Reduced expression of N-methyl-D-aspartate receptor and calcium signaling genes in gray matter is associated with cancer-related cognitive impairment

Abstract

Purpose

Cognitive decline is common after cancer, but little is known regarding the etiology of this adverse effect, especially in terms of molecular mechanisms.

Methods

This prospective study obtained brain imaging and cognitive testing from 50 newly diagnosed women with primary breast cancer prior to any cancer treatment and 53 female controls. Participants completed up to seven assessments for a total time span of 9.7 ± 0.92 years. Imaging transcriptomics was used to measure the expression of genes in the brain involved in N-methyl-D-aspartate (NMDA) and calcium-mediated neurotransmission.

Results

GRIN2A, GRIN2B, and CACNA1C were significantly expressed in gray matter in both groups (R² ≥ 0.092, p ≤ 0.015). GRIN2A (t = −2.72, p = 0.018) was significantly lower in the cancer group compared to controls across timepoints. GRIN2A declined over time in patients, and this was significantly different compared to controls (χ2 = 9.73, p = 0.005). Cognitive scores were significantly lower in patients compared to controls (p ≤ 0.004). In patients, GRIN2A was significantly associated with cognitive performance over time (p ≤ 0.009).

Conclusion

These findings suggest that gene expression involved in neurotransmission is disrupted in the brain among patients with breast cancer and may contribute to cognitive changes. Our results provide novel molecular insights regarding the roles of non-CNS cancer pathology and treatments in the brain related to NMDA signaling and pro-survival/plasticity-related pathways. Our findings also point to potential treatments for cognitive effects of cancer.

Introduction

Cancer-related cognitive impairment (CRCI, known colloquially as “chemobrain”) is one of the most prevalent and burdensome symptoms, affecting up to 84.4% of survivors [1]. CRCI has been associated with diminished quality of life and increased cancer mortality [2, 3]. Notably, survivors have reported cognitive difficulties up to 20-years post-treatment [4]. Despite growing recognition of CRCI, its etiologies remain unclear. Inflammation, neurotoxic injury, suppression of neuroprogenitor cells, and accelerated aging are among the candidate mechanisms [5]. However, studies focused on molecular mechanisms have been limited. Challenges include the difficulty in translating findings from preclinical models and the incomplete correspondence of peripheral assays with the central nervous system environment.

Imaging transcriptomics is a non-invasive, in vivo method for examining how gene expression differs across brain regions, revealing molecular features linked to disease-related changes in brain structure and function [6]. Unlike single nucleotide polymorphism (SNP) studies, imaging transcriptomics offers more direct and functional insights by linking gene expression to brain organization. CRCI is consistently associated with abnormalities in large-scale brain networks that support integrated cognitive functions such as executive control and memory [7–9]. These networks uniquely rely on glutamate receptor and calcium-mediated neurotransmission. N-methyl-D-aspartate receptors (NMDARs) are glutamate receptors that flux high levels of calcium to support the persistent neuronal firing required to maintain advanced cognition, such as executive function [10]. Calcium drives cyclic adenosine 3′,5′-monophosphate production in a feedforward signaling cascade that is essential for synaptic plasticity and transcription [11].

GRIN2A and GRIN2B genes encode NMDAR -GluN2A and -GluN2B subunits, respectively [10]. CACNA1C encodes CA_v1.2 calcium channels and cAMP-activating proteins [12]. Dysregulated activation of NMDAR/calcium signaling pathways is associated with various cognitive disorders and neurodegenerative diseases [10, 13]. Abnormal calcium signaling is also associated with increased neuroinflammation [10], which is putatively elevated in CRCI [14]. In our unique, 10-year prospective, longitudinal study, we measured GRIN2A, GRIN2B,and CACNA1C expression in gray matter. We hypothesized that the expression of these genes across time would significantly differ in patients compared to non-cancer controls and would affect cognitive function. Given the novelty of these analyses and the complex interplay of biological systems under investigation, we explored nonlinear relationships and did not make a priori predictions regarding direction.

Methods

Participants

We prospectively enrolled 50 females newly diagnosed with primary breast cancer (stages 0 to IIIA) prior to any cancer treatment (surgery with anesthesia, chemotherapy, radiation) and 53 non-cancer female controls. Participants completed up to seven different assessments for a total time span of 9.7 ± 0.92 years, on average (Table 1). In the first phase of the study, three assessments occurred: at enrollment, 1 month after completing adjuvant chemotherapy treatment (or yoked interval for the control group), and again 1 year later. Participants were recruited through physician referrals and community-based advertisements. During the second phase of the study, we annually re-evaluated a consecutive sampling of participants (N = 35 per group) for the duration of the project, resulting in an additional four assessments (Table 2). The sample size reduction reflected resource limitations. The second phase sample did not differ significantly from the original in terms of demographic or clinical characteristics at baseline (p ≥ 0.852). Participants were excluded for history of chronic health conditions affecting cognition, major sensory deficits, or limited English proficiency. Demographic and medical history data were obtained from self-report. This study received Stanford University Institutional Review Board approval, followed the Declaration of Helsinki ethics, and obtained written informed consent from all participants.

Table 1.

Demographic data at enrollment

	Patients N = 50	Controls N = 53	Statistic	p value
Age (years)	49.48 ± 8.9	49.93 ± 9.9	t = 0.113	0.910
Education (years)	16.14 ± 2.9	16.91 ± 2.1	t = 1.60	0.114
Postmenopausal	42%	49%	X2 = 0.503	0.478
Stage at diagnosis (0, IA, IB, IIA, IIB, IIIA)	0%,15%,0%,33%, 38%,14%
Radiation therapy	64%
Hormone blockade therapy	72%
Number of chemotherapy cycles	Median = 6, IQR = 4, range = 4 to 20
Chemotherapy regimens	ACT 49%, TC 39%, CAF 4%, FEC 4%, AC 4%

A doxorubicin (Adriamycin®), C cyclophosphamide, T paclitaxel (Taxol®), F 5-fluorouracil, E epirubicin

Table 2.

Number of participants at each assessment. The time in months represents the average interval between the assessment and baseline. There was a longer interval between the third and fourth assessments than planned due to the COVID-19 pandemic which then required a slight acceleration of the final two assessments. The longer interval between the fourth and fifth assessments was due to relocations of the principal investigators

Timepoint	1	2	3	4	5	6	7
Time in months	0	6	18	75	99	109	117
Chemo	50	36	42	28	22	17	9
Control	53	44	44	37	31	26	19

Imaging transcriptomics

A high-resolution, anatomic brain MRI scan was obtained for each participant: TR = 8.5, TE = minimum, TI = 400 ms, FOV = 22 cm, slice thickness = 1.5 mm, 124 slices, 256 × 256, and scan time = 4:33 min using a 3 T GE Signa HDx whole body scanner (GE Medical Systems, Milwaukee, WI). Additional scans were collected that are not reported here.

We extracted gray matter volumes from the MRI for each participant using voxel-based morphometry [15]. Imaging transcriptomic analysis was carried out using Multimodal Environment for Neuroimaging and Genomic Analysis (MENGA) [16], which utilizes the Allen Human Brain Atlas (AHBA). The gray matter volume was resampled into AHBA coordinates with a 5-mm resolution. The expression data for each of the six AHBA donors, for each gene, was sampled from each of 169 brain regions defined by AHBA structure-level parcellation implemented via MENGA’s internal library. All images were registered to MNI ICBM152 space (1 mm isotropic) prior to analysis. MENGA resampled each image into each AHBA donor’s coordinate system using the MNI coordinates provided by the Allen Institute for each tissue sample. Regions were excluded if image or genomic data were missing for any donor, resulting in 89 of 169 regions with complete data available for regression. Principal components analysis was performed on the 89-region by 6-donor expression matrix for each gene to identify components explaining at least 95% of between-donor variance. The number of components retained was 2 for GRIN2A (cumulative variance = 95.1%) and GRIN2B (cumulative variance = 96.6%), and 5 for CACNA1C (cumulative variance = 98.0%), with degrees of freedom of (2,86) and (5,83) respectively for the subsequent regression. The component scores were then entered as independent variables into a multiple linear regression with the spatially corresponding gray matter volumes as the dependent variable.

Probe-to-gene mapping accuracy was assessed using the MENGA platform [16]. For each gene, probe accuracy was quantified as the average Pearson correlation coefficient between the selected probe and all other available probes for that gene, computed across all six AHBA donors, providing a measure of between-probe spatial consistency. Expression values for each probe were normalized to z-scores within each donor independently, using the within-donor mean and standard deviation computed across all available brain tissue samples (~ 617 samples per donor on average) as normalization factors. A data-driven selection procedure identified the probe with the most symmetric and least negatively skewed z-score distribution, pooled across all six donors, as the representative probe for that gene and the same probe was applied across all donors. The selected probes and their between-probe consistency values were as follows: GRIN2A, probe A_24_P381844 (r = 0.922, 5 candidate probes); GRIN2B, probe A_23_P151264 (r = 0.869, 3 candidate probes); and CACNA1C, probe A_23_P373031 (r = 0.828, 3 candidate probes). Gene expression reliability across donors was measured using an intraclass correlation coefficient (ICC).

Cognitive assessment

We administered Trails 1 and 5, referred to as Trails A and B, of the Comprehensive Trail Making Test. These tests measure processing speed, attention, and executive function [17]. Other cognitive tests were also administered but we a priori selected Trails A and B for this study to focus on executive function and to reduce comparisons. Executive function is one of the most profoundly affected cognitive domains in CRCI, and impairments in this area are among the most debilitating [5, 18].

Data analysis

Data were visualized for normality and outliers. Missing data were assessed using Little’s MCAR test and visualizations, which indicated missing at random. We fit generalized additive models (GAMs) using penalized regression splines, with smoothing parameters estimated by restricted maximum likelihood. GAMs are well-suited for irregular or sparse longitudinal data [19]. They stabilize estimates at timepoints with fewer observations by borrowing strength from neighboring values through smooth basis functions. The distributions and trajectories of the brain transcriptome are unknown. Thus, we modeled potentially nonlinear changes over time, using spline functions with a restricted basis dimension (k = 5) to limit the complexity of the smooth and prevent overfitting in sparsely sampled timepoints. Models included a random intercept as well as covariates for age, education, and gray matter volume.

We fit separate smooth functions of time for each group, allowing nonlinear outcome trajectories to vary flexibly between groups. To determine group-by-time effect, we refit the model using maximum likelihood and compared it to a model that included a shared smooth for time to test a group-by-time interaction effect using a likelihood ratio.

To explore the effects of gene expression on cognitive performance, we fit GAMs predicting Trails A/B score as a function of genes that showed significant main effects of group or group by time, with age and education as covariates. These GAMs were fit only in patients.

All analyses were performed in the R Statistical Package (v4.5.1) with p < 0.05. All p values were corrected using the Benjamini-Hochberg false discovery rate (FDR) within each family of tests (gray matter R², gene GAMs, gray matter GAM, cognitive GAMs, and correlation GAMs). For GAMs, only the main effect p values were corrected.

Results

As shown in Table 3, GRIN2A/B and CACNA1C were significantly expressed in gray matter in both groups (R² ≥ 0.092, p_FDR ≤ 0.015). Gene expression showed high reliability (ICC ≥ 0.829).

Table 3.

Correlation of gene expression with gray matter volume across participants. P values were corrected for false discovery rate (FDR)

Gene	Cancer			Control
	GRIN2A	GRIN2B	CACNA1C	GRIN2A	GRIN2B	CACNA1C
R2	0.377	0.094	0.477	0.372	0.092	0.476
F	26.01	4.45	15.13	25.45	4.38	15.08
p (FDR)	< 0.001	0.015	< 0.001	< 0.001	0.015	< 0.001
ICC*	0.941	0.956	0.829
ICC SD	0.014	0.013	0.043

ICC intraclass correlation coefficient, SD standard deviation

^*ICC values are computed across Allen Human Brain Atlas donors (i.e., gene expression reliability), not for sample groups

GRIN2A (t = −2.72, p_FDR = 0.018) was significantly lower on average in the cancer group compared to controls across timepoints (Fig. 1). CACNA1C was also lower but did not survive FDR correction (t = −2.11, p = 0.036, p_FDR = 0.060). GRIN2A declined over time in patients, and this was significantly different compared to controls (χ2 = 9.73, p_FDR = 0.005). Change in GRIN2B (χ2 = 0.00, p_FDR = 1) and CACNA1C (χ2 = 0.973, p_FDR = 0.393) were not different between groups. Age and gray matter volume contributed significantly to all models (F > 3.72, p < 0.013). Higher education was associated with increased GRIN2A/B expression (p < 0.047).

Fig. 1.

Longitudinal trajectory of neurotransmission-related genes in gray matter and gray matter volume. Expression of GRIN2A was significantly lower (t =  − 2.72, p_FDR = 0.018) and declined significantly over time (χ2 = 9.73, p_FDR = 0.005) in patients compared to controls. GRIN2B was not different between groups and did not change significantly over time. CACNA1C was marginally lower in patients compared to controls (t =  − 2.11, p_FDR = 0.060) and longitudinal trajectories did not differ between groups. Total gray matter volume was not different between groups (t =  − 1.41, p_FDR = 0.318) and did not change significantly over time. Plots show the partial effects controlling for all other predictors in the model. The solid line indicates the estimated smooth function for the generalized additive model with the shaded area representing the 95% confidence interval

Gray matter volume was lower in patients but did not differ significantly between groups (t = −1.41, p_FDR = 0.318) nor did it change significantly over time in either group (p_FDR > 0.529, Fig. 1).

Trails A (t = −4.29, p_FDR < 0.001) and Trails B (t = −3.14, p_FDR = 0.004) scores were significantly lower on average in patients compared to controls across timepoints (Fig. 2). Performance trajectories did not differ between groups (χ2 < 0.157, p_FDR > 0.721). Higher education was associated with higher performance in both models (p < 0.003) and age contributed to the Trails A model in a nonlinear manner (F = 5.89, p = 0.001).

Fig. 2.

Longitudinal trajectories of cognitive performance. Patients demonstrated significantly lower scores on the Trails A (t =  − 4.29, p_FDR < 0.001) and B (t =  − 3.14, p_FDR = 0.004) tests of executive function compared to controls but trajectories of performance did not differ between groups. Plots show the partial effects controlling for all other predictors in the model. The solid line indicates the estimated smooth function for the generalized additive model with the shaded area representing the 95% confidence interval

In patients, a model of Trails A performance that included gene expression and demographics (age and education) showed significantly higher goodness-of-fit compared to a demographics only model (χ2 = 22.19, p_FDR = 0.003). GRIN2A (F = 4.39, p = 0.009) and age were significant predictors (F = 4.82, p = 0.003). The partial effects of GRIN2A expression and age on Trails A performance were nonlinear (Fig. 3).

Fig. 3.

Predictors of cognitive performance in patients with breast cancer. In patients, GRIN2A expression (F = 4.39, p = 0.009) and age were significant predictors (F = 4.82, p = 0.003) of Trails A performance. GRIN2A (F = 6.25, p < 0.001), age (F = 5.69, p = 0.001), and education (t = 4.20, p < 0.001) were significantly associated with Trails B performance. Plots show the partial effects controlling for all other predictors in the model. The solid line indicates the estimated smooth function for the generalized additive model with the shaded area representing the 95% confidence interval

For Trails B, the model that included gene expression demonstrated significantly higher goodness-of-fit compared to a demographics only model (χ2 = 37.38, p_FDR < 0.001). GRIN2A (F = 6.25, p < 0.001), age (F = 5.69, p = 0.001), and education (t = 4.20, p < 0.001) were significantly associated with Trails B performance over time. The partial effects of age and GRIN2A were nonlinear (Fig. 3).

To determine if the effects of GRIN2A were domain specific, we conducted exploratory, post hoc GAMs with the additional cognitive tests that were not selected a priori for examination in this study. These included the Rey Auditory Verbal Learning Test Immediate and Delayed Recall (verbal declarative memory) [20], and the Controlled Oral Word Association test (verbal fluency) [21]. None was significant (F ≤ 1.85, p ≥ 0.145, uncorrected).

Discussion

GRIN2A/B encode NMDAR subunits which are ionotropic glutamate receptors critical for excitatory neurotransmission. NMDARs are unique from other glutamate receptors in several ways that impact cognitive functions. They allow passage of calcium in addition to sodium and potassium ions. Additionally, they are both ligand- and voltage-gated, requiring the coincidence of presynaptic glutamate and postsynaptic depolarization to pass calcium, which supports long-term synaptic information storage [22]. NMDARs have slower and longer-lasting amplitudes compared to other glutamate receptors, allowing sustained neural activity over time in the absence of sensory stimuli [10, 23]. These properties make NMDARs critical for executive skills, consistent with our findings that GRIN2A was significantly associated with Trails A/B performance.

Trail Making Test performance is frequently impaired in patients with breast cancer [24, 25]. Executive dysfunction is particularly problematic for cancer survivors given its strong association with treatment compliance, health behaviors, and fall risk, among others [26–29]. Our results indicated that Trail Making Test performance was lower, on average, in patients compared to controls. This group effect does not suggest a deficit at every assessment. However, a post hoc t-test revealed that patients showed significantly lower Trails A (t = −2.19, p = 0.031) performance at pre-treatment baseline. Trails B was also lower but not significantly so (t = −1.72, p = 0.090). Without assessments obtained prior to disease onset, the source of baseline group difference cannot be determined with certainty. Lower scores in patients may reflect pre-existing cognitive differences or early disease-related effects of tumor biology, systemic inflammation, or psychological distress associated with diagnosis.

The partial associations between GRIN2A and cognitive performance were nonlinear, which is not surprising given that biological systems tend to be complex. Associations showed biphasic or inverted U-shaped relationships, which could suggest that low levels of gene-brain co-localization may not be sufficient to influence downstream pathways, so cognition would not change until expression passed a certain point. Once above the threshold, effects may accelerate. At higher co-localization, the system may reach a plateau, so additional increases have diminishing or no effect. Certain levels of co-localization could be compensatory, while other levels of co-localization reflect dysregulation. There is likely an optimal range of gene expression in the brain that does not conform to traditional linear measurements.

Post hoc analyses provided preliminary evidence that the association between GRIN2A and cognitive performance may be domain specific. GRIN2A was significantly associated with Trails A and B performance, measures of processing speed and executive function, but not with declarative memory or verbal fluency. These findings should be interpreted cautiously given the limitations in the scope of the cognitive battery and sample size, and replication in studies using more comprehensive neuropsychological assessments is warranted.

Cancer treatments suppress neuroprogenitor cells while aging and inflammation disrupt mechanisms that protect neurons from the toxic effects of dysregulated calcium levels [30, 31]. This may result in increased neuron death (i.e., decreased gray matter), and, thus, lower gene expression. Prior longitudinal studies of gray matter volume in patients with CRCI indicate acute atrophic changes followed by recovery [32, 33]. We observed that total gray matter volume showed an overall decreasing trend but was not significantly lower compared to controls. Prior studies involved a much shorter follow-up timeframe compared to ours (1–2 years vs. 10 years) but included regional analyses. We noted that gray matter volume contributed significantly to all three gene models. The partial relationship between gray matter and gene expression was nonlinear. Thus, gray matter changes may help explain differences in gene expression, but not sufficiently, or they may be driven by regional fluctuations that were not evaluated here.

GRIN2B did not differ between groups before or after FDR correction, nor did it change over time in either group. Although this gene was significantly and reliably expressed in gray matter, it was associated with the lowest variance explained (9.2%–9.4% vs. 37%–48%). However, the lower co-localization of GRIN2B compared to GRIN2A is expected given that GluN2B receptors dominate during early development but are gradually replaced by GluN2A receptors with age [34, 35]. GluN2B receptors continue to play critical roles in synaptic plasticity and neurotransmission throughout the lifespan but have different trafficking and maintenance mechanisms in mature synapses compared to GluN2A [36].

Pathological imbalances between GluN2A/B subunit interactions, localization, and trafficking are believed to play critical roles in several neurodegenerative conditions [37]. Functional synaptic NMDARs (highly enriched in GluN2A in mature neurons) support pro-survival and plasticity-related signaling pathways. When synaptic GluN2A is lost, the overall NMDAR activity becomes dominated by extrasynaptic NMDARs, which include GluN2B subunits. Extrasynaptic NMDARs are mechanistically linked to neuronal dysfunction and cell death pathways, often by suppressing the pro-survival signaling of their synaptic counterparts [37, 38]. Reduced GRIN2A co-localization in gray matter may result in fewer available GluN2A subunits. Lower activity of pro-survival and plasticity-related signaling pathways results in negative feedback loops that suppress GRIN2A activity in gray matter.

GluN2A/B subunits also interact with different scaffolding proteins, which determines their stability and location in the mature neuron. If a disease or injury affects the anchor proteins responsible for keeping a receptor at the synapse (e.g., PSD-95 binding sites), GluN2A would be the primary target because it has been trafficked in to define the mature synapse [36, 39]. GluN2B has a mechanism for internalization and, as noted above, is largely extrasynaptic [40] and thus might be affected differently or less acutely by an impairment in the GluN2A-anchoring system. There is evidence showing that chemotherapies, including those used to treat breast cancer, functionally impair the PSD-95 complex and PSD-95 has been proposed as a mechanistic target for treating CRCI [41, 42].

In addition to supporting PSD-95 as a target for treating CRCI, our findings provide other translational insights for possibly addressing CRCI in the future. For example, neuron-derived exosomes (NDEs) have the potential to act as intervention targets for neurodegeneration by delivering neuroprotective cargo such as microRNAs that promote repair [43]. Integrating transcriptomic data with exosomal microRNA cargo profiling would enable pathway-level inference, indicating whether NDE microRNAs regulate the same NMDA/calcium signaling genes disrupted in CRCI.

This study represents the first investigation of brain transcriptome changes associated with CRCI. Other innovative aspects of this research include the pre-surgical baseline and the approximately 10-year longitudinal span of follow-up assessments. However, there are also several limitations that should be considered. Attrition was substantial and missing not at random (MNAR) is an inherent possibility in any longitudinal study of this duration. Importantly, participants in this sample were long-term survivors in good health, and dropout was attributable to logistical factors including relocation and the sustained participation burden of a nearly 10-year study rather than disease-related outcomes. This is more consistent with the MAR assumption than with an outcome-contingent missing data process. Nonetheless, sensitivity analyses under explicit MNAR assumptions are recommended for future larger-scale investigations. The trajectory estimates at the sparsest timepoints (T6 and T7) carry greater uncertainty relative to earlier assessments, as reflected in the widened confidence intervals in Fig. 1, and should be interpreted with appropriate caution. Even though the AHBA transcriptomic atlas is the most comprehensive tool to date [6], it was developed based on only six donors. The AHBA atlas is predominantly male and there currently are no sex-balanced brain transcriptome atlases available despite well-known sex differences in gene expression. We utilized normalized expression values, which removes donor-level differences in absolute signal [44]. However, development of sex-balanced brain transcriptomic atlases should be a priority for the field. The atlas includes expression patterns for thousands of brain regions but does not cover the entire brain. Genes were selected based on specific hypotheses and with consideration of statistical power, but other genes may also be relevant for CRCI. Similarly, we chose to focus on specific neuropsychological tests, but other assessments may yield different results. We did not have sufficient samples to examine regional effects. The patient group received heterogeneous treatments including diverse chemotherapy regimens, radiation, and hormone blockade therapy which may independently contribute to our observed results. The present study was not designed or powered to separate these contributions. Future research involving larger groups should independently replicate our findings.

Despite these limitations, our findings that co-localization of neurotransmission-related genes in gray matter is reduced in patients with breast cancer align with prior evidence on the roles of these genes in brain function and neurodegeneration. These results provide novel insights into the potential molecular mechanisms underlying CRCI, addressing a critical gap in the current literature. Furthermore, this study establishes a strong foundation for ongoing and future investigations into the effects of cancer and its treatments on the brain transcriptome. Patients diagnosed with cancer experience cognitive deficits because of the tumor biology and the additive negative effects of cancer treatments on their cognition. These longitudinal data present a unique opportunity to understand the trajectory of CRCI long-term and its underlying molecular mechanisms. This research also offers broader insights into cognitive decline seen in other disorders where there is an interaction between gene expression, aging, disease pathophysiology, or stressors/injuries.

Author contribution

Conceptualization: Shelli Kesler, Oxana Palesh; Methodology: Shelli Kesler, Oxana Palesh, Douglas Blayney, Alexa De La Torre Schutz, Oscar Franco-Rocha; Manuela Kogon, Leah Tolby, Ruth Nyagaka; Formal analysis and investigation: Shelli Kesler, Oscar Franco-Rocha, Alexa De La Torre Schutz; Writing—original draft preparation: Shelli Kesler, Oscar Franco-Rocha, Sarah Braun, Manuela Kogon; Writing—review and editing: Shelli Kesler, Oscar Franco-Rocha, Manuela Kogon, Sarah Braun, Leah Tolby, Alexa De La Torre Schutz, Douglas Blayney, Oxana Palesh; Funding acquisition: Shelli Kesler, Oxana Palesh; Resources: Shelli Kesler, Oxana Palesh; Supervision: Shelli Kesler, Oxana Palesh, Douglas Blayney.

Funding

This work was supported by research grants from the National Institutes of Health (1R01CA226080, 1R01CA172145, 2R01CA172145). The funder did not play a role in the design of the study; the collection, analysis, and interpretation of the data; the writing of the manuscript; and the decision to submit the manuscript for publication.

Data availability

The derived data that support the findings of this study are available from the corresponding author, upon reasonable request. Raw data (i.e., MRI scans) are not available due to data sharing and privacy restrictions.

Declarations

Ethics approval

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Institutional Review Board of Stanford University (protocol #8123). All participants provided written informed consent.

Competing interests

The authors declare no competing interests.