Large-scale neurophysiology and single-cell profiling in human neuroscience

Are we ready for Human Neuroscience?

Modern neuroscience began with Paul Broca’s discovery of language cortex based on observations that lesions in the human brain disrupt normal cognitive functions. Subsequently, Wilder Penfield went on to map the functional areas of the human cerebral cortex through intraoperative stimulation to localize language, memory, and movements. Subsequent mechanistic studies to understand brain function beyond localization have primarily been advanced over the last few decades in animal models. While revealing fundamental principles about the formation and function of the neural circuit, this has left a gap in our understanding of the regulators in higher functional domains such as language and higher sensory processing. Uncovering their genetic, cellular, and molecular basis of cognitive specializations observed in humans (Box 1) has prompted renewed interest in studying the human brain. However, studies of the human brain are challenging due to the lack of scalable tools to manipulate cell behavior or gene expression, and confounding variables of individuality.

What makes us human?

Important insights into brain function have been gained from animal models. Mice offer the opportunity to genetically access and manipulate neural circuits. Comparative studies can therefore provide insights into the cellular, molecular and genetic underpinnings of brain function. However, humans are distinct in many fundamental ways.

Genetics:

While humans and mice share most of the protein coding genes, there are a number of genes that are not conserved to mice or even non-human primates. Such mutations have been defined for only a subset of the genome, but now that the complete genome can be resolved using telomere-to-telomere sequencing¹³⁴, a full list of human-specific mutations will be defined. In addition, comprehensive understanding of human genetic diversity is being uncovered through large scale collaboration projects such as the Pangenome project¹³⁵. Findings from these studies will likely have important ramifications for our understanding of human neurodiversity, evolution, and susceptibility to disorders.

Cell types and circuits:

Although comprehensive and cross-species molecular comparisons are yet to be fully conducted, initial studies suggest that broad cell types can be largely found across species. What differs across species are cell type specific molecular profiles as well as spatial distribution, including across the layers of the cerebral cortex¹³⁶, and deeper analyses have begun to unravel differences in cell subtype diversity^137,138. Classical comparative studies have underscored differences in abundance of pyramidal neurons in upper cortical layers of the cerebral cortex^139,140. Finally, differences in long range connectivity have long been appreciated across species¹⁴¹.

Neurodevelopment:

Human development is protracted. Key differences include expansion of neural stem cell populations¹⁴², expanded number of cortical areas, including disproportionate expansion of prefrontal areas^44,139, neoteny of synaptic connections^143,144, and adaptations of cortical glutamatergic neurons^51,137.

Functional:

Human brain has evolved to enable complex cognitive functions, including language, abstraction and creativity. Several hypotheses have been put forth as to how these specialized functions emerge, including co-option of ancestral circuits, or emergence of new circuits, as discussed in many excellent reviews^145–147. It is evident that human brains are highly specialized to these complex behaviors, and understanding how these processes occurred during evolution is critical to understanding of what makes us human.

In this perspective, we highlight emerging efforts to advance human neuroscience through our personal experiences as a collaborative team focused on cellular-level function and structure, including neurosurgeons (EC, AL), a neurologist (MP), and basic neuroscientists (TN, AL). We believe that the convergence of neuroscience research techniques that enable mapping of the human brain at large-scale and with cellular-resolution has finally arrived for the fields of human neurophysiology and anatomy to herald the burgeoning field of human neuroscience. Specifically, we describe how parallel breakthroughs in the surgical operating room, including the ability to perform large-scale recordings of neural activity in single neurons of awake behaving participants, coupled with long-term culture methods of resected brain tissue, opens up a new unprecedented opportunity to study the entire spectrum of human neuroscience, from network to microcircuit, to cell types and their molecular features. Given our collective expertise, we do not focus specifically on advancements in cognitive and systems neuroscience, but rather how we can leverage human tissue to study cellular function and genetic diversity. Importantly, as we make strides in understanding the genetic, cellular, and molecular bases of human cognitive specializations, we face important questions about societal implications and ethics related to human neuroscience research. We describe paradigms that may allow the scientific community to proactively engage in broader societal domains including the concepts of informed consent and data sharing, especially in a field where patients are experimental subjects, direct benefit from the research is currently limited, and importantly, where ramifications of future advancements in machine learning and artificial intelligence analyses are thus far unknown.

Large-scale neurophysiology of human behavior

Non-invasive imaging technologies such as functional magnetic resonance imaging (fMRI) have provided important insights into neural dynamics across the entire human brain but are substantially limited in resolution. On the other hand, technologies that can directly record neural activity from the human brain have improved dramatically over the past decade, increasing in both scale and spatiotemporal resolution.

Historically, traditional tungsten microelectrodes used in intraoperative recordings measure activity from a few neurons per electrode. Despite their limitation of scale, these devices have high spatiotemporal resolution which enabled direct measurements of activity from single neurons such as in the basal ganglia for placement of deep brain stimulation (DBS) electrodes in Parkinson’s disease. Recordings from stereotactically-guided depth electrodes in stereoelectroencephalography (SEEG) and placement of craniotomy-based surface grid electrocorticography electrodes (ECOG) in epilepsy surgery are able to capture large neuronal networks, but lack the spatiotemporal resolution necessary to detect single action potential spikes. Instead, these latter modalities capture an aggregate activity of inputs to local neuronal networks via local field potentials (LFPs). Thus what has been lacking in the field of human neurophysiology is a means of capturing single units at large-scale.

Why is single-unit activity critical to more fully understanding network function? For starters, inferences of neuronal spiking activity from LFPs alone are indirect¹ and do not address fundamental questions of functional organization, circuit connectivity, or information processing². Indeed, how oscillations (as detected by LFPs) contribute to cell-type specific neuronal firing patterns and vice versa are an area of active investigation in model systems^3–5. In rodents, specific inhibitory neuron subtypes respond preferentially to various aspects of sensory discrimination tasks, while different oscillation frequencies are associated with a range of disparate cognitive functions (i.e. theta in working memory, beta in rule application)^6–11. For example, inhibition of prefrontal parvalbumin interneurons in mice reduces gamma oscillations and negatively impacts rule-shifting tasks⁴. How these findings from model systems translate to human neuroscience is an open question, notwithstanding the cognitive functions that are unable to be readily studied in model systems such as language. From a neuroengineering perspective, multiscale decoding algorithms that incorporate both LFP and individual spikes outperform single-scale decoding using either modality alone in BMI simulations¹².

Previous approaches to obtain higher yield of single-unit recordings include electrode designs, such as microwire bundles that exit the depth electrode tips (Behnke-Fried electrode), or the microelectrode Utah array, which permit action potential recordings from dozens of single units. Indeed, some groups have been able to leverage these techniques to collect thousands of single neuron activity, though requiring large patient cohorts and hundreds of separate individual probes¹³. These methods however, are limited and have not advanced significantly since their introduction several decades ago. For electrode technologies to be applied in human neuroscience, they must address several critical issues, including 1) spatial resolution (i.e. action potential of single cells), 2) temporal resolution on order of microseconds, 3) brain coverage (local ensemble/cortical column, or across disparate areas), 4) robustness (ability to record for minutes intraoperatively, vs chronically for months and years), and 5) safety (biocompatibility, injury from insertion).

A recent fundamental advance is the Neuropixels probe, which features a thousand electrodes on a narrow shank, with the potential to record from hundreds of neurons from a single probe insertion. Neuropixels have transformed electrophysiology research in rodents and nonhuman primates¹⁴. We have recently demonstrated its feasibility of simultaneous large-scale data collection from single neurons in human intraoperative recordings after overcoming challenges such as probe fractures, insertion difficulties, electrical noise, and brain movement-related artifacts^15–17. Vertical probe orientation permits hundreds of cellular recordings within a cortical column, across all cortical layers. Recordings appear robust within seconds of insertion, and thereby overcome challenges of previous single unit recordings with low yield. Importantly, the data acquisition system is turn-key and constantly improved by a community of active researchers. Recent Neuropixels advances have overcome existing limitations in non-human primate neurophysiology such as large-scale sulcal recordings or simultaneous wide-spread recordings from multiple interconnected brain regions such as primary motor, supplemental motor area, and basal ganglia for studying motor planning and control¹⁸.

Using Neuropixels, we have been interested to address questions about stimulus feature encoding in speech that have been difficult to address using LFP recordings alone. Because Neuropixels samples spiking activity through all cortical layers, we now have the capability of studying how neuronal activity is processed as it flows through cortical layers within individual cortical column. While model systems have begun to address questions of cortical columnar processing, many of these techniques require either calcium dyes, viral transduction of calcium or voltage indicators, 2-photon imaging, or implantation of large lenses, none of which have the temporal resolution nor safety feasibility of Neuropixels for human research¹⁹. By integrating population-level LFP data with the spiking activity of individual neurons and information processing through cortical layers, Neuropixel recordings afford us opportunities to study cellular mechanisms of how auditory circuits encode for the multitude of stimulus features in human language within cortical columns and across cortical layers^17,20, Future work will prioritize scaling Neuropixels probes so that recordings can be carried out in parallel from multiple simultaneous brain regions, sampling tens of thousands of single neurons, and ideally fully compatible with chronic recordings. One limitation of the current Neuropixel technology is that an individual probe can only collect data from a single cortical area. To facilitate larger cortical coverage, emerging single-neuron recording techniques could be distributed over a wider range of the cerebral cortex. Another complementary approach includes high-density micro-electrocorticography^20,21, which focuses on recordings from the cortical surface using new materials and fabrication processes that are common to advanced micro-electromechanical systems (MEMS) devices, allowing electrode arrays to be fabricated using chip printing lithography on highly flexible and conformal polymer substrates. Such approaches have the advantage of covering significant cortical areas, while maintaining the safety of avoiding cortical surface penetration. Recent demonstrations include the ability to record action potentials, as well as very focal cortical events like high-frequency oscillations (HFOs), on the scale of ten of microns in resolution²². Together, surface uECOG and penetrating laminar multi-electrode arrays such as Neuropixels will allow for unprecedented three-dimensional measurements from cortical and subcortical structures to understand how distributed neuronal activity gives rise to patterned activity and human behaviors²³.

An atlas of cell types and circuits

Relating the activity of single neurons measured in the context of an in vivo circuit or network to the molecular identity of that neuron represents a fundamental and unresolved challenge in neuroscience. A deeper appreciation for the cellular building blocks of a neuronal circuit will not only provide more mechanistic understanding for how different types of information are processed in the human brain, but also potentially provide opportunities to modulate the physiology or transcriptome in diseases with cell-type specific treatments. Approaches such as a priori genetic labeling of molecular cell types followed by in vivo activity recordings have provided some of the first proof of concept applications, but face challenges in applications to non-model organisms, including humans²⁴.

Cells in the brain have been classically categorized into types based on morphology, anatomical position, as well as intrinsic physiological properties such as action potentials, resting membrane properties, and various channel properties. More recently, single cell transcriptomics and epigenomics have enabled capturing a more complete spectrum of cellular properties thought to define cell types (Figure 1). We initiated these efforts in the brain tissue by profiling dozens of cells²⁵, then went on to report hundreds and thousands of cells^26,27. Breakthroughs in technology for capturing single cells using microfluidics have opened up a practical path towards census- scale efforts in neuroscience²⁸, many of which are currently ongoing. Recently, a complete census of the mouse brain has been completed, reporting that the adult mouse brain contains approximately 5,000 cell types^29,30. An initial draft atlas of the human brain revealed over 400 clusters and 3000 subclusters of cell types, with many showing regional diversity in neurons and macroglia unique to their respective forebrain, midbrain, and hindbrain regions³¹. How this complexity compares between the mouse and human brain human brain will likely be resolved over the next few years. Advances in technology that enable microfluidics-free cell capture³² will further accelerate these efforts by enabling routine capture of tens to hundreds of millions of cells in a given study.

Figure 1: The Central Dogma of neuroscience.

Cells in the human brain share the same genomic information but develop distinct patterns of gene expression. In turn, gene expression profiles correlate, and in many cases, regulate the abundance of proteins that build the cells and determine their distinct properties, such as morphology and physiology. Ensembles of neuronal cells give rise to synaptically-coupled neuronal networks that generate network-level properties. Coordinated activity of neurons in neural circuit give rise to motor output and cognition. Importantly, neuronal activity feeds back to control gene expression through activity-dependent molecular and cellular changes.

Ongoing efforts seek to characterize the diversity of cell types in the human brain using single cell or single nucleus transcriptomics³³, chromatin state profiling, and epigenomics^34–36. New and emerging efforts are beginning to extend the spectrum of methodologies to incorporate measurements of lipids, metabolites, proteins³⁷, opening up new opportunities for deep molecular characterization of cell types in the human brain. It is important to recognize that many features of cell identity, such as the subcellular transcriptome, are not captured in those studies due to the limitations of current technologies^38,39. We also must embrace the fact that a consensus definition of a cell type is currently lacking⁴⁰, but even if the question of the absolute number of cell types in the brain cannot be fully addressed, we can be successful in defining their transcriptomic marker genes. The impact of census-scale transcriptomic studies have already led to the development of technologies that allow us to visualize transcriptomic cell types in tissue^38,41–43 and begin to precisely localize them within a common coordinate framework⁴⁴.

Linking cellular function and networks

Molecular cell states have primarily been assayed using postmortem human tissue specimens. However, the time needed to harvest and freeze postmortem tissue may result in significant changes in gene expression⁴⁵. Furthermore, studies of postmortem tissue fail to capture the functional properties of cells, many of which could be fundamental to cell type definition (Figure 1). The extent to which transcriptomic cell types contribute to network dynamics is an open question, which is not even fully addressed in model organisms^24,46. Thus, the advent of approaches that allow neurosurgical specimen to be maintained viable in culture for multiple days offers an alternative opportunity to study cell types within a functional microcircuit^24,46,47 (Figure 2).

Figure 2: Central approaches to advance the study of human brain.

Intraoperative recordings on neural activity while patients are doing cognitive tasks can be conducted ethically and with patient consent to gain insight into neural correlates of human behavior (top left). Tissue resected from the brain can be maintained in ex vivo culture condition to facilitate single-cell electrophysiology and molecular profiling (bottom left). In parallel, postmortem tissue specimens provide generalizable insights across brain regions and broader patient population (right).

Human brain tissue resected from surgery has already begun to reveal new insights into how human neurons process information by leveraging human-specific neurophysiology such as unique dendritic compartmentalization and computation^48–51. Since many of the epilepsy surgeries require the patients to be awake, ex vivo approaches also afford a rare opportunity of investigating the genetic and cellular composition of tissue that has been functionally mapped as the patient undergoes cognitive testing. For example, transcriptomic studies of resected tissue functionally mapped with episodic memory tasks have linked specific gene expression networks with oscillatory signatures of memory formation across multiple frequency bands⁵². Indeed, the ability of the ex vivo approach to dissect the cellular architecture of brain tissue whose networks have been functionally mapped by Neuropixels or ECOG during human cognitive tasks (i.e. language), is a key distinction between human neuroscience and other model systems.

Studying neurosurgically resected tissue in culture also provides a window to study subtype-specific synaptic connectivity maps, which will be critical in linking cell types to circuit and network function. Although advances in electron microscopy (EM) will soon enable whole mouse brain connectome studies⁵³, achieving the scale necessary for human brain may be a lot more challenging. Moreover, strategies to visualize molecular cell types in EM workflows are less well developed. Virally introducing synaptically-connected molecular barcodes in ex vivo slices, and then profiling their connectivity with next-generation single-cell sequencing technologies, may enable us to overcome the limitations of cell-type identities and scale seen in EM connectome strategies⁵⁴. Although ex vivo virally-mediated barcode tracing strategies inherently narrows study to the local microcircuit within each slice, the integration of cellular identity and single-cell transcriptomics with their synaptic connectivity maps may highlight intrinsic differences in how endogenous network activity motifs differ in disease states^55–57.

To take a step further, we need to find efficient ways to relate in vivo cellular activity to cell type identity, such as developing approaches that can register the location of a cell during in vivo recording and determine its identity ex vivo. By preserving the location of the cells from in vivo recordings to their ex vivo positions, large multidimensional datasets that incorporate neurophysiology with single-cell sequencing techniques (transcriptomics, chromatin state profiling, and epigenomics) can be built. This can to some degree be accomplished using extracellular electrodes⁵⁸. An alternative approach may be to leverage in silico models, which provide biorealistic approximations of brain function. Success of biorealistic in silico models trained with data from humans provide a representation, and predictions from the models can be tested as validation^59,60.

Several issues need to be considered with the ex vivo tissue culture approach including 1) the generalizability of findings from tissue collected in epilepsy surgeries, 2) the additional technical variation of cell health in tissue culture, and 3) the constraints of studying a microcircuit in isolation. Since the majority of our tissue comes primarily from epilepsy patients, an important caveat is that conclusions from these studies may not routinely be generalizable to the broader patient population. However, this limitation does not necessarily discount the utility of insights derived from this patient cohort. At our institution, the majority of the patients undergo focused temporal lobectomies where the primary epileptogenic pathology is in subcortical structures such as the hippocampus and amygdala, and not necessarily in the cortex. Moreover, because experimental conditions are compared across epilepsy patients rather than control cohorts, other control methodologies can be used, such as matched post-mortem gene expression samples from epilepsy and control individuals⁵². In fact, the heterogeneity of epilepsy patients may itself pose a unique research opportunity since this patient population often harbors co-morbidities such as depression, anxiety, or autism, which may also be of interest. Furthermore, multiple non-neuronal tissue types such as endothelial cells of the cerebrovascular system may also be studied in tissue from epileptogenic patient population and be more broadly generalizable⁶¹. Similar considerations of generalizability would also apply for tissue collected during resections of deep brain tumors. Overall then, ex vivo human neuroscience approaches can be used to stimulate growth in the field and promote hypotheses which can be supported with other techniques including gene association, exome sequencing, or proteomic studies. Regarding technical variations introduced by tissue culture, neurons from cultured brain slices have been demonstrated to have stable electrophysiological properties over several weeks and are similar to acute slices as evidenced by patch clamp electrophysiology^47,62,63. Finally, although the ex vivo approach results in microcircuits lacking its native in vivo inputs, we can still study questions of subtype composition, synaptic connectivity, cell-intrinsic physiology, and endogenous network motifs, which altogether can provide mechanistic insight into the larger intact network.

Lastly, human ex vivo tissue culture models can provide a number of unique opportunities to advance therapeutic opportunities. First, defining cellular states directly in disease tissue can reveal unexpected cellular, genetic, and intercellular phenotypes that could reveal unexpected disease transformations, changes that correlate with disease progression and inform cross-model studies of animal and in vitro systems⁶¹. The process of developing and validating new vectors for gene therapy such as cell-type specific enhancer discovery⁶⁴ or AAV capsid engineering by directed evolution could be accelerated by incorporating human tissue-based experiments, thereby lowering the risks of such tools failing to translate from mice or monkey to human⁶⁵.

Human neuroscience and clinical practice

Thus far, we have outlined a case for the need and the approaches to advance human neuroscience at the cellular and network levels, and how to connect those dimensions to typical brain function. While these advances yield exciting insights into what makes us human, they also highlight the unmet need to develop strategies to precisely modulate the neural network at the gene, cellular, or circuit level towards the benefit of neurological and psychiatric patients. This journey starts with the need to systematically define cellular function and dysfunction in the context of disease and, with the rapid gains in basic science knowledge, to develop new precision tools for cell- or circuit- specific treatments.

Genetic mapping and precision medicine.

Genomic, transcriptomic, and proteomic studies have defined a starting point to understand neuropathological mechanisms of many neurological and psychiatric conditions. Gene association and exome sequencing studies have begun to identify genetic risk architecture of disease such as with motor neuron disease⁶⁶ and neurodegenerative diseases^67,68, epilepsy^69,70, autism spectrum disorders and neurodevelopmental disorders^71,72, and schizophrenia^73,74. Thanks to the rapid advances in sequencing technologies, genetic diagnosis⁷⁵ is being applied to improve diagnostic capabilities. Identifying variants in risk genes for complex conditions such as Autism Spectrum Disorder and Epilepsy have been increasingly incorporated in clinical work-up^76,77 and can provide guidance when symptoms point to a suspected diagnosis.

However, genomic variants identified at the plasma level explain a small fraction of cases in many neurological and psychiatric cases⁷⁸. The ability to access human brain tissue, such as via neurosurgical resections, has revealed a unique set of mutations that emerge outside the germline and led to a growing appreciation for somatic mosaicism in the role of typical brain development and in neurological disease^79–82. Mutations in this scenario emerge later in development after the zygote state, leading to organ tissues that are mosaic, with only a subset of cells containing the mutation. Such somatic mutations have specifically been observed in tissue obtained from neurosurgical patients, such as those with focal epilepsy who have undergone resection to remove a lesion, such as in cortical dysplasia or hemimegalencephaly^83,84. This has led to the identification of new disease risk genes for focal cortical dysplasias and HME, including AKT3⁸⁵ and DEPDC5⁸⁶, and a converging role of the mTOR pathway in neurodevelopmental-associated epilepsy⁸⁷. These findings would not have been possible without the interrogation of human neurosurgical material.

Gain- or loss-of-function of genes identified in these studies increases the risk for a particular condition, and this information can guide pharmaceutical development, or even repurposing of old drugs. For example, Dravet Syndrome is caused by loss of function mutations in a gene SCN1A, which encodes a sodium channel. Identification of the causal gene has led to practical ways to improve patient outcomes including better treatment strategy (avoiding certain seizure medications) and use of novel pharmacologic agents, including fenfluramine and cannabidiol⁸⁸.

Furthermore, mutational burden information can be integrated with cell type-specific gene expression information profiles to highlight patterns of select cell type vulnerability to disease-causing mutations. One example of this is the identification of TREM2 as a genetic risk factor for Alzheimer’s Disease⁸⁹. TREM2 is selectively expressed in microglia⁹⁰, a finding which has sparked substantial efforts to understand the role of the innate immune system in Alzheimer’s Disease. Of course, gene expression in any one cell type may not be necessary or sufficient to explain the full spectrum of disease symptoms, but identifying profiles of gene expression may represent an important step towards systematic interrogation of cell type specific contributions patient specific mutations to disease processes. Such information can also guide efforts to develop improved cellular models that could serve as platforms to interrogate disease mechanisms and serve as preclinical models.

Although vast majority of single cell reference atlases of human cell type diversity do not extensively characterize mRNA isoform, an important yet underestimated ramification of cell type specific atlases is the potential to identify pathogenic and cell type specific mRNA isoforms. Antisense oligonucleotides (ASOs) are single-stranded synthetic short molecules that bind to target mRNA and lead to target specific knockdown. ASOs have been used for treatment of Duchenne muscular dystrophy^91,92 and spinal muscular atrophy⁹³. The biggest limitations to the use of ASOs are ensuring correct sequence matching to target genes (to optimize gene targeting of all isoforms and prevent off-target effects) and delivery, given these molecules are short-lived and cannot cross the blood-brain barrier. Isoform-resolved single cell transcriptomic atlases could accelerate the identification of potential targets for this emerging class of therapeutic tools.

The ability to integrate the increasing datasets available^94–97 into machine-learning enabled formats⁹⁸ will enable deep learning approaches to combined the genetic, clinical and environmental factors into a more cohesive framework to create tailored and accurate therapeutic strategies.

Cell-type specific molecular access

Emergence of new tools for cell type-specific molecular access to monitor, target and manipulate specific cell types in the human brain open up opportunities to pursue mechanistic and therapeutic studies. Gene therapy offers a direct way to repair or replace a defective gene, an approach that is highly suited for monogenic conditions. Adeno-associated viral (AAV) vectors represent an approach that has been successfully applied to target nervous system tissue^99,100. AAV vectors can infect many cell types, show low immunogenicity, and enable long-term expression of introduced transgene without requiring host chromosome integration. One successful example in neurology is spinal muscular atrophy, a devastating neurodegenerative neuromuscular disease¹⁰¹ caused by mutations in SMN1 or SMN2¹⁰². The positive results from the initial cohort has led to enthusiasm about the potential for in vivo gene therapy, such as via AAV vectors, and has served as a guide for gene therapy trials in other conditions¹⁰³. There are currently trials to treat conditions such as Canavan disease (with mutations in ASPA that makes aspartoacylase), Batten disease (a lysosomal storage disorder due to the loss of tripeptidyl peptidase 1), and aromatic-L-amino acid decarboxylase (AADC) deficiency^104,105. However, most AAVs used in current clinical trials have very broad tropism, and as we begin to better understand neurological disease at the cell type and circuit level, it will become increasingly important to develop cell type specific AAVs. Advances in the engineering of AAV capsids and identification of cell-type specific enhancer elements will allow for more precise AAV vector targeting^106–108 and applicability across species⁶⁴. These next generation of vectors will allow for the delivery of a variety of effector molecules not limited to causal disease genes, but will be increasingly tailored to cargos such as optogenetic or chemogenetic actuators of neural activity.

Ethics and data in human neuroscience

Human neuroscience research depends critically on patient volunteers. Participation should involve informed consent, which describes both the risks for tissue collection or neural device implantation, and also how the data generated can be used and shared with the research community. This is especially important since the immediate benefit from surgical patients participating as research subjects are currently limited and the risks are potentially severe, in both obvious and unpredictable ways. For tissue collection, it is paramount patients understand that participation in tissue collection does not change their surgical plan. Only tissue that is clinically indicated for resection are collected for research. The ethical frameworks for human research with neural devices have been reviewed extensively by the NIH BRAIN Initiative Neuroethics Working Group and discusses the analysis of risk, informed consent, and posttrial responsibilities to research participants undergoing implantation of neural devices¹⁰⁹. With regards specifically to data sharing, the Working Group state that the social and scientific value of data sharing should be weighed against the robust analyses of privacy risks, and that sharing data may be a key posttrial responsibility. In practice, many types of human neurophysiology data are intrinsically difficult to de-identify, particularly the high-dimensional datasets obtained from recording technologies described earlier in this perspective. However, informed consent should recognize that unforeseen advances in machine-learning (ML) and artificial intelligence (AI) may risk patient de-identification, particularly as neurophysiology data begin to be superimposed by the patient’s corresponding transcriptomics, epigenomics, and proteomics datasets. The merging and analyses of this complex dataset will require the computing and analytical power of ML and AI. In this context then, what constitutes as informed consent will need to be revisited as we increasingly rely on these technologies, and the degree to which their conclusions arise from a spectrum of “white-box” or “black-box” approaches (i.e. explainable vs unexplainable models)¹¹⁰. In the near future, an important discussion for all stakeholders is what safeguards to have in place as AI advances towards artificial general intelligence¹¹¹. One way to maintain faith in the informed consent process is for the academic research community to make transparent any partnerships between academics and industry technology companies, particularly since private industries are significant drivers of AI advances.

Recently, Research Opportunities in Humans (ROH) Consortium has developed a series of best practices around engaging patients in research to maintain the integrity of clinical care and ensure voluntariness of participation¹¹². However, further discussions are needed especially around data sharing practices. Considerations around the privacy and ownership of the research data have been quite advanced in other fields such as genomics, where public benefits to datasets involving large numbers of study participants have been quite clear. What are the ramifications for broad dissemination of data? Who should have access to the data? What are the public benefits and what are the risks? Answers to those questions are less clearly defined in neuroscience compared to genomics.

Sharing research findings based on data derived from study participants is generally viewed as safe because the raw data is de-identified, or that the underlying raw data did not carry information that would pose a possibility of re-identifying the study participants. New data sharing policies that set FAIR (findability, accessibility, interoperability, and reusability) principles at the centerpiece of federally funded research are set to revolutionize biomedical research¹¹³, but also prompt a reflection upon the risks of data sharing. Even when data is de-identified, with the advent of very large-scale datasets, the possibility of re-identification becomes a point of concern. Nowhere is this more clear than in the area of genomics, where naturally occurring human genetic variation can be used with some degree of confidence to re-identify an individual¹¹⁴, which is particularly important as commercial providers of genome sequencing create incentives to share genomic material of large fragment of the population¹¹⁵. To account for the possibility of re-identifying study participants, consent forms now often implement explicit language informing the participants about the potential risk that the data generated as a result of their participation could be used to re-identify them, such that an individual signing a consent form is aware of the ramifications. However, as discussed above it is difficult to fully predict how future technologies or analytical approaches might open up new avenues to use the data. To limit the risk of potential misuse of such information, researchers can make study data available through controlled access archives, which impose restrictions on how the data can be utilized. At this point in time, these nuanced considerations have been largely limited to genomic data because of the clear evidence that genetic sequence information is unique to every individual and have implications to clinical management.^116,117.

In genomics, population-scale studies have demonstrated the importance of sequence-level variation as it relates to identifying individuals with genetic risk who could be targeted for screening and early detection, which in the case of cancer mutations can substantially increase survival¹¹⁸. In neuropsychiatric disease, recent investigations in somatic copy-number variants have detected potential genetic loci in increasing schizophrenia risk¹¹⁹. In terms of affecting medical management, multiple examples exist where information of a patient’s genetic background can lead to a more precise treatment plan, such as changes in clopidogrel metabolism based on CYP2C19 genotype¹²⁰ and increasing progression-free survival using targeted IDH inhibitors in IDH-mutant glioma¹²¹. Thus, genomics information has been transformative thanks to two properties: genetic sequence information is unique to every individual, and such information has clear implications for disease risk assessment, prevention, and treatment strategy. Another revolution poised to take advantage of genomic variant detection includes recent developments in CRISPR therapeutics¹²².

While neuroscience data has currently no clear equivalents of these two properties, ongoing large-scale data generation efforts will begin to test this assumption. Single cell genomics datasets are already beginning to uncover molecular states of cells associated with diseases such as Alzheimer’s Disease^123–125. Advent of systematic taxonomies of cell types in the human brain will provide a quantitative structure to systematically compare these cellular phenotypes across disorders¹²⁶, and the knowledge generated as a result of such comparisons will likely lead to new therapeutic interventions that precisely target the underlying cells and circuits. Although the full potential of single cell atlases has not been fully demonstrated, there are multiple ways in which such data will likely contribute to addressing treatment needs in neurological disorders, as discussed above.

What is less apparent is how data collected during intraoperative neural recordings should be managed. It is clear that neural activity data carries important information about the underlying disease states, and large-scale datasets collected across hundreds of individuals will likely reveal common patterns of neural function that could provide inroads into personalized therapy. In this scenario data sharing would ultimately benefit individual patients. However, as transcriptomic and epigenomic information are superimposed onto the physiology datasets, human neuroscience may begin to uncover the genetic and physiological codes that underlie our cognitive functions on personal and population levels. The research and broader community will need to be prepared about how to explore these questions since there is real potential for group-level harm and using data to affirm existing biases against certain groups of people. It will require the neuroscience community to engage not just patients, but the general public about the ramifications for broad dissemination of data, the public benefits and risks to different degrees of data access, and the perspectives of different types of communities (race, ethnicity, faith, etc) on how to conduct and apply the findings from human neuroscience research. At first glance, it is sensible that the data sharing of neural activity would be well served to follow similar principles to genomic information. This approach ensures that all findable, accessible, interoperable, and resuable datasets collected from an individual could be removed if requested by the patient. However, patient and community engagement will be just as critical to support a diversification of neuroscience research participants which is required for understanding of human brain function that is inclusive of all people.

The success of human neuroscience

Together, understanding fundamental principles of human brain structure and function, mechanistic interrogation of genetic, cellular and neurophysiological properties of neurons and glia, and the development of tools to precisely monitor, target and replace cells of the brain, we will begin to deliver on the ambitious goals of the BRAIN initiative. But how can we best capitalize on these discoveries and technological advances? What is the future of human neuroscience now that we begin to gain the right tools and mechanisms to study the human brain, and how should we do this responsibly and ethically?

One of the most transformative lessons from the human genome project has been that in order to enable rapid data and knowledge sharing, the scientific community needs to adopt the use of accessible data sharing platforms and data browsers, which organize data in a common reference. Similar efforts are beginning to organize around proposals for cellular taxonomies¹²⁶, but at the current pace of data production, we also predict an even more pressing need to accelerate efforts for adoption of cloud-enabled data repositories¹²⁷, adoption of tools for integrative analysis and annotation, including large language models. Focus of such analyses must include not only the definition of cell types and states but also phenotypes, as opposed to reference taxonomies, as well as development of comparable tools for epigenetic datasets that compare chromatin accessibility. Non-discrete, continuous sources of variation related to cellular trajectories that may exist during development or state transition may pose significant analytical challenges. Responsible data sharing and use policies for genetic variant information must be considered¹²⁸. There is an emerging but unmet need to develop standards of data sharing in neurophysiology that could serve as a foundation to accelerate secure data platforms for analysis of human neurophysiological datasets derived from surgically implanted probes, especially at the time when commercial entities begin to consider adoption of such technologies.

The second key issue is that an approach to science will have to fundamentally change from single investigator driven projects to community-driven efforts, as evidenced by similar collaborative human neuroscience efforts to ours across the United States (i.e. Massachusetts General Hospital, Stanford, Baylor, Southwestern, Cedar-Sinai). This will bring several major benefits. Partnerships involving basic science researchers, computational scientists, physician scientists, and physicians will galvanize projects that interrogate questions in human neuroscience across scales, and across models. We argue that such partnerships should include engagement efforts of communities of interest, including patient groups and study participants who ultimately consent to share their information, data, and tissue information with the trust that such data will be used in a responsible manner, and which could ultimately lead to a better understanding of neurological conditions and neurodiversity, and to development of tools to modulate neural functions. We believe that a revolution in biomedical science has already begun, as evidenced by the formation of BRAIN Initiative Network efforts¹²⁹ to tackle similar challenges in mice and non-human primates, and to begin to set the stage to address these questions in human.

Finally, for all the advances to happen efficiently and equitably, patient volunteers and their families must be invited to the table. Scientists and physician-scientists can ensure respectful and affirming community engagement through partnership with patient advocacy groups, meeting with community partners, and offering accessibility through outreach and educational events^130–132. These practices require preparation of team leaders and members and getting the appropriate skills to perform these tasks. Engagement involves creating trust, allowing for dynamic communication, and ultimately giving back to the community, through the discovery and increased awareness of the neuroscience and neuropathology¹³³. Healing happens with the potential for therapy but also with the acknowledgement and sharing of gained knowledge.

We propose that the era of human neuroscience has begun and will lead to transformative discoveries in fundamental and clinical neuroscience over the next decade, and critical to its success is the ability to build a Human Neuroscience community that embraces the principles of data sharing that have enabled the genomics community to succeed.