: field of view = 220 × 220 × 5 mm³, repetition time = 3 s; inversion time = 1.02 s; voxel size = 3.44 × 3.44 × 5 mm³; labeling slab thickness = 100 mm; gap = 22.5 mm; four eTEs: 0, 40, 80 and 160 ms; and scan duration = 1.2 min. The echo time (TE) of TRUST was 6.5 ms in Study 1 on the Siemens scanner and 3.6 ms in Study 2 on the Philips scanner. The difference in TE was due to the slightly different implementations of the TRUST pulse sequence on the two MR vendors. A previous report on multi-vendor comparison of TRUST MRI has demonstrated that the performance of these two implementations is highly compatible.⁴⁴

Data processing

The TRUST MRI data were processed with in-house MATLAB (Mathworks, Natick, MA) scripts, following established procedures reported in the literature.¹⁷ Briefly, after motion correction, pair-wise subtraction between control and labeled images was conducted to obtain the difference images, on which a preliminary region of interest (ROI) was drawn encompassing the SSS. Four voxels inside the ROI with the highest signal intensities were selected as the final mask. Blood signals in the final mask were spatially averaged for each eTE and were fitted monoexponentially as a function of eTEs to obtain the T₂ of blood. Blood T₂ was then converted to Y_v using well-established calibration plots, considering hematocrit of the subject.⁴³ In Study 1, hematocrit was assumed to be 0.42 for males and 0.40 for females; whereas in Study 2, hematocrit used the subject-specific measured values. In addition, as a metric of the quality of Y_v data, the width of the 95% confidence interval of 1/T₂ was quantified from the monoexponential fitting procedure, and was hereafter referred to as ΔR₂. TRUST scans with ΔR₂ ≥ 5 Hz would suggest a poor data quality and were excluded from statistical analysis.

Statistical analysis

In Study 1, linear regression analyses were performed using OEF as the dependent variable and EtCO₂ as the independent variable. Age was included as a covariate, since the association between age and OEF has been reported in multiple studies.^33,45–47 Gender was not considered as a covariate because it has been shown to be not associated with OEF in recent literature.³³ A P value less than 0.05 is considered significant.

Based on the data in Study 1, power calculation was performed for Study 2. The power calculation used the slope and EtCO₂ standard deviation measured in Study 1. For the standard deviation of the dependent variable, OEF, we used results in Study 1 but also considered the fact that Study 2 was performed in elderly individuals who may have larger measurement noise and that the TRUST sequence was performed only once. The power calculation used the “Power and Sample Size Calculation” software⁴⁸ and revealed that the power to observe a significant association between EtCO₂ and OEF was 97%.

In Study 2, since we have multiple candidate modulators of OEF, step-wise linear regression analyses were performed to investigate the dependence of OEF on physiological parameters. In these analyses, OEF was the dependent variable and the candidate independent variables were EtCO₂, SBP, DBP, and HR. Age was always included in the regression model. The candidate independent variables were added to the regression model in a step-wise manner, starting from the most significant one until there were no more variables with a P < 0.05.

Results

Figure 1 shows a representative data set of TRUST MRI. Subtraction of labeled image from the control yields strong venous blood signal in the SSS, which decays with eTEs. The scatter plot shows the averaged blood signal as a function of eTEs as well as the monoexponential fitting results. Table 1 summarizes the physiological measurements.

Figure 1.

A representative data set of TRUST MRI. Subtraction of the labeled image from the control yields strong venous blood signal in the SSS, which decays with eTEs. The control and labeled images are only shown for eTE = 0 ms, while the zoomed-in views of difference images are shown for all four eTEs. The plot on the right displays averaged blood signal as a function of eTEs. The blood T₂ obtained from monoexponential fitting and the corresponding Y_v are also shown.

eTE: effective echo time.

Table 1.

Summary of physiological measurements.

Parameter	Study 1	Study 2
EtCO2 (mmHg)	38.4 ± 2.6	34.3 ± 4.3
SBP (mmHg)	−	145.5 ± 19.5
DBP (mmHg)	−	85.4 ± 12.3
HR (beats per minute)	−	70.8 ± 10.5
Hematocrit	−	0.41 ± 0.03

HR: heart rate; DBP: diastolic blood pressure; SBP: systolic blood pressure.

For data from Study 1, one TRUST scan of one subject was excluded due to a ΔR₂ ≥ 5 Hz, and the remaining two TRUST scans were used to compute the OEF of this subject. The OEF and EtCO₂ of the young subjects in Study 1 were 38.0 ± 3.7% and 38.4 ± 2.6 mmHg, respectively, which were in good agreement with previous reports on young healthy subjects.^32,41 Figure 2 shows a scatter plot of OEF and EtCO₂ across subjects. Univariate linear regression showed that OEF was negatively correlated with EtCO₂ (P = 0.02, R²= 0.52). The association between OEF and EtCO₂ remained significant (P = 0.04) after adding age as a covariate, as shown in Table 2. Individuals with a higher EtCO₂ tend to have a lower OEF, at a rate of −0.90 ± 0.36% per mmHg CO₂.

Figure 2.

Scatter plot of OEF and EtCO₂ for young healthy subjects (N = 10). The dashed line shows the fitting curve.

OEF: oxygen extraction fraction; EtCO₂: end-tidal carbon dioxide.

Table 2.

Multiple linear regression using OEF (%) as the dependent variable in young healthy subjects, R²= 0.59.

Variables	β ± SE	P
Age	0.33 ± 0.28	0.29
EtCO2	−0.90 ± 0.36	0.04

β: standardized beta coefficient; SE: standardized error; EtCO₂: end-tidal carbon dioxide.

For data from Study 2, all TRUST scans had ΔR₂ < 5 Hz. The OEF and EtCO₂ of the elderly subjects were 47.7 ± 6.8% and 34.3 ± 4.3 mmHg, respectively. A two-sample t-test assuming equal variance revealed that the elderly subjects in Study 2 had a lower EtCO₂ (P = 0.01) and higher OEF (P = 0.0002) than the young subjects in Study 1. These observations are in agreement with previous reports on these physiological variables.^45,46,49 The SBP, DBP, and HR of the elderly subjects were 145.5 ± 19.5 mmHg, 85.4 ± 12.3 mmHg and 70.8 ± 10.5 beats per minute, respectively. Figure 3 displays a scatter plot of OEF and EtCO₂ across subjects, showing a strong negative correlation (P = 0.0002, R²= 0.45). Table 3 summarizes the results of stepwise linear regression analyses. It was found that OEF is negatively associated with EtCO₂ (P = 2 × 10⁻⁶), but positively associated with age (P = 0.002) and SBP (P = 0.01). Importantly, EtCO₂ alone can explain about 50% of the inter-subject variations in OEF, which is similar to the findings in Study 1. Furthermore, age, EtCO₂, and SBP together can account for a total of 73% of the normal variations in OEF. HR and DBP did not enter the final stepwise model (P > 0.1). However, further investigation suggested that HR, DBP, and SBP contributed similar sources of variations to OEF, in that when only one of these three was included in the model (without the other two), their effect on OEF was significant (see Supplementary Table S1 and S2).

Figure 3.

Scatter plot of OEF and EtCO₂ for cognitively normal elderly subjects (N = 25). The dashed line indicates the fitting curve.

OEF: oxygen extraction fraction; EtCO₂: end-tidal carbon dioxide.

Table 3.

Stepwise linear regression using OEF (%) as the dependent variable in cognitively normal elderly subjects.

Variables	β ± SE	P	ΔModel R2
Age	0.40 ± 0.12	0.002	0.13
EtCO2	−1.19 ± 0.18	2 × 10−6	0.49
SBP	0.12 ± 0.04	0.01	0.11

β: standardized beta coefficient; SE: standardized error; EtCO₂: end-tidal carbon dioxide.

Discussion

This work demonstrated that across normal subjects, OEF is inversely correlated with EtCO₂. From the slope of the linear regression results, it appeared that one unit (i.e. mmHg) increase of EtCO₂ corresponds to approximately one unit (i.e. %) decrease in OEF. Furthermore, EtCO₂ alone can account for about 50% of inter-subject variations in OEF. These findings were reproducible in two different populations (young and elderly subjects) and on two different MR systems (Siemens and Philips). In addition, OEF is positively associated with age and SBP. By accounting for EtCO₂, age, and SBP, inter-subject variations in OEF can be reduced by 73%, which may significantly improve the sensitivity of OEF as a disease biomarker.

As an important index of the homeostasis of the brain’s oxygen metabolism, OEF has found clinical utility in a large number of diseases. For example, increased OEF has been demonstrated to be a strong and independent predictor of stroke risk in patients with cerebral arterial occlusive diseases.^1–4 Decreased OEF values have been reported in patients with mild cognitive impairment⁷ and patients with multiple sclerosis,⁸ while elevated OEF has been found in patients with moyamoya disease,⁵ hepatic encephalopathy,¹² or end-stage renal disease.¹¹ Alterations in OEF have also been reported in sickle cell disease patients, although it is still controversial whether OEF is elevated^5,14 or diminished.¹³ It is also possible that OEF can be used for treatment monitoring and thereby developing new interventions (i.e. endpoint in clinical trials). However, due to considerable inter-subject variations in OEF even in the healthy population,^32–35 the detection of disease-related abnormalities in OEF is mostly based on a group level. In this work, we demonstrated that by accounting for EtCO₂ alone, inter-subject variations in OEF can be reduced by about 50%. Given that EtCO₂ can be easily measured with a commercial capnograph device, recording of EtCO₂ during OEF measurements may be a cost-effective approach to enhance the sensitivity of detecting disease-related abnormalities on an individual level. Another approach to evaluate the effect size of EtCO₂ on OEF is to study the extent of EtCO₂ variations in daily life. Based on our data, EtCO₂ variations within the same scan session is 0.5 ± 0.4 mmHg (mean ± SD). A previous study reported that between-day variations in EtCO₂ were 1.06 mmHg.⁵⁰ EtCO₂ was found to increase during intensive exercise,^51,52 but returned to pre-exercise level a few minutes after exercise.⁵¹ Food consumption could also have an effect on EtCO₂. Patik et al. reported a slight increase in EtCO₂ (by about 2 mmHg) after fast-food meal,⁵³ while Zwillich et al. found no significant change in EtCO₂ after carbohydrate-rich or protein-rich meal.⁵⁴ Therefore, it is reasonable to expect that an individual’s OEF would change by 2–3% just by day-to-day variations.

The mechanism underlying the observed inverse correlation between OEF and EtCO₂ is likely related to the vasoactive effect of CO₂. CO₂ is known to be a potent vasodilator. Higher EtCO₂ is expected to result in a higher cerebral blood flow (CBF). In an intra-subject setting, it is well established from previous ¹⁵O-PET and MRI studies that increases in blood CO₂ content result in an augmentation in CBF.^38,41,49,55 Several prior reports have also noted a similar relationship in an inter-subject setting, in that there exists a positive association between EtCO₂ and CBF across individuals.^56,57 If brain metabolic rate is similar (or not as variable) across subjects, it can be readily derived that OEF will be lower in individuals with a higher EtCO₂, as OEF is related to the ratio between metabolism and CBF. There are also reports that a higher EtCO₂ may result in a lower metabolic rate of the brain due to a tissue acidosis effect of CO₂.^41,58,59 This effect would also result in a lower OEF. There is also a possibility that a lower OEF may result in decreased activity in chemoreceptor neurons in the brain that innervates the respiratory neurons in the brain stem, resulting in a reduced respiration and thereby a higher EtCO₂.⁶⁰ However, in studies of cerebral arterial occlusion patients, the results of EtCO₂ were mixed, with some showing slightly higher⁶¹ and others showing slightly lower⁶² EtCO₂ in occlusion patients.

The influence of EtCO₂ on other cerebral physiological parameters has also been reported. Besides the association between EtCO₂ and CBF,^56,57 it has also been suggested that, across subjects, baseline EtCO₂ is inversely correlated with cerebrovascular reactivity measured with BOLD MRI and CO₂ inhalation.⁶³

It should be emphasized that the present study reports physiological associations on an inter-subject level. Intra-subject correlations between OEF and EtCO₂ have been reported previously. For example, it is known that, when an individual’s arterial CO₂ is elevated by inhalation of hypercapnic gas or by breath-holding, OEF decreases as a result of CBF increase.^24,40–42 However, to our best knowledge, no prior studies have reported this relationship in an inter-subject setting. We note that it is not trivial to extend intra-subject observations to inter-subject, because inter-subject variations in physiological parameters are often caused by multiple factors. In the case of OEF, many other sources, such as basal metabolic rate, intake of caffeine, vasoactive medications, and so on, could have resulted in variations in OEF across individuals. Therefore, the observations in this study are novel in that it identifies EtCO₂ as the leading physiological factor that explains normal inter-subject variations in OEF.

Besides the negative correlation between OEF and EtCO₂, in Study 2, we also found that after adjusting for EtCO₂, OEF is positively associated with age and SBP. The increase in OEF with age agreed with previous reports using ¹⁵O-PET oximetry^47,64 or TRUST MRI.^33,45,46 Increased OEF with age is thought to be attributed to a decrease in CBF with age,^45–47 as well as additional effects of an elevation in cerebral metabolic rate of oxygen.^45,46 As for the blood pressure, it has been previously reported that OEF is positively associated with mean arterial blood pressure,^65,66 which is thought to be a compensatory response to a diminished CBF at high blood pressures.^65–67 It is worth noting that in univariate linear regression analyses, OEF showed no significant associations with age (P = 0.07) or SBP (P = 0.26), whereas the associations became significant after accounting for EtCO₂ (P = 0.002 for age; P = 0.01 for SBP). This finding provides an early indication that consideration of EtCO₂ reduces noise in the data and helps reveal the dependence of OEF on other physiological (and potentially pathological) factors.

We observed that elderly subjects in Study 2 had significantly lower EtCO₂ than the young subjects in Study 1. This is consistent with previous literature that EtCO₂ in subjects more than 50 years of age was significantly lower than young subjects less than 30 years of age.⁶⁸ Additionally, De Vis et al. reported that young subjects (aged 28 ± 3 years) had a mean EtCO₂ of 39 ± 3 mmHg and elderly subjects (aged 66 ± 4 years) had a significantly lower EtCO₂ of 35 ± 3 mmHg,⁴⁹ which agreed well with our observations.

There are a few limitations of the present work. First, the measurements of OEF in this work were on a whole brain level, without spatial information. Thus, it is not clear whether the association between OEF and EtCO₂ is region dependent. As new region-specific OEF techniques become available,^{19,20,22,26,30,31} spatial specificity of this relationship should be studied. However, it should be pointed out that the strengths of the TRUST MRI technique used in this study include its short scan time (1.2 min), excellent test–retest reproducibility,³² and scalability across multiple sites³³ and across different MR vendors.⁴⁴ A second limitation is that the studies performed in this work were focused on normal subjects. Future work is needed to investigate the extent to which this relationship can hold in patients with brain diseases. For example, in diseases where vascular reserve is exhausted or severely impaired such as in chronic obstructive pulmonary disease^69–71 and arteriostenosis,^62,72 the dependence of OEF on EtCO₂ will be attenuated or even abolished. Therefore, caution needs to be taken when applying the relationship reported in the present study in diseased populations. A third limitation is that blood pressure in this study was measured at seated position, while the MRI-OEF measure was obtained at supine position. There have been some reports that blood pressure measured at different positions could yield slightly different values.^73–78 We had chosen to conduct the blood pressure measurement at seated position to be consistent with standard clinical practice and to reduce patient burden. We expect that, as long as the measurement is always performed at seated position, the observed relationship will be applicable.

In conclusion, the present work suggests that across normal subjects, OEF is inversely correlated with EtCO₂, and about 50% of the inter-subject variations in OEF can be attributed to their EtCO₂ differences. We also found that OEF is positively associated with age and SBP. If EtCO₂, age, and SBP information can be obtained from the subjects, one can reduce the variation of OEF data by 73%, which is expected to substantially enhance the utility of OEF as a disease biomarker.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Institutes of Health grant R01 MH084021, R01 NS106711, R01 NS106702, R01 AG047972, R21 NS095342, R21 NS085634, P41 EB015909 and UH2 NS100588.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Authors’ contributions

DJ acquired the data, analyzed the data, and drafted the article. ZL, PL, and CX acquired the MRI data. SS, KH, and GP recruited subjects, acquired, and analyzed cognitive test data. SY, PR, and MA analyzed and interpreted the cognitive and clinical data. HL designed the study, guided the experiments and data analyses, and revised the article.

Supplemental material

Supplemental material for this paper can be found at the journal website: http://journals.sagepub.com/home/jcb