Boosting Schizophrenia Genetics by Utilizing Genetic Overlap With Brain Morphology

BACKGROUND: Schizophrenia is a complex polygenic disorder with subtle, distributed abnormalities in brain morphology. There are indications of shared genetic architecture between schizophrenia and brain measures despite low genetic correlations. Through the use of analytical methods that allow for mixed directions of effects, this overlap may be leveraged to improve our understanding of underlying mechanisms of schizophrenia and enrich polygenic risk prediction outcome. METHODS: We ran a multivariate genome-wide analysis of 175 brain morphology measures using data from 33,735 participants of the UK Biobank and analyzed the results in a conditional false discovery rate together with schizophrenia genome-wide association study summary statistics of the Psychiatric Genomics Consortium (PGC) Wave 3. We subsequently created a pleiotropy-enriched polygenic score based on the loci identi ﬁ ed through the conditional false discovery rate approach and used this to predict schizophrenia in a nonoverlapping sample of 743 individuals with schizophrenia and 1074 healthy controls. RESULTS: We found that 20% of the loci and 50% of the genes signi ﬁ cantly associated with schizophrenia were also associated with brain morphology. The conditional false discovery rate analysis identi ﬁ ed 428 loci, including 267 novel loci, signi ﬁ cantly associated with brain-linked schizophrenia risk, with functional annotation indicating high relevance for brain tissue. The pleiotropy-enriched polygenic score explained more variance in liability than conventional polygenic scores across several scenarios. CONCLUSIONS: Our results indicate strong genetic overlap between schizophrenia and brain morphology with mixed directions of effect. The results also illustrate the potential of exploiting polygenetic overlap between brain morphology and mental disorders to boost discovery of brain tissue – speci ﬁ c genetic variants and its use in polygenic risk frameworks.

Schizophrenia is a highly heritable complex brain disorder with a polygenic architecture, involving numerous common polymorphisms with small effects (1). While genome-wide association studies (GWASs) have identified hundreds of genetic variants associated with schizophrenia, together these explain only a fraction of its heritability (2). Identifying more variants will increase our understanding of the underlying disease mechanisms and improve genetic prediction, with potential for clinical utility. However, this is currently not feasible, as uncovering a significant portion of the heritability with conventional approaches will require sample sizes of more than a million individuals (3).
The complexity of the genetic architecture of schizophrenia has made it difficult to determine a genetic relationship with abnormal brain morphology phenotypes associated with the disorder (4). Patients with schizophrenia have on average a thinner cerebral cortex compared with healthy peers as well as reductions in cortical surface area (5). Subcortically, schizophrenia is associated with smaller hippocampal and amygdala volumes, larger lateral ventricles (6), and smaller cerebellum (7). In line with the clinical and genetic heterogeneity of schizophrenia, imaging studies have documented substantial brain structural heterogeneity (8), and specific regional deviations are shared by a only small percentage of individuals (9).
It is possible to boost the power of schizophrenia genetic studies by leveraging auxiliary genetic information contained in a related trait (10), such as brain morphology. Surprisingly, large-scale investigations have shown no or low genetic correlations between schizophrenia and brain structure (4,11). However, substantial genetic overlap may remain undetected by standard measures of genetic correlation owing to mixed directions of effects of shared genetic variants across the two traits, canceling each other out (12). Conditional false SEE COMMENTARY ON PAGE 258 discovery rate (cFDR) analysis uses overlapping variant associations, regardless of direction of effects, to re-rank the test statistics in a primary phenotype conditional on the associations in a secondary phenotype (10,13). The idea behind this is that in the presence of cross-trait enrichment, a variant with strong associations with both traits is more likely to represent a true association. This approach has been shown to allow for discovery of many novel variants (13), which have been found to replicate well (14,15). Further, owing to the distributed effects of genetic variants involved in brain morphology, the data can be analyzed more powerfully through a multivariate approach, capitalizing on the shared genetic signal across regional brain measures by considering the brain as an integrated unit (16), thereby providing greater statistical power and better alignment with the underlying biology than univariate approaches.
Here we characterized the polygenic overlap between schizophrenia and brain morphology, using a comprehensive set of brain morphology measures combined in a multivariate framework, and investigated whether this overlap can be leveraged to allow for discovery of common genetic variants contributing to brain-linked schizophrenia risk. We further sought to improve the polygenic prediction of schizophrenia by leveraging the newly identified genetic overlap.

Participants
We used data from the UK Biobank (UKB) population cohort, under accession number 27412. The composition, set-up, and data-gathering protocols of UKB have been described elsewhere (17). After preprocessing, our sample size for the primary neuroimaging analyses, excluding individuals with schizophrenia, was 33,735, with a mean (SD) age of 64.33 (7.49

Genetic Data Preprocessing
We used UKB v3 imputed data, quality controlled as described by the UKB genetics team (18). We carried out standard quality check procedures, including removing individuals with more than 10% missingness, lead single nucleotide polymorphisms (SNPs) with more than 5% missingness, with an INFO score below 0.8, and failing the Hardy-Weinberg equilibrium test at p = 1 3 10 29 . We set a minor allele frequency (MAF) threshold of 0.005, leaving 9,061,072 SNPs. This MAF was chosen in accordance with previous imaging genetics work with UKB data by our group (16). Extensive information on the TOP genetic processing is given in Supplement 1.

Image Acquisition
For the genetic analyses of neuroimaging data, we used UKB magnetic resonance imaging data released up to March 2020. T1-weighted scans were collected from four scanning sites throughout the United Kingdom, all on identically configured Siemens MAGNETOM Skyra 3T scanners (Siemens AG). The UKB core neuroimaging team has published extensive information on the applied procedures (19).
For the association of brain morphology with schizophrenia diagnosis, we used T1-weighted data from TOP, collected at the Oslo University Hospital on three scanners: a 1.5T Siemens MAGNETOM Sonata (Siemens AG), a 3.0T Signa HDxt (GE Healthcare), and a 3.0T GE 750 (GE Healthcare). For details on scanning protocols, see Brandt et al. (20) and Kaufman et al. (21).

Neuroimaging Data Preprocessing
We applied the standard recon-all-all processing pipeline of FreeSurfer v5.3, performing automated surface-based morphometry and subcortical segmentation (22,23). From the output, we extracted 175 global, subcortical, and cortical morphology measures. Table S1 in Supplement 1 contains all the measures included. We included both the left and the right hemisphere measures, if applicable.
For the imaging GWAS of UKB data, we first selected all individuals with White European ancestry, based on selfidentification as White British and similar genetic ancestry. A cross-ethnic analysis including 5737 non-White participants and yielding highly similar results is provided in Supplement 1.
We excluded anyone with an ICD-10 schizophrenia diagnosis, indicated by an F2 code; individuals with bad structural scan quality, indicated by an age-and sex-adjusted Euler number (24) 3 standard deviations below the scanner site mean; or individuals with a global brain measure 5 standard deviations from the sample mean. Finally, we removed one of each genetically related pair of individuals, defined by a threshold of 0.0625 determined by genome-wide complex trait analysis (25).
We followed a similar preprocessing procedure for associating brain measures with schizophrenia in TOP. White European ethnicity was based on self-report. DSM-IV diagnosis of schizophrenia was determined based on the Structured Clinical Interview for DSM-IV Axis I disorders. Controls were individuals without brain damage or history of a severe psychiatric disorder themselves or first-degree relatives.
For both samples, separately, we regressed out age, sex, scanner site, Euler number, and the first 20 genetic principal components from each brain measure. We further regressed out estimated intracranial volume for subcortical volumes, mean thickness for regional thickness measures, and total surface area for regional surface area measures. We then Genetic Overlap of Schizophrenia With the Brain applied rank-based inverse normal transformation (26) to the residuals.

Genetic Analyses
A univariate GWAS on each of the 175 processed brain morphology measures was carried out using the additive model in PLINK2 (27). We ran a multivariate GWAS on all preresidualized brain morphology measures using the Multivariate Omnibus Statistical Test (MOSTest). MOSTest yields one set of summary statistics that captures the significance of its association across all measures for each SNP. See van der Meer et al. (16) for an extensive description of MOSTest, its validation, and application to brain morphology data. The MOSTest software is openly available via https://github.com/precimed/ mostest.
We further ran a GWAS on nonmelanoma skin cancer for use as a negative control, given that this is a heritable trait (28) with no significant association with psychiatric disorders (29). We made use of the UKB data, categorizing all individuals with an ICD-10 code C44 as cases, restricted to White Europeans without brain disorders. We ran a logistic regression with PLINK2, covarying for age, sex, and 20 genetic principal components.
We used summary statistics from the Psychiatric Genomics Consortium Wave 3 (PGC3) schizophrenia GWAS (30) to check for genetic overlap between brain morphology and schizophrenia. We selected meta-analyzed summary statistics excluding the TOP sample, preventing sample overlap, containing 13,025,668 SNPs for 50,965 individuals with schizophrenia and 68,049 controls.
We conducted cFDR analysis, conditioning the PGC3 schizophrenia GWAS on the brain morphology GWAS, through the pleioFDR tool using default settings (https://github.com/ precimed/pleiofdr). PleioFDR builds on statistical pleiotropy, defined as the presence of associations with both traits (31). We set an FDR threshold of 0.01 as whole-genome significance, in accordance with recommendations. Independent significant SNPs and genomic loci were identified from the resulting GWAS summary statistics in accordance with the Functional Mapping and Annotation (FUMA) SNP2GENE definition (32). We also used FUMA (https://fuma. ctglab.nl/) to map SNPs to genes, using default settings. Given that cFDR output is incompatible with FUMA, we first identified lead SNPs according to FUMA definitions by clumping these summary statistics at an FDR threshold of 0.01 (https://github. com/precimed/python_convert) before using the FUMAs predefined lead SNPs option.
We carried out gene-based analyses through MAGMA v1.08 for the GWASs of brain morphology and schizophrenia (not for cFDR, as this is not valid), with a SNP-wide mean model and the 1000 Genomes Phase 3 EUR reference panel.
Hypergeometric tests were applied through FUMA GENE2FUNC function to calculate overlap between the sets of mapped genes and differentially expressed genes in each of the 54 tissues available in the Genotype-Tissue Expression (GTEx) v8 database. Differentially expressed genes were defined as genes with log2 transformed, normalized expression values (reads per kilobase million, zero mean) with p value # .05 after Bonferroni correction and absolute log fold change $0.58 in a given tissue, the default set by FUMA.

Polygenic Scoring
We calculated PGSs through PRSice v2 (33). We first clumped each of the summary statistics (skin cancer, brain morphology, schizophrenia, and cFDR) through PLINK, using an MAF threshold of 0.05, p-value threshold of 1, linkage disequilibrium threshold of 0.1, and distance threshold of 10 mb. We then ran PRSice, calculating PGSs for the individuals in TOP (and UKB replication sample) based on the PGC3 summary statistics, using the most significant lead SNPs from each of the four clumped summary statistics as determined by the indicated PGS threshold.
We used logistic regression to calculate the log odds and Nagelkerke R 2 of the PGSs at each threshold, with schizophrenia diagnosis as outcome in TOP and UKB separately. We included age, sex, batch, and the first 20 genetic principal components as covariates. See Supplement 1 for robustness analyses, whereby we created PGSs based on different GWASs to predict schizophrenia and intelligence.

Statistical Analyses
All downstream analyses were run in R v3.6.1 (R Foundation for Statistical Computing). Graphs were created using ggplot2 (34), and brain maps were created through pysurfer https:// pysurfer.github.io/. Coding applied for this study and an overview of the steps involved in the novel polygenic scoring approach, pleiotropy-enriched PGS (pleioPGS), are available at https://github.com/norment/open-science.

RESULTS
There was considerable overlap between the loci and genes identified in the PGC3 schizophrenia GWAS and those discovered through the multivariate analysis of 175 brain regions. There were 571 independent unique genetic loci associated with brain morphology at p , 5 3 10 28 . Of these, 32 loci were also among the 164 genome-wide significant loci (i.e., 19.5%) associated with schizophrenia ( Figure 1A). Through MAGMA (32,35), we found 1586 genes significantly associated with brain morphology after multiple comparisons correction, of which 252 were also among the 508 genes (i.e., 49.6%) identified through the schizophrenia GWAS. Supplements 2 and 3 list all discovered loci and genes.

Conditional FDR
The cFDR analysis, conditioning the schizophrenia GWAS on the multivariate brain morphology GWAS, further indicated genetic overlap ( Figure 1B). Through cFDR, we identified 428 independent significant loci, of which 267 were novel compared with the schizophrenia GWAS ( Figure S1 in Supplement 1). Supplement 2 lists information on all identified loci.
As follow-up of the main analyses using multivariate data, we checked for significant genetic overlap between schizophrenia and the individual univariate brain measures.

Functional Annotation
There were 50 lead SNPs with a Combined Annotation-Dependent Depletion score above 12.37, categorizing them as deleterious (36), versus 20 of the schizophrenia GWAS lead SNPs. There were no significant differences in distributions of minimum chromatin state scores or RegulomeDB scores (37), indicating that the SNPs identified through either cFDR or the schizophrenia GWAS were equally likely to be located in regulatory regions.

Gene Mapping
The lead SNPs identified by the cFDR analysis were mapped onto 1114 genes, of which 645 genes were not mapped through the schizophrenia GWAS. Among the top novel genes identified were RORA, encoding a nuclear receptor essential for development of the cerebellum (38) and for regulating circadian rhythm and immune function (39), hailed as a strong potential drug target (40); BCL11A, a transcription factor important for cortical development by regulating axon outgrowth (41), coupled to intellectual disability (42); and ALG14, encoding a catalyst of glycosylation, mutation of which is known to cause early, severe neurodegeneration (43) and behavioral problems (44). Information on each mapped gene is listed in Supplement 4.
Further, the set of genes mapped through the cFDR loci was significantly enriched for genes with known differential expression in the brain, particularly in the cerebellum and frontal cortex, but not in nonbrain tissues. This contrasts with the genes identified through the brain morphology GWAS (n = 2245) and schizophrenia GWAS (n = 660), which showed less specificity of expression in brain tissue ( Figure 2). As we have previously shown that the brain morphology GWAS does produce highly significant associations with genetic pathways central for brain development (16), we speculate that these findings speak to the nature of this test, capturing specificity, which is lost for the multivariate brain morphology data owing to the large number of genes identified.
The cFDR-identified genes were significantly overrepresented in Gene Ontology (GO) nervous system Miami plot, depicting the 2log 10 (p values) on the y-axis, of each genetic variant by their genomic position per chromosome on the x-axis, as identified through the multivariate brain morphology GWAS (top half) and schizophrenia GWAS (bottom half). The red dashed lines indicate the genome-wide significance threshold of 5 3 10 28 . Note that the y-axis is clipped at 2log 10 (p value) = 100. (B) Quantilequantile plot conditioning the schizophrenia GWAS on the multivariate brain morphology GWAS. Successive leftward deflection from the null distribution (dotted line) of the observed p-value distributions at more stringent significance thresholds (color-coded lines) indicates genetic overlap between the two traits. GWAS, genome-wide association study; SNPs, single nucleotide polymorphisms. Genetic Overlap of Schizophrenia With the Brain development gene sets and previously identified through GWASs of bipolar disorder, neuroticism, cognitive ability, and intelligence. This was also the case when rerunning these analyses on only the 645 novel genes ( Figures S3 and S4 in Supplement 1).
We additionally compared our gene list with two studies coupling transcriptomics data to GWAS summary statistics to boost identification of schizophrenia genes not linked to the PGC set of significant loci. Of the non-PGC identified genes, 16 were among the 62 additional genes reported by Gusev et al.

Polygenic Scoring
We hypothesized that a pleiotropy-informed approach, identifying SNPs that influence schizophrenia conditioned on brain morphology in independent samples, can reduce noise in the creation of PGSs owing to locus selection being based on two traits, lowering the likelihood of chance findings. To test this, we created PGSs to predict schizophrenia diagnosis for participants of the TOP clinical cohort. We first selected the most significant lead SNPs identified by 1) the multivariate brain morphology GWAS; 2) the schizophrenia GWAS; 3) the cFDR (schizophrenia/multivariate brain morphology) analysis; or 4) nonmelanoma skin cancer, as a control measure. We then calculated the PGSs for each of these sets of lead SNPs, based on the regression coefficients from the schizophrenia GWAS summary statistics (excluding the TOP cohort). We calculated the scores for a number of lead SNPs rather than for significance thresholds to allow for direct comparison between the approaches, at equal numbers of SNPs in each set. Here, the 200 SNPs threshold corresponds approximately to wholegenome significance, and the 30,000 SNPs threshold corresponds approximately to the number of lead SNPs identified in the schizophrenia GWAS at nominal p = .05, the threshold commonly applied for schizophrenia PGSs (30). As shown in Figure 3, at these thresholds, the cFDR-based pleioPGSs showed best prediction. Expressed as variance in liability, pleioPGSs explained 0.013 with the top 200 SNPs versus 0.008 for regular PGSs. With the top 30,000 SNPs, this was 0.077 versus 0.069. Figure S5 in Supplement 1 shows the explained variance in schizophrenia diagnosis per 1000 lead SNPs for each locus selection strategy through a sliding window approach. This graph illustrates that the variance explained by each subsequent window goes down strongest for the pleioPGS and the standard schizophrenia-based PGS, while there is no difference between the first and the last window of SNPs based on the skin cancer GWAS, indicating the lack of information captured by selecting loci based on this trait.
We subsequently sought to determine the robustness of these findings by rerunning the PGS analyses with different settings. In each case, the pleioPGS approach performed as good as or better than the conventional PGS approach. We first tested the use of a different replication sample: when calculating scores on 576 UKB participants with a schizophrenia diagnosis versus 317,139 individuals without brain disorders, the pleioPGS explained 0.022 (top 200 SNPs) and 0.063 (top 30,000 SNPs) variance in liability compared with 0.012 and 0.059 for regular PGSs. Second, we used an earlier subset of the schizophrenia GWAS to generate the PGSs, namely, the smaller (and thus likely noisier) PGC2 schizophrenia GWAS (2), with the TOP cohort excluded. With this, the explained liability in TOP was 0.017 and 0.057 for pleioPGS versus 0.009 and 0.050 for regular PGSs. Third, we checked how the approach generalizes to a different outcome measure: we first ran a GWAS of fluid intelligence on 100,000 UKB participants, followed by a cFDR analysis whereby we conditioned this GWAS on the multivariate brain morphology GWAS. We then created PGSs based on the resulting summary statistics and predicted intelligence in a random holdout set of 18,515 participants. Here, the pleioPGS approach explained 0.023 and 0.059 variance at the two thresholds versus 0.020 and 0.055 for regular PGSs. Therefore, with each of these three analyses, we saw that pleioPGS performed better at lower thresholds and that it reached a plateau earlier, indicating that it may allow for improved prioritization of genetic variants to study and predict brain traits (see Supplement 1 for more details and the full results).

DISCUSSION
With the present study, we provide evidence of polygenic overlap between schizophrenia and brain morphology that has not been detected with standard genetic approaches (4,11). We showed that this shared genetic architecture may be used to discover novel genetic variants that may play a role in schizophrenia, identifying 267 loci not discovered by the PGC and implicating schizophrenia-associated genes with high relevance for brain morphology. Lastly, we found that PGSs based on the shared genetic architecture between schizophrenia and brain magnetic resonance imaging morphology can be used for predicting schizophrenia. The present identification of genetic overlap between schizophrenia and brain morphology probably results from the application of analytical approaches more in line with the complex nature of the polygenic architecture of the brain and brain-related disorders such as schizophrenia. In particular, the distributed nature of genetic signal relevant to schizophrenia across the brain supports the use of multivariate methods (16), and the mixed directions of effects involved in the genetic overlap support the use of methods that do not rely on globally consistent directionality of effects (13,16). The high polygenicity of nearly all clinically relevant traits suggest that there are a large number of neurobiological mechanisms involved (3), and there is often large genetic overlap between many pairs of traits, while there is a relatively low genetic correlation (12). This indicates that some of these mechanisms strengthen both traits, while others inhibit the expression of one trait and strengthen the other. Given this, the approaches put forth here are able to better characterize the complex genetic architecture beyond genetic correlation, providing a more accurate model of the genetic relationship between brain-related traits and disorders (1), essential for our understanding of the etiology of schizophrenia and other mental disorders.
The high yield in terms of genome-wide significant loci from the cFDR analyses further support the value of considering genetic overlap between schizophrenia and brain morphology, providing important leads for understanding the neurobiological underpinnings of this disorder. Through the cFDR approach, we more than doubled the number of loci identified compared with the largest schizophrenia GWAS available to date (30). Conditioning a schizophrenia GWAS on a multivariate brain morphology GWAS allows for the identification of genetic determinants, particularly of brain-linked schizophrenia risk. This is further supported by functional annotation of these cFDR loci, with the mapped genes being linked to brain tissue-specific gene expression. A subset of the genes not reported by the PGC were also found in studies identifying schizophrenia risk genes on the basis of brain transcriptome data (45,46), thereby increasing the confidence in these findings through converging evidence. Further experimental interrogation of the identified loci may uncover biological insights that can lead to better understanding of the disease mechanisms and inform the development of novel treatment regimens. In addition, the specificity attained by our current approach may be valuable to parse schizophrenia genetics. Here we applied a brain morphology GWAS to boost the identification of nervous system genes associated with schizophrenia; we expect that other novel genes may be found when conditioning the schizophrenia GWAS on other system-specific GWAS, such as immune-related disorders or traits.
We further found that a pleioPGS has the potential to improve PGSs. We note that this was specifically the case for more stringent significance thresholds, in line with the reasoning that this approach brings about greater specificity, being particularly relevant for identifying the subset of loci that are most relevant for brain-linked schizophrenia risk. This approach appears robust, as the improved performance is also seen when using a subset of the main schizophrenia GWAS, i.e., less statistical power, as well as when using another independent schizophrenia test sample. We also found a similar pattern of results when analyzing another brain-related trait, fluid intelligence. This suggests that the concept underlying the pleioPGS may be widely applicable. This approach is based on the selection of loci identified by conditioning the genetics of one trait on that of another trait, leading to better out-ofsample prediction, possibly by reducing noise in GWAS signal and identifying more biologically relevant genetic variants. This is conceptually similar to a previous report that PGSs based on genetic variants specific to placental tissue are more predictive of schizophrenia than PGSs based on nonspecific variants (47). We are currently following up on this approach with the investigation of more pairs of traits.
It should be noted that the presence of statistical pleiotropy leveraged in this study to improve gene discovery and prediction can be partly driven by patterns of linkage disequilibrium and MAF (31). This warrants replication of the findings in new samples in the future. We also note the discrepancy in demographics between the elderly population-based UKB and younger clinical TOP sample, which may have contributed to larger error in the polygenic scoring analyses. Further, experimental studies are needed to clarify the molecular mechanisms underlying the current statistical associations. Last, we note that our aim was to illustrate the broad concept of leveraging genetic overlap to aid discovery and prediction. We thereby recognize that we did not explore other tools to achieve this, as that falls outside the scope of the current work, leaving that for future studies. Regardless of these limitations, our approach may be of particular clinical interest in the context of comorbidity; by conditioning a disease on comorbid traits or disorders, we may be able to improve the prediction of clinical subgroups.
To conclude, we provide evidence in this study for polygenic overlap between schizophrenia and brain morphology. This enabled us to identify 267 novel loci and 645 genes with specific relevance for the nervous system in schizophrenia. It also made it possible to develop enriched PGSs based on overlapping genetic architecture between brain morphology and schizophrenia. This suggests that the current approach can be used to attain greater specificity in identifying the underlying biological pathways and improve our understanding of the disease mechanisms of schizophrenia. This work was partly performed on the TSD (Tjeneste for Sensitive Data) facilities, owned by the University of Oslo, operated and developed by the TSD service group at the University of Oslo, IT Department (USIT) (tsd-drift@ usit.uio.no). Computations were also performed on resources provided by UNINETT Sigma2, the national infrastructure for high performance computing and data storage in Norway.

ACKNOWLEDGMENTS AND DISCLOSURES
DvdM conceived the study. DvdM and TK pre-processed the data. DvdM performed all analyses, with conceptual input from TK, AAS, OF, JMS, OAA, Genetic Overlap of Schizophrenia With the Brain and LTW. All authors contributed to interpretation of results. DvdM drafted the manuscript, and all authors contributed to and approved the final manuscript.
The data incorporated in this work were gathered from public resources and the Thematically Organized Psychosis study. The code is available via https://github.com/precimed (GPLv3 license). Correspondence and requests for materials should be addressed to d.v.d.meer@medisin.uio.no.
OAA has received speaker's honorarium from Lundbeck and Sunovion and is a consultant to HealthLytix. AMD is a founder of and holds equity in CorTechs Labs, Inc., and serves on its Scientific Advisory Board. He is a member of the Scientific Advisory Board of Human Longevity, Inc., and receives funding through research agreements with GE Healthcare and Medtronic. The terms of these arrangements have been reviewed and approved by the University of California San Diego in accordance with its conflict of interest policies. All other authors report no biomedical financial interests or potential conflicts of interest.