Functional genomics and systems biology in human neuroscience

Abstract

Neuroscience research has entered a new phase of key discoveries in the realm of neurogenomics due to strong financial and intellectual support for resource-building and tool development. The previous challenge of tissue heterogeneity has been surmounted by the application of techniques that can profile individual cells at scale. Moreover, the ability to perturb genes, gene regulatory elements, and neuronal activity in a cell type-specific manner has been integrated with gene expression measures to uncover functional underpinnings of the genome at a systems level. While these insights have necessarily been grounded in model systems, now is the opportunity to apply these approaches in humans and human tissue due to further advances in human genetics, brain imaging, and tissue collection. We acknowledge that there will likely always be limits to the extent to which we can apply our genomic tools developed in model systems to human neuroscience; however, as we describe here, the neuroscience field is now primed with the optimal foundation for tackling this ambitious challenge. Importantly, the application of systems-level network analyses to these datasets will facilitate a deeper appreciation of human neurogenomics that cannot otherwise be achieved from directly observable measures.

Introduction

The human brain is considered one of the final frontiers in the biological sciences, and functional genomics and systems biology can provide unique insights into molecular mechanisms at genome and brain-wide scales. Over the last decade, the (Brain Research Through Advancing Innovative Neurotechnologies) BRAIN Initiative and the Human Brain Project (and others; see below) have spurred innovation into the study of basic mechanisms and translational approaches to understand the nervous system. These investments have aided in fast-tracking neuroscience research to catch up with, and arguably surpass, other ongoing long-term research programs such as those in cancer. Since we last discussed the pace of functional genomics and systems biology in neuroscience almost 15 years ago¹, there has been remarkable progress in both tool development and discoveries. For example, deep transcriptomic mapping at the cell-type level has been carried out to generate brain atlases as well as detailed brain transcriptomic datasets from hypothesis-testing studies. These cell-type genomic datasets have been linked with circuit mapping or physiological measures and genomic tools have been harnessed to directly manipulate brain circuits or behavior (e.g.,²). However, many of these insights have been derived from and applied to model systems such as rodents or non-human primates. Because we cannot directly experiment on humans in the same manner as in model systems, the direct applicability of many of these findings to the human brain remains unknown. Thus, the neuroscience field needs to be more thoughtful about how we leverage model system data to understand human brain function and consider ways we can use summative network analysis to tease additional meaning from available human snapshot datasets.

As detailed below, our understanding of the human brain is slowly being updated and refined from cell types to circuits to behavior by embracing the windfall of technological resources that have become available in the last decade. Advances in genomics have been integrated with other approaches such as viral tools or electrophysiological measurements in single cells or circuits across brain regions. Foundational improvements in genomics have been derived from increases in the length of reads in sequencing data. For example, longer sequencing reads can be harnessed to determine the impact of gene isoform usage in cell type function in an RNA-sequencing experiment. The cost of sequencing per byte of information has also decreased, yielding greater depth of data for approachable costs viable to single labs. As another example, this has resulted in greater numbers of whole genomes being sequenced with improved insights into a number of brain disorders using genome wide association studies (GWAS)^3–5. The isolation of single cells, whether by flow cytometry or droplet-based methods and followed by any number of genomic approaches, has underscored the complexity of the human brain at cellular resolution^2,6. These findings include characterization of the heterogeneity of rare cell types such as vascular cells or microglia^7,8, cell types or cell states related to disease^8–12, the relationship between chromatin state and gene expression at cellular resolution^13,14, the identification of human-specialized cell types^15–17, developmental lineage information with or without cellular barcoding^18–20, and the presence of somatic single nucleotide variants in single neurons^21,22.

Unlike model systems or even other tissues in humans, the human brain is challenging to access in living individuals (with some exceptions). Thus, neuroscientists have gone to great lengths to 1) determine the similarities and differences between the human brain and the brains of other mammals, at multiple scales, for the purposes of interpreting modeling results and therapeutic testing; 2) develop ex vivo systems that capture a portion of living human brain tissue; and 3) examine molecular, developmental, and functional properties of human brain cells derived from pluripotent stem cells. Even though neuroscience had to initially play “catch-up” to other fields for the past decade with respect to embracing and applying functional genomics approaches, we believe neuroscience is now at the forefront of all fields with respect to developing and integrating functional genomics into most types of research questions. There is now a strong foundation in basic principles of mammalian brain development and function from a genomics and systems biology perspective that has been discovered in model systems. Moreover, the mainstay of neuroscience, electrophysiology, has been paired with measures of genetic output (optogenetics)²³ or genome-wide expression (Patch-seq)²⁴. Brain disease-focused researchers (e.g., in autism and neurodegenerative disorders) have also leaned into applying these genomic approaches in their arsenal of tools. Thus, we believe the next era of functional genomics should be focused directly on the human brain. Ongoing advances in human genetics can be coupled with our growing appreciation of human brain genomics and function. In this perspective, we outline the progress that has been made and where we think there are still exciting opportunities for expansion.

Surveying the transcriptome

In the last decade, high throughput genomics technologies have transformed the data resources that inform neuroscience research, especially with regard to human neuroscience. These technologies have become a mainstay in neuroscience research and across the community, spurred in part by the immense multi-billion dollar investments of the NIH BRAIN Initiative. These efforts across three consortium and dozens of ancillary programs have focused heavily on tool development for the characterization of single-cells within the brain²⁵. Such efforts brought together individuals from across disciplines and enabled the marriage of multiple modalities in the characterization of the brain. Both individual and multiomic measurements have expanded our understanding of how cell type-specific properties are correlated to one another^14,25–30 through studies mostly focused on mouse brain atlasing. Through these efforts, we now understand that snapshot measurements using conventional single-cell transcriptional approaches can capture much of the cell type and functional diversity that exists within the brain. Across human samples from both the normal brain and in the contexts of brain diseases and disorders, these datasets are transformation in their ability to provide a reference (Box 1) for hypothesis generation, validation of model system experiments, and development of new analytical methods. However, important details of brain function can only be captured by integrating information between modalities. The takeaways from these human and model system efforts highlight 1) electrophysiological and morphological features cannot be fully predicted by transcriptomics²⁶, 2) spatial context allows for more heterogeneity than transcriptional analysis alone³¹, and 3) epigenetic profiling provides essential nuance, especially when considering state transitions or development¹⁴.

BOX 1. Relevant Human Single-Cell Dataset Repositories.

Key examples of published human brain single cell datasets together with the location and type of data.

Dataset	Repository	Type(s) of Data	Sample Publications
Human Brain Development	NeMO, GEO, Individual Archives	sc/snRNA-seq, scATAC-seq	Nowakowski et al 201769, Fan et al 2018111, Zhong et al 2018112, Polioudakis et al. 2019113, Fan et al 2020114, Ziffra et al 202114, Eze et al 2021115, Bhaduri et al 20216, Smith et al 2021116, Ramos et al 2022117, Braun et al 2022118, Herring et al 2022119, Cameron et al 2023120
Adult Human Brain	Allen Brain Atlas data portal, GEO, NeMO	snRNA-seq, snATAC-seq	Hodge et al 201972, Krienen et al 2020121, Bakken et al 2021122, Ma et al 202216, Caglayan et al 202317
Neurodevelopmental and Neuropsychiatric Disorders	PsychENCODE Knowledge Portal, GEO	sc/snRNA-seq	Wang et al 201836, Velmeshev et al 201912, Gandal et al 202210
Alzheimer’s Disease	Alzheimer’s Disease Knowledge Portal (subset of Synapse)	Bulk profiles and snRNA-seq	Hodes et al 2016123, Mathys et al 201911
Other Neurodegenerative Disorders	dbGAP, GEO	snRNA-seq	Tryka et al 201438, Schirmer et al 2019124, Bressan et al 2023125
Interactive Browsers	UCSC Cell Browser, CellXGene	sc/snRNA-seq, scATAC-seq, others	Speir et al 202135, Tabula Sapiens Consortium et al 202237

In human samples, beginning with bulk transcriptomics, characterizations of steady state expression have been foundational in understanding the brain and the spinal cord, highlighting important gene programs activated across brain regions and during development. While the nature of how atlases are generated has evolved, the initial RNA-based ISH and microarray atlases of the adult and developing human brain from the Allen Brain Institute^32,33 are still widely and effectively used as a reference for benchmarking studies across human analyses and systems. With the advent of single-cell approaches, this concept of an atlas has expanded immensely. Efforts from the BRAIN Initiative, the Human Cell Atlas, the CZI Tabula Sapiens, and other initiatives by individual labs have begun to coalesce on a transcriptomic definition of how many cell types likely exist in the mouse and human brains, and the newest efforts are achieving these analyses in the context of spatial cell type distributions as well³⁴. These data are collected in repositories such as dbGAP, the NeMO archives, CellxGene browsers, Synapse or the UCSC Cell Browser^35–38 and are detailed further in Box 1.

Ultimately, these studies of steady state expression rely entirely on transcriptomic measurements and have not yet fully expanded to the other measurement modalities being used in rodent systems. Some of these challenges derive from more complicated methodologies being used in human samples, such as the notable phenomenon of lipofuscin in adult human brains when trying to implement spatial approaches³⁴. Others, such as electrophysiological measurements, require access to specialized samples and long-term culture conditions^39,40 to enable relevant characterizations.

As a result, existing surveys of human cells have been highly transcriptome-centric, with some measurements of open chromatin or methylation state^41,42. While these measurements in the single-cell space have generally risen in quality, making broad cell typing possible, there remain limitations regarding the utility of these datasets. For example, the advent of single-nuclei sequencing⁴³ has enabled study of hard to dissociate tissues, especially adult brain samples and banked human specimens. Yet, this approach (as well as single cell profiling) can still suffer from excessive ambient RNA content, a phenomenon that releases cytoplasmic RNA into the cell suspension prior to analysis, impeding the accuracy of the single-cell nature of the experiment⁴⁴. These issues have been particularly challenging when studying glial ratios in samples, because these populations have an increased probability of mis-typing due to their naturally lower numbers of expressed genes. Additionally, even when the modalities represent samples well, gaps in terms of brain regions, developmental stages, or certain disease states exist across model organisms. These gaps are especially pronounced in the context of human dataset acquisition where sample availability and scope of existing consortia have not yet yielded comprehensive cell type atlases.

The ongoing consortia efforts pursuing atlases of the human brain present an interesting paradox to researchers working in spaces that would benefit from these atlases of the normal or diseased brain. The resources and distribution methods of consortia can be hypothesis generating and incredibly useful to individual labs, but the missing data points across regions, time points, and disease states can diminish the utility of current human datasets. However, the cost and the chance to be usurped by larger groups decreases the incentive and the opportunity for smaller groups to pursue the filling of these gaps. In the field, transparency about the planned scope of large institutes and consortium efforts paired with advances in lower cost technologies for sampling single-cell profiles would help fill these gaps more efficiently. Importantly, researchers studying human biology and disorders are the most well positioned to vet and optimize data generated from these technologically challenging approaches and are well positioned to plug the holes in terms of “missing” datasets and modalities related to human neuroscience.

Characterizing the Synapse with Omics

The nervous system is uniquely characterized by complex information flow through synapses, junctions between neurons and sometimes other cell types. The human brain is thought to contain on the order of 600 trillion to a quadrillion synapses⁴⁵, and synaptic dysfunction or even minor dysregulation is implicated in numerous developmental, psychiatric, and neurodegenerative disorders. Importantly, compared to other species, humans have more synapses per neuron⁴⁶. Thus, tools to study synaptic biology are necessary to truly understand brain biology. Moreover, in light of the thousands of cell types being identified from the atlas efforts described above, the characterization of cell type-specific synaptic interactions with other populations is a tantalizing link between phenotypes, behavior, and steady state measurements. Transient synaptic connections are a hallmark of key developmental processes as well, notably between the thalamocortical afferents and transient cell types and substructure of the developing cortex for example^47,48, establishing the study of synapses via genomic biology a unique and essential feature of next generation neuroscience.

Notable progress has been made in this sphere. Bulk proteomic characterizations have been generated from various species and across disease contexts, especially in the context of normal brains and neurodegenerative diseases. However, recent work is also emphasizing the interesting heterogeneity to be found between cell type-specific proteomic studies of the synaptic proteome. Given the challenges of measuring electrophysiological properties accurately in human tissues, studies of synapse biology at the bulk or single cell level provide a unique glimpse into human-specialized processes from preserved tissues. For example, recent efforts exploring proteomics of synapses during development highlighted potential insight into mechanisms of neoteny in humans⁴⁹. In the context of aging and neurodegeneration, profiles have identified disease associated synaptic proteins in ALS⁵⁰ and Alzheimer’s disease⁵¹, as well as synaptic proteins subject to age-related changes⁵² or even those that seem to be resistant to these types of degradation^52,53.

Synapses present a unique challenge, because not only is there a large number of synapses in the human brain, but the specific properties of these synapses are likely quite diverse. Given the size of the human brain, synapses are often located far from the cell body and associating a synapse with the originating cell is incredibly challenging. New tools to characterize single-cell gene expression of synapses with droplet-based capture approaches⁵⁴ are beginning to change our ability to understand the composition and heterogeneity of synapses as well as how they are dysregulated in disease states such as Alzheimer’s. However, these approaches do not solve the question of what cell type the interacting cells are and are also somewhat controversial regarding whether they accurately measure synapses^55,56. Thus, additional innovations in this space are required to fully understand the heterogeneity of synapses within the human brain. Other strategies to interrogate this problem are emerging, for example, recent innovations in capturing single dendrites highlighted key properties of local translation in rat neurons⁵⁷. Importantly, these results are highlighting the opportunities to gain information from sub-cellular analyses that move beyond nuclear transcripts⁵⁴, emphasizing the need to engage in more comprehensive synaptic single-cell analysis, especially because this is accessible in the human context.

Beyond the contents of what exists in a synapse specifically, understanding where this information is being shared has important implications for how individual cell types communicate with cells from different regions. This will enable a better understanding of how single cells can contribute to circuits that ultimately underlie behavior and may be dysregulated in cases of disease. Early pioneering efforts in rodents to characterize this important link between cells and function harnessed the unique properties of rabies viruses that infect the nervous system and propagates via retrograde axonal transport⁵⁸. Engineered modifications of this system eliminated the extreme impacts of the virus and instead allowed for targeted tracing of an individual, genetically labeled cell, usually from the rodent brain⁵⁹. As single-cell technologies have emerged, these types of approaches can now be paired with single-cell analysis enabling linkage between an individual cell and the cells it communicates with. Use of these approaches in neurons and between glial populations have expanded our understanding of how presynaptic networks are structured and what proportion of interaction partners belong to specific classes⁶⁰, including identifying a role for axon guidance molecules in glial interactions⁶¹. These approaches are now being extended to the spatial context, integrating both cell type and connection location with in situ sequencing on brains previously infected with custom synaptic tracers⁶². Parallel tools with some cell specificity are adeno-associated viral strategies that label cell types of interest⁶³ and can be used in ex vivo slices to map projections and morphologies⁶⁴. However, an important caveat of these tracing tools in the context of human biology is that they can only be applied to slices of live tissue, limiting their utility to intra-cortical relationships, for example. These tools, while still evolving rapidly, extend the ability to fully understand synaptic function but thus far require manipulatable in vivo systems for full scale characterization, such as model organisms ranging from rodents to non-human primates. However, human neurons are characterized by a number of unique properties including decreased intrinsic excitability, increased apical dendrite length, and more circuit compartmentalization⁶⁵. Given the challenges of directly assessing these properties in human explant cultures, progress in the ability to utilize xenotransplantation of human pluripotent stem cell derived developing neurons into mouse brains is an exciting approach to explore the development of human neuronal morphology and circuits^66,67. Using these tools, it has been observed that human neurons mature within the rodent brain but intrinsically maintain their neoteny (prolonged developmental timeline)⁶⁵; these types of experimental models are promising systems in which to apply the expanding toolkit of single-cell approaches to study synaptic biology. To derive value from these tools directly in the context of human biology, it is imperative to adapt these tools to live postmortem culture models of human tissue, explore their utility in in vitro human model systems, and connect the tracing findings from rodent models to other measures of gene expression and synaptic proteomes. These adaptations will determine whether any of these characteristics can be inferred from measures that are acceptable in the human context.

Moving from lists to networks

As detailed above, technological improvements together with consortia and initiatives have resulted in vast amounts of genomic data, much of which are relevant to the human brain. The efforts have yielded catalogs of genes that are associated with cell types, developmental trajectories, and/or disease status. However, one of the greatest challenges posed by these datasets is determining how to consolidate and prioritize the information. In other words, which genes are most crucial for a particular developmental process, phenotype, or disease? Since a single gene cannot determine a specific cell type, we need to consider genes in groups or networks. In the era of bulk tissue RNA-sequencing, this challenge was facilitated by the use of approaches to mine the relationships among patterns of gene expression in a way that was agnostic to other known genomic or biological information. Weighted gene co-expression network analysis (WGCNA) is a popular approach to systematize brain genomic datasets¹. This method groups genes into modules based on their patterns of co-expression. Within each module, the genes that appear to be most linked to the pattern of co-expression can be prioritized as hub genes. It is worth noting that when starting with a bulk genomics dataset, WGCNA is useful for pulling out co-expression signatures related to cell types⁶⁸. For single cell datasets, WGCNA is useful to apply to already defined major cell types or to datasets with dynamic cell type-specific expression, as in the case of a developmental dataset (Figure 1). For example, in the developing human brain where dynamic patterns of gene expression make defining subtypes of cells challenging, modules can be generated by cell type then re-clustered to obtain cleaner subclusters⁶⁹. In other studies of human primary tissue and/or organoids, WGCNA has been applied to observe how defined networks change across developmental time or between species^70,71. The use of WGCNA and related co-expression methods will be key to apply to the realm of underexplored human brain regions and disease states. The re-invigoration of gene networks into the human neurogenomics field could provide additional color into how biological processes are not necessarily always completely parallel to cell type. We would like to note, however, that while WGCNA is still relevant in the era of single cell modalities, application of WGCNA to single cell datasets may be limited when cell type heterogeneity is already apparent and defined. Moreover, because single cell datasets tend to be more sparse with many zeros (whether due to technical or biological reasons), the implementation of WGCNA may be challenging due to sparsity or the number of data points. Sparsity can be addressed by deeper profiling per cell type (e.g. with SMART-seq or Fluidigm)^15,72, imputation of missing data points or pseudobulking values for genes across cell types^73,74.

Figure 1: Mining human brain single cell datasets to infer regulatory mechanisms.

Human brain tissue can be queried for genomic information at the RNA and chromatin level. However, these datasets require further applied analysis to understand the dynamic nature of the datasets. WGCNA can be used to understand cell type composition and contributions as well as the dynamic nature of cell type development. Gene regulatory networks (GRN) can be applied to infer which transcription factors (TFs) might be important for co-regulation of sets of genes in a given cell type. Since these regulatory mechanisms cannot be directly tested in vivo in human, they can instead be tested in either human brain slice culture or model systems (e.g., rodent or monkey) using CRISPR-based tools.

Other methods such as gene regulatory networks (GRN)⁷⁵ are therefore important and needed to mine single cell datasets to determine functional relevance. At its simplest level, implementation of GRN essentially identifies the genes regulated by transcription factors and builds this out at scale into a network. In a scRNA-seq dataset, this translates to determining transcription factor motifs at cell-type resolution (Figure 1). Thus, one can graph putative co-regulated (via the same transcription factor) genes in a cell type-specific manner. GRN can not only be applied to scRNA-seq but also to single cell chromatin data such as snATAC-seq. The layering and integration of both RNA and chromatin datasets is key for building functional networks, especially in human brain datasets, since perturbations to test functional outcomes are limited. The combination of RNA and chromatin at the cell type level facilitates identifying genes and regulatory mechanisms (via motif and variant mining in areas of the genome that are differential across brain states (development, disease, etc.)) and correspond to changes in gene expression. While applying some of the latest GRN methods to multimodal datasets has been shown to improve the predictive power of the GRN approach overall (e.g., MIRA, SCENIC+, and Dictys), there are still limitations even to these recent, benchmarked approaches in that some types of data such as validated TF binding sites or enhancer-promoter links are not sufficiently observed⁷⁶. One of the important challenges for all these approaches with respect to human datasets, is the relatively constrained ability to confirm any of the outcomes. Both WGCNA and GRN are predictive rather than definitive of causation. In other words, these are hypothesis-generating approaches that need to be independently validated. For GRN-generated TF networks, it might be possible to validate the target genes of a specific TF using cell type specific transposase and tagmentation approaches such as single cell CUT&Tag⁷⁷. We envision such approaches could adapt the use of antibodies to TFs (compared to validated antibodies for histone modifiers) and apply the approach to human brain samples.

Beyond frequently used approaches such as WGCNA and GRN, additional methods are emerging to mine multimodal genomic datasets. A causal inference approach, CoCoA-diff, has been applied to single cell RNA-seq data from Alzheimer’s disease brain tissue, facilitating the prioritization of cell type specific genes that are likely to be drivers of disease state⁷⁸. One recent machine learning approach that included human brain tissue is the Geneformer tool⁷⁹. Single cell RNA-seq data from human brain tissue was included within the large training dataset Geneformer used, and the networks predicted the contribution of copy number variants (i.e., dosage) of developmental disease genes with higher accuracy in neurons compared to random cells from throughout the body. There is much excitement about applying artificial intelligence approaches such as deep learning to single cell transcriptomics⁸⁰. However, there are still challenges (in general⁸¹) and unknowns in how such tools might predict functional networks in a system like the human brain that is not amenable to perturbation except in select ways.

Hypothesis testing of integrated datasets can be carried in human datasets via in silico approaches such as CellOracle⁸² or genomic perturbation methods such as PerturbSeq⁸³ or massively parallel reporter assays (MPRAs)^84,85. These manipulations in a human context, however, require accessible models of the human brain and either immense scale in terms of cell number or limited gene sets because of the number of cells per guide RNA required for robust conclusions. These types of manipulations are the ideal scenario in which stem cell-derived models of human neurons or organoid systems of normal development or neurodevelopmental disorders can enable relevant perturbations. However, because of the in vitro nature of these models, benchmarking existing datasets from primary samples and including ample cell line and technical replicates is essential to ensuring that the phenotypes observed are not artifacts of the system.

Finally, many studies, especially those funded through BRAIN-initiative grants, are tasked with setting goals across the modalities of RNA and chromatin as already mentioned, but now also include spatial approaches due to available technology. Since single cell approaches necessarily dissociate tissue to profile individual cells, the implementation of spatial transcriptomics can be used to either independently generate or validate cellular genomic datasets across human brain subregions (e.g., a human cortical section)⁸⁶. A limitation of human spatial transcriptomics is the sheer size of the human brain relative to the imaging window. However, one could imagine tiling together datasets across larger regions of the human brain to make key spatially-relevant insights. Another consideration is the sheer amount of imaging time that might be needed (and computationally intensive analysis) to cover large sections of the human brain. However, as these hurdles are surmounted, spatial transcriptomics could be scaled to slice cultures of larger areas of human brain and integrated with additional approaches that interrogate functional measures such as connectivity and viral gene manipulations, further bridging the genotype to phenotype gap.

Integrating genetic and phenotypic data

An ultimate goal of studying the human brain is to piece together the mechanisms that underlie our behavior, particularly with regard to those traits that distinguish us from other species such as language. One way to achieve this is by profiling genetic and genomic information across a wide range of individuals with varying phenotypes. This can be achieved by comparing datasets from neurotypical individuals with those who have brain disorders with distinct behavioral symptoms such as neurodevelopmental (e.g., autism spectrum disorder (ASD) or schizophrenia) or neurodegenerative (e.g., Alzheimer disease (AD)) disorders). Since it is not feasible to perturb humans genetically like we do in model systems, we can instead compare data from neurotypical individuals to those with disease to infer correlations between genes and behavior in the human brain (Figure 2A). These studies necessarily either use peripheral tissues such as blood or saliva (for DNA variants) in living individuals or postmortem tissues for brain access. While DNA polymorphism and copy number variant information have revolutionized our understanding of the contribution of both rare and common variants to risk for these brain diseases^87–89, it is still not entirely clear how DNA alterations in these risk factors impact brain cells in a functional manner. In other words, how can we infer causality from DNA all the way to behavior/disease via cells and circuits? This is a significant gap in knowledge that the neurogenomics field should address over the next decade. In addition, we are challenged to identify how developmental pressures and/or brain activity intersect with genetic risk in the absence of datasets derived directly from brain tissue. Thus, researchers need to be innovative in how postmortem brain tissue can be harnessed to understand functional outcomes. One logistical challenge in leveraging postmortem tissues is that humans have diverse lifestyles and demographics, and thus available datasets are not always well matched (e.g., by age, other demographics, drug or treatment conditions, or brain region). Large scale datasets can mitigate some of these issues. For example, the UK Biobank dataset currently contains genetic data from hundreds of thousands of individuals with corresponding demographic information as well as some phenotypic information and brain imaging from tens of thousands of individuals⁹⁰. These datasets have been used to make a number of important insights into how genetic variants correlate with both structural and functional brain measures and mining these datasets has identified specific genes that may underlie these phenotypes as well as link to particular brain disorders⁹¹. One disappointing outcome from these and related studies has been the inability to identify significant genetic signals that correspond to task-based fMRI measures⁹², in other words, the genetics underlying human behavior. It is not clear if further increases in sample size or refinement of imaging methods or behavioral tasks will alter this outcome.

Figure 2: Integrative approaches to use human brain tissue to uncover multimodal systems related to genomics.

A) Brain imaging can be combined with peripheral DNA profiling to infer genetic variants that may underlie brain size or function. B) Brain gene expression and chromatin state data from post-mortem tissues can be integrated with functional MRI measures to understand how gene expression patterns may underlie brain activity. C) Human post-mortem tissues can be used to confirm findings from all other approaches. D) Ex vivo brain slice cultures can be acquired from surgical patients and use to integrate physiological and gene expression measures from the same individual.

Another recent innovation is the integration of human brain imaging with postmortem brain gene expression datasets (Figure 2B). In these comparisons, the datasets are derived from two different cohorts of people but the relatively stability of these measures across neurotypical populations has led to key insights about genomic underpinnings of brain morphology, size, lamination, and cell types⁹³ as well as functional resting state measures^94,95. These functional imaging datasets typically rely upon large repositories of imaged cohorts such as from the IMAGEN or Human Connectome Project for neurotypical individuals or disease relevant consortia such as ABIDE for ASD. Many of these integrative studies also use RNA-seq data from the human brain atlas generated by the Allen Brain Institute⁹⁴, but an increasing number of studies are beginning to add additional human postmortem RNA-seq datasets^95,96. Moreover, the inclusion of imaging and/or genomic datasets from individuals with a number of disorders such as depression, schizophrenia, ASD, and other neurodevelopmental disorders has resulted in the identification of genes that have patterns of expression that vary in a disease context^96–99. Interestingly, these are not necessarily the same genes that are associated with genetic variant risk, suggesting a complex interplay of genetic risk and functional genomic outcomes. Beyond the use of data from disease cohorts, there are efforts to correlate genomics and brain imaging with specific behavioral repertoires relevant to humans such as language¹⁰⁰. All these studies have laid the groundwork for human brain gene prioritization as well as insights into how groups of genes correlate together with respect to a number of brain functions or disease states. A caveat to these studies is that the postmortem tissue data are derived from separate cohorts of individuals than those who underwent imaging. Therefore, the demographics of each population can be challenging to match especially if disease or developmental datasets are considered.

Another important source of gene expression datasets (e.g., RNA-seq and/or ATAC-seq) derived from human brain tissue is surgically resected brain tissue. The use of these tissues precludes issues surrounding independent group correlations as this tissue can be obtained from individuals who undergo prior brain imaging (structural or functional), intracranial stimulation or electrophysiological measures such as electroencephalograms (EEG) (Figure 2D). For example, a within-subjects study identified the gene expression patterns associated with specific oscillatory activity during a memory recall task, uncovering specific genes that are likely important for normal human memory function¹⁰¹. Of course, a major caveat to the use of surgically resected tissue is that all the individuals providing tissue undergo operations for a medical condition such as medication-resistant epilepsy or cancer. The use of rapid autopsy tissue could fill a gap with respect to the availability of tissue from individuals with non-brain disease and access to the entire brain rather than where surgery dictates. Regardless, neither surgical nor rapid autopsy sources are likely to provide an abundant resource for brain tissue for genomics from individuals with cognitive disorders like ASD or AD. Therefore, between group correlations need to be made for integrating genomics with brain function in these cases.

At the cell type level, biophysical phenotypes can be integrated with genomic measures through approaches such as Patch-seq in which individual cells undergo patch clamp measurements followed by scRNA-seq²⁴. Patch-seq has been used at scale to compare human and mouse cortical cells, identifying features that have been modified in human brain cells and might be at risk in certain disorders². These efforts could be further expanded by manipulating disease-relevant genes or cell types in organotypic slice culture derived from surgical resections and/or rapid autopsy brain tissues. Human brain slice cultures can be kept viable for several weeks and are amenable to viral transduction^40,102. Again, applying these techniques to brain tissue from individuals with specific brain disorders like ASD or AD would be limited. But high-quality gene expression and regulation measures from postmortem tissues can be obtained, and information applied to physiological studies such as key disease-relevant gene manipulation in the organotypic slice culture.

Importantly, for all these integrated studies, detailed phenotypic information related to the donor of the tissues is key in order to correlate genomic features with either demographics, disease-relevant phenotypes, or technical considerations. Coordinated brain banks that have high standards of best practices are making these tissues readily available to qualified researchers (e.g., NIH NeuroBioBank for control as well as several types of brain disorders and Simons Foundation BrainNet for ASD and control tissue). A challenge for these brain banks is to collect enough samples across the developmental and age-related epochs that are impacted by disease state. Even though the post-mortem tissue samples will reflect steady-state levels of expression, the use of Hi-C and ATAC approaches have already given insight into some aspects of molecular/chromatin features that are dynamic^{13,36,103–105}.

An option to fill in any missing knowledge of how DNA variants lead to alterations in behavior, via changes in cell type expression patterns or brain circuits in humans, is to harness the power of non-human primate (NHP) models. NHP models that are ethically amenable to gene manipulation (e.g., marmoset or macaque) are still evolutionarily distant from humans and harbor many genomic differences at the cell type level^16,17. However, their brain gene expression patterns and circuits also share many features with the human brain²⁵. Thus, each model system such be selected based on the specific hypothesis that relates to human behavior. For example, a songbird model could be ideal for a study of vocal learning over even great apes, which do not have observable vocal learning.

Human data sharing

As the genomic tools to study neuroscience improve and expand in scale to address technical challenges associated with human samples, the number of studies utilizing donated surgical, biopsied, or postmortem tissues is expanding rapidly. Numerous organizations, including funding agencies such as the NIH are simultaneously pushing for increasingly open access to this data, including the raw sequencing reads. Scientifically this is reasonable because tools for alignment, parsing isoforms, and other applications to improve differential gene expression are improving and expanding while traditional repositories such as dbGAP or SYNAPSE are onerous to navigate bureaucratically. More streamlined access would allow speed and efficiency for tools that are yet to be conceptualized, much less developed.

However, as the tools to utilize these raw sequences improve, so also do the technologies to infer identifiable characteristics within these data. For example, while short read, low coverage 3’ enriched sequencing once was impossible to use for the identification of SNPs, tools and models now make relatively high confidence SNP calling in these sequencing datasets feasible and have been used for scientifically exciting applications such as demultiplexing identity in sequencing runs¹⁰⁶, tracing potential lineage relationships between cells^107,108, or identifying tumor subclones in single-cell studies¹⁰⁹.

These same tools, when paired with the expanding repository of human sequencing in scientific and social contexts, are quickly approaching adaptability for identifying not only individuals but also their kin. This poses ethical risks with regards to most consent forms where deidentification is a condition of the research use of a sample and increasing societal risks to individuals and their relatives. Spelling these risk factors out directly (Figure 3) is not alarmist, but rather is specific in terms of the harms that can arise when sequencing data is not fully protected; more comprehensive access must come with legal protections against these harms. The considerations on family are particularly salient when dealing with postmortem tissues that are archived with older consent forms; not only might the research uses extend beyond those originally planned, but new technologies may expand the impact of identifying previously unknown genetic risk factors to surviving relatives who may not have been consulted during the donation process.

Thus, we as scientists and the broader community must consider what types of firewalls to protect data and community guidelines of scientific conduct must be maintained when housing, sharing, using, or publishing these sensitive data types. Additionally, when thinking about what aspects of the data or metadata will be shared, protections for future potential uses of this data must be considered in policy and scientific decision making now.

Conclusions and future directions

Data integration across modalities remains a challenging frontier in human brain genomics. However, as access to human brain tissues with associated phenotypic information increases, the gap between genes and behavior should narrow. Advances in chemistry and brain imaging may facilitate mechanisms to query gene expression in vivo. Certainly, the possibility to image the activity of a living human brain at single cell resolution might be achievable¹¹⁰ and resultant data could be integrated with independent genomic measures.

Most importantly, the neuroscience community has come a long way from reluctant use of genomic methods and hypothesis-generating approaches to embracing technology that profiles the brain agnostically at cellular resolution or integrates electrophysiological measures with gene expression. While significant progress has been achieved in these domains within model systems, in the next decade, neuroscience research is certain to push the boundaries of what is possible in observing and manipulating human brains. These new approaches will enable additional characterization of human biology, and existing integrative analysis approaches overviewed here can continue to drive cohesive biological understanding of existing and emerging datasets. The functional understanding of how genes influence the human brain across scales from cells to circuits will usher in a new era of neuromodulation and gene manipulations. Thus, we propose that a continuation of efforts to delve into genomic features of the human brain and how these features vary across cells, circuits, and behaviors will be rewarded with a rich appreciation of how the human brain evolved, develops, functions, and dysfunctions. Determining how these genetic features directly affect behavior via cells and circuits is a challenging feat in the human brain that may require new technological advances. These achievements will not only enhance how we understand human behavior but also have therapeutic value.