Increasing resolution in stress neurobiology: from single cells to complex group behaviors

Abstract Stress can have severe psychological and physiological consequences. Thus, inappropriate regulation of the stress response is linked to the etiology of mood and anxiety disorders. The generation and implementation of preclinical animal models represent valuable tools to explore and characterize the mechanisms underlying the pathophysiology of stress-related psychiatric disorders and the development of novel pharmacological strategies. In this commentary, we discuss the strengths and limitations of state-of-the-art molecular and computational advances employed in stress neurobiology research, with a focus on the ever-increasing spatiotemporal resolution in cell biology and behavioral science. Finally, we share our perspective on future directions in the fields of preclinical and human stress research.


Introduction
Stress is an important risk factor in the development of neuropsychiatric disorders including major depression, anxiety, post-traumatic stress disorder (PTSD), and other mood disorders (Davis et al., 2017;Musazzi et al., 2018;Musazzi & Marrocco, 2016;Sanacora et al., 2022).Elucidating the underlying cellular and molecular mechanisms responsible for the pathophysiology of psychiatric disorders requires the generation and implementation of preclinical animal models.Although unable to fully recapitulate the multidimensionality and complexity of stress-related psychiatric disorders in humans, they represent valuable tools to shed light onto the mechanisms underlying mental health disorders and develop appropriate pharmacological strategies.
Unraveling the complexity of the neurobiological circuits and molecular pathways underlying a healthy or abnormal stress response requires the combination and integration of cellular, molecular, and behavioral data.While traditional approaches lack in-depth spatial and temporal resolution, recent technological advancements have made it possible to improve these aspects considerably (Gururajan et al., 2018).For instance, single-cell transcriptomics allows to probe thousands of genes simultaneously and to dissect the contribution of distinct cell types involved in the stress response.Likewise, the implementation of activity-dependent labeling methods combined with brain clearing techniques, enables to ascertain which cells are activated following specific stressors, and to reconstruct the brain circuits involved in a specific stress-response.Like all "omics" and high-throughput techniques, the implementation of these strategies generates large amounts of data.It is thus fundamental that they are complemented by appropriate computational and statistical tools.As a consequence, the advancement in molecular and cellular neuroscience techniques prompted a growth in the fields of computational science and the development of suitable data analysis software (Wang et al., 2020).In turn, the remarkable computational innovation stimulated a paradigm shift in the context of behavioral phenotyping, bringing about methods to automatically detect and analyze behaviors of interest (Shemesh et al., 2013).This now makes it possible to assess at the behavioral level the specific effect of different types of stressors (e.g.physical, psychological), stress paradigms (acute, chronic), developmental ages (e.g.early life, adolescence, adulthood, old age), and sex (males and females) in a time-effective manner, while considerably reducing manual scoring-related bias.In this commentary, we explore strengths and limitations of state-of-the art methodologies employed in the field of stress neurobiology, focusing on molecular (in vivo) techniques, as well as computational (in silico) tools for both single-cell transcriptomic data analysis and automatic behavioral tracking systems, with an emphasis on the ever-increasing spatiotemporal resolution (Figure 1).While far from devoid of problems, we believe that the correct integration of molecular and computational techniques will greatly contribute to elucidating the role of stress in neuropsychiatric disorders and in designing suitable treatment options.

Single cells: increasing spatial and temporal resolution of the brain
Cells are the essential building blocks of life and are therefore a crucial component to understand the biological mechanisms responsible for health and disease.Understanding the molecular profile of individual cells, for instance on the RNA or protein level, will enhance our understanding of the mechanisms by which stressors are perceived and processed into molecular, neuroendocrine, and behavioral responses under healthy and pathological conditions (Gururajan et al., 2018).In this section, we discuss the current state of the field of stress neurobiology from a molecular neurobiology perspective and provide our view, as early career scientists, on future directions.
First, we examine the ever-increasing resolution of the field at the micro level, in which there is an increased emphasis on information about genes, their cellular entity, and their surroundings.Currently, there is a wide variety of different techniques to investigate the molecular profile of cells.Traditional approaches include western blot, immunohistochemistry, in situ hybridization, and quantitative realtime PCR, among others.Over the years, these techniques have provided great insights into many cellular processes associated with a response to stress across different organs and brain regions.However, these techniques are limited to a small number of genes or proteins that can be measured by a single experiment and require large amounts of input material.Over the last 20 years, the development of RNA sequencing technologies allowed for the quantification of thousands of genes in a single experiment by genome-wide analyses, shaping our understanding of the intricate mechanisms of the stress response in human and animal tissues.RNA sequencing has triggered a paradigm shift, in which a hypothesis-generating approach is utilized to investigate the role of novel molecular targets and their link to stress-related disorders.In addition, RNA sequencing techniques allow for high-throughput quantification of a variety of RNA species, including long-noncoding RNAs and microRNAs, which have now emerged as important regulators of the physiological response to a stressor (Issler et al., 2020;Lin et al., 2021;Lopez et al., 2018).However, these experiments require large amounts of RNA, from thousands of cells, and unfortunately, lack cell-type specificity.New developments in the field of molecular genomics now allow for single-cell and cell-type specificity using single-cell RNA sequencing (scRNA-seq) (Wen et al., 2022), which has emphasized and highlighted the contributions of different cell types in relation to a stressful event.For example, in 2021 we performed a cell type-specific, molecular characterization of all three components of the hypothalamic-pituitary-adrenal (HPA) axis, under baseline and chronic stress conditions (Lopez et al., 2021).In contrast to standard "bulk" RNA-sequencing methods, scRNA-seq allowed us to perform an unbiased characterization of cell types from the paraventricular nucleus of the hypothalamus (PVN), pituitary, and adrenal glands.We use the term "unbiased" here because currently, most cell-types are classified using a handful of established marker genes from the From macro to micro: increasing resolution in stress neurobiology.State-of-the-art molecular and computational advances employed in stress neurobiology research, with a focus on the ever-increasing spatiotemporal resolution in cell biology and behavioral science.The response to stress can be explored at different levels.For example, the type of stressor (physical [Phys] vs. psychological [Psych]), duration of the stressor (acute vs. chronic), developmental stage (early life, adolescence, adulthood, old age), or sex (male vs. female).Created with BioRender.com.literature, rather than using a comprehensive strategy that allows for an objective classification of different cell types that is based on unique transcriptional signatures from hundreds/thousands of genes.It is possible that a more reliable molecular signal of stress response remains elusive because of lack of an unbiased classification of cell types and the absence of better markers for their identification.Since then, other studies have demonstrated the importance of understanding how different cell types respond to an acute or chronic stressor (Dournes et al., 2020;H€ ausl et al., 2021;Kwon et al., 2022;Short, 2021;von Ziegler et al., 2022).Most importantly, these studies have provided extensive datasets as valuable resources for researchers and clinicians interested in the organism's nervous and endocrine responses to stress and the interplay between these tissues.Nevertheless, it is important to point out that scRNA-seq technologies are not without limitations.For example, the results from these experiments are often contingent on analysis parameters and other unaccounted variables, such as sample preparation, dissociation protocols, as well as proportions and sensitivities of cell types.Therefore, it is highly recommended to engage in independent validation of the primary findings to ensure that robust and consistent findings are reported.This may be cost prohibitive, but it is important as many investigators may use these datasets to generate new hypotheses and interpret previous findings.Another limitation of scRNA-seq is the depth of sequencing, which is much less than bulk RNA-seq.An alternative approach to explore cell-type specificity is to isolate a population of cells using a specific marker.This can be accomplished using flow cytometry or antibody-bound methods to enrich for a target population and sequence at greater depth, using bulk RNA-seq.However, it is important to point out that, this approach requires prior knowledge of target cell-type markers, and lacks the exploratory capabilities of single-cell experiments.
Unfortunately, due to the dissociation of single cells from target tissues, standard scRNA-seq techniques lack spatial information (Tian et al., 2022).Over the years, many studies in the stress field have shown the importance of cellular organization and function using well-established techniques, such as immunohistochemistry (IHC) (Hamilton et al., 2018) and fluorescent in-situ hybridization (FISH) (Engelhardt et al., 2021).However, these techniques only allow for a small number of genes to be detected and quantified and thus require specific hypothesis-driven gene targeting approaches.Interestingly, current developments in FISH have allowed for a significant increase in the number of targets, with more than 2.000 genes that can be labeled in one slice using enhanced electric FISH (Borm et al., 2022).Furthermore, recent improvements in sequencing methods now allow for single-cell transcriptomic analyses with spatial information and resolution (Moffitt et al., 2022).Importantly, these new technologies do not aim at replacing non-spatial techniques, but can often be seen as complementary.Along these lines, spatial transcriptomics in particular can also help annotate already available single-cell expression data.For example, Maynard et al. (2021) analyzed gene expression across the six layers of the human dorsolateral prefrontal cortex.They not only identified genes that were differentially expressed in specific layers, but also used their data to improve the annotation of previously obtained, non-spatial datasets.Approaches like these could add information on existing data in other regions of the brain, the HPA axis, and the immune system, to name a few.Considering the advancements in FISH and scRNA-seq, in the future we can expect a significant increase in the molecular resolution at which we can assess how stress exposure influences changes in the expression of genes and their respective cell types.
A major drawback of these techniques is that the main outcome of the experiment remains a snapshot of the stress response in a tissue of interest, at a specific moment in time.This is a critical limitation, as the effects of stress exposure can vary substantially across different time points.Most importantly, these techniques cannot distinguish between those cells which are engaged directly during and after exposure to stress to those that remain unengaged.Obtaining activity-dependent information will be critical when investigating the response to a stressor.In the next section, we highlight several techniques that have been developed to capture the activity-dependent state of cells after exposure to a stimulus, within and across brain regions.

Activity sensors: understanding the individual role of brain cells and circuits in stress
Lack of spatiotemporal resolution of neuronal activity is a major problem for the precise dissection of brain circuits (Gururajan et al., 2018).Exposure to a stressful event triggers cellular activation in multiple temporal waves across different cells within a set of brain regions, which ultimately drives a neuroendocrine and behavioral response.The activity state of cells is an important proxy to investigate the cellular response system (Kawashima et al., 2014).Stressors can activate a spatially scattered subset of cells within homologous brain regions, which emphasizes the importance to distinguish cells based on their activity patterns.In response to cellular activation, different cell populations will use electrical and chemical synapses to communicate with other cells.Chemical synapses release one or several different neurotransmitters (NTs) and neuromodulators (NMs), many of which are related to the stress response system, such as norepinephrine and corticotropin-releasing factor (Deussing & Chen, 2018;H€ okfelt et al., 2018).Several techniques have been developed to explore the activity of neuronal networks, such as microdialysis and mapping of brain networks using immediate early genes.These techniques have provided important insights into the different brain regions activated in response to specific stressful events, but have limited celltype specificity and high spatiotemporal resolution.
Promising and recently developed techniques are now aiming to provide new information to explore the activity of individual cells and neuronal circuits within a network.For example, genetically encoded GPCR activation-based (GRAB) sensors (Feng et al., 2019), reviewed by Wu et.al (Wu et al., 2022) to investigate in-vivo fluctuations of neurotransmitters (NTs) and neuromodulators (NMs).GRAB sensors are highly selective to the NT or NM of interest, and upon binding will change their conformation, so a fluorescent signal can be detected.The GRAB sensors can target specific cell populations by using cre-dependent labeling, which opens up the possibility to investigate the contributions of different cell types within the system.In addition, they are able to detect NT and NM fluctuations within the millisecond time window.This makes the GRAB sensors a strong tool for investigating the cellular response to stress in the brain, as it has high molecular selectivity and temporal sensitivity.Another promising technique is the implementation of activity-dependent labeling methods.Genetic labeling of neurons, with a specific response feature, is an emerging technology for the precise dissection of functionally heterogeneous brain circuits.Immediate early gene mapping has been widely used for decades to identify brain regions that are activated by external stimuli, however high spatiotemporal resolution has proved to be time consuming and extremely laborious (Franceschini et al., 2020).A recent study used multiple cohorts of mice, sacrificed at different time points, after exposure to a particular stressor and highlighted the importance of timing (the temporal component) in stress research, as they observed a specific time-dependent pattern of c-Fos protein expression across different brain regions (Bonapersona et al., 2022).However, the expression of c-Fos is ubiquitous across neuronal populations (Cruz-Mendoza et al., 2022), which limits the information that can be gathered regarding specialized functions of particular neuronal types.In addition, the statistical analysis for such brain-wide analyses using different time-dependent cohorts is complex and highly variable due to the individual differences between cohorts.The recent characterization of the promoter and enhancer elements responsible for neuronal activity-dependent transcription has opened new avenues for the dissection of active neurons, allowing for the characterization of neural ensembles and circuits in greater detail (Kawashima et al., 2014).Using activity-driven labeling, it is now possible to label and track activated cell populations in a specific time window through the brain using viruses and genetic mouse lines, such as the enhanced synaptic activity responsive element E-SARE (Kawashima et al., 2013) or targeted recombination in active populations (TRAP) and TRAP2 (Allen et al., 2017).Most importantly, these innovative techniques can be combined with in-vivo tracking tools, such as electrophysiology, optogenetics, DREADDs, calcium imaging, as well as GRAB sensors to provide a deeper understanding of a healthy and abnormal stress response.Currently, only a handful of studies have investigated the stress response system using activity-dependent labeling and in-vivo tracking tools (Koutlas et al., 2022;Niu et al., 2022;Ramirez et al., 2015).For example, an interesting study using first-generation TRAP mice was conducted by Ramirez and colleagues, who showed that the reactivation of dentate gyrus cells, which were previously labeled during a positive experience, can rescue stressinduced depression-like behaviors (Ramirez et al., 2015).However, more recent studies utilize a new generation of TRAP mice (TRAP2), which allow for a more specific signal, only in neurons of interest.More specifically, using TRAP2 mice, Koutlas and colleagues showed that stress-activated neurons in the ventral tegmental area have different electrophysiological properties, as compared to non-activated neurons in the same region (Koutlas et al., 2022).Furthermore, combining calcium imaging tools with activitydependent labeling would allow for exclusive investigation of cellular plasticity from stress-responsive cells and exploration of their activity properties at different time scales, from immediate (acute) to long-term (chronic) effects.An important remark is that the "tagging" of different cell types simultaneously is not possible, which limits the identification and contributions of cell type-specific effects.
While current advances in the field of neurobiology using FISH, scRNA-seq, GRAB sensors, and activity-dependent labeling methods have been aimed at increasing molecular resolution (the micro level), these techniques by themselves do not inform at the level of circuits and networks, as well as interactions across brain regions and communication with other peripheral systems (the macro level).These topics will be discussed in the next section.

Brain and body: investigating whole systems to better understand the stress response
A stressful event triggers a cascade of cellular responses in many different brain regions, and peripheral systems, which in turn influence each other (Dedic et al., 2018).It is crucial to consider the entire brain and body as a holistic entity, to further understand different systems and characterize novel pathways related to stress exposure and response.Biochemical and neuroendocrine data, such as circulating levels of glucocorticoids (GCs) have been used as an important parameter to measure the stress response in animal models and humans.Nowadays, advances in multiplex immunoassays can provide a more holistic view of biological markers (e.g.GCs, cytokines, catecholamines, vasopressin, among others), and even distinguish markers related to different types of stressors (e.g.acute versus chronic stress) (Ataallahi et al., 2022, Tighe et al., 2015).In addition, several physiological measurements are now used to investigate how stress responses can promote energy reallocation to support survival.For example, metabolic cages allow for quantification and exploration of several physiological parameters, such as weight, respiratory exchange rate, and energy expenditure, which have been found to be differentially altered between different stress paradigms (Kuti et al., 2022).Most of these measurements are readily available from numerous human psychiatric and preclinical studies.Being able to integrate this data, into the spectrum of single cell -whole brain studies can increase translatability across species and studies.
Moreover, a possible way to obtain a more comprehensive view of stressed-induced alterations in the brain is by using a series of slices through the entire brain and labeling the expression of immediate-early genes using IHC or FISH, which has provided interesting insights into a brain-wide analysis of different cellular targets related to stress exposure (Scharf et al., 2011;Silva et al., 2019).However, as previously stated, these approaches are extremely laborious, time-consuming, and are limited to a small number of genes that can be detected and quantified.Another approach is the use of magnetic resonance imaging (MRI), in which whole brain activity can be obtained in a single experimental setup, which provides insight into the activation and communication of particular regions across the brain (Mandino et al., 2019).Unfortunately, using MRI animals can only be tested under deep anesthesia, which severely limits external manipulations, such as natural exposure to stressors during experimental recordings.An alternative method to investigate whole-brain activity using cerebrovascular fluctuations is functional ultrasound imaging (fUS), which enables in-vivo recordings without anesthesia (Deffieux et al., 2021).Next to investigating the blood flow changes across brain regions, it is crucial to be able to investigate activity patterns across the brain at higher resolution in order to trace and investigate the activated circuits at the single-cell level, which cannot be achieved using techniques, such as MRI and fUS.A technique that can explore cellular activity across the entire brain, while providing single-cell resolution is brain clearing.Brain clearing has been rapidly advancing with different methods, such as CLARITY (Chung et al., 2013) and iDISCO (Renier et al., 2014).These techniques help us visualize protein expression throughout a cleared tissue, such as the entire brain, at an incredible cellular resolution.Nevertheless, we believe that to successfully capture the complexity of the stress response, the combination of these different techniques will be crucial.For instance, combining cell type specific methods, with activity-driven labeling tools, and brain clearing techniques will provide a more well-rounded view of the brain during or after exposure to stress.Ultimately, this will allow us to use a more unbiased method to investigate specific brain regions, cell types, and cell populations related to the stress response system.An excellent showcase for combing these tools are recent studies published by Niu et al, in 2022(Niu et al., 2022), and Davoudian et al., 2023(Davoudian et al., 2023).In the first, the authors start their study using whole-brain imaging after restraint stress then narrowed their focus to a few identified stress-responsive regions, including a novel target in the claustrum.Subsequently, they labeled a stressresponsive neuronal ensemble in the claustrum, using activity-dependent labeling tools and observed that the silencing of this neuronal network, using DREADDS, resulted in attenuation of anxiety-related behaviors, whereas the activation of the same network elicited those behaviors.Similarly, in 2023 Davoudian and colleagues employed whole-brain serial twophoton microscopy and light sheet microscopy to map the expression of the immediate early gene, c-Fos, in male and female mice, following administration of ketamine and psilocybin.Their systematic mapping approach produced an unbiased list of brain regions impacted by both treatments.
Furthermore, in the future important topics, such as the influences of sex on the stress response system, the molecular mechanisms and circuits involved in treatment response, or the connection between the central and peripheral nervous systems can be investigated in greater detail using such an approach.For reference, Brivio et al, summarize most of the studies that have established sex differences in the neurobiological and behavioral effects of stress exposure (Brivio et al., 2020).These new tools will significantly improve our understanding of the molecular mechanisms and cellular circuits responsible for the development of stress-related psychiatric disorders.However, the generation and analysis of these increasingly more complex and larger datasets have created great statistical and computational challenges in our field, hence the need for the development and integration of computational tools in the field of stress neurobiology.

Digging deeper: leveraging computational advances to increase resolution in individual data modalities
Many of the questions that the stress neurobiology field is currently trying to address require a joint collection of many data modalities to reach sound conclusions.As technology advances, more data becomes available in different areas such as genomics, transcriptomics, proteomics, circuits, and behavior, to name a few.This renders an apparent need for developing standardized ways of taking advantage of this increased resolution, without losing sight of the big picture (their interaction).Furthermore, not only does this increase in resolution and data volume have value in itself, but it also carries the potential to incentivize the development of new tools that leverage computational developments, tailored to the tasks at hand.For example, the field of behavioral and computational science has witnessed an increasing number of statistical and machine learning tools designed to tackle different arising problems and automate laborious tasks, which has a huge impact in how research is being done to study the molecular mechanisms and behaviors associated with a stress response.In this section, we discuss the current state of the computational field, from a stress research perspective, and illustrate our view on where we think research could move next.
We will start by discussing the field at the micro level (that is, increased resolution in molecular data modalities) which allows us to inspect closer aspects of biology that were inaccessible before.As discussed earlier, increasing data volume and resolution in transcriptomics, has sparked a plethora of tools and methods that can make such analyses manageable.On the single-cell side, programs like SCANPY (Wolf et al., 2018) and SEURAT (Satija et al., 2015) have succeeded in making state-of-the-art processing and analysis accessible to a broad spectrum of researchers.To date, these have accumulated thousands of citations, and the user basis continues to grow.In addition, new extensions that handle new data modalities continue to be released and maintained, such as spatial transcriptomics (Palla et al., 2022), which is helping make unprecedented progress in the study of both tissue organization and cellular communication.Moreover, tools that leverage the ever-growing public datasets to automate even further workloads (for example, automatic cell annotation) are an example of the positive feedback loop these tools generate (Fischer et al., 2021).This level of standardization portrays substantial benefits for several related fields, and stress research is not an exception, with high implications for basic understanding of cell composition and gene expression in relevant tissues to novel drug targets and the development of new treatments.As an example of the latter, in 2022, using a combination of automatic behavioral tracking techniques and state-of-the-art scRNA-seq methods, we identified cell-type-specific molecular signatures, and a previously unknown mechanism of action, for the sustained antidepressant effects of ketamine in glutamatergic neurons of the ventral hippocampus of adult mice (Lopez et al., 2022).We expect that, in the near future, these technologies will continue to shed light not only into cellular mechanisms underlying the action of drugs used in stress-related disorders, but also hint at new potential pharmacological targets that could be exploited in the future.Finally, while still unexplored in the stress field, to the best of our knowledge, the technical advances in spatial transcriptomics could accelerate these findings by providing access to crucial information on cellular distribution and communication within a given tissue.
So far, we have focused on areas in which breakthroughs in the experimental domain have triggered an increase in data volume, which in turn sparked the need for new computational approaches (either completely novel or borrowed from other computational and statistical fields).The case of behavioral analysis, however, follows the opposite trend: here, it was the thoughtful application of recent computational techniques, such as convolutional neural networks (CNNs) and other computer vision advances, which led to a rapid increase in data collection, and ultimately to a drastic change in how research is being carried out and the types of questions people can ask.In the next section, we discuss how precision behavioral tracking is an emerging and exacting new field in neuroscience research.

Precision tracking: automated systems to dissect the behavioral language of rodents in stress research
In 2013, Shemesh et al. (2013) developed an automatic phenotyping system based on video color recognition, known as the "Social Box" (SB).Here, the authors described how social behavior in mice develops in a semi-natural environment, using a set of techniques that aim to quantify behavioral traits in an automated way, thus freeing researchers from the burden of laborious manual quantification.The authors automatically tracked several groups of mice in their home environment and investigated how individual behavior is strongly interdependent in their groups.In a follow-up study in 2019, Forkosh and colleagues developed a model, using the SB system, that captures and outlines stable personality traits in mice (Forkosh et al., 2019).Although undoubtedly insightful, this work and many that followed (Anpilov et al., 2020;Forkosh et al., 2019;Karamihalev et al., 2020;Shemesh et al., 2016) were limited to tracking the central position of each animal.Furthermore, in this and other contemporary approaches, animal identification relied on dedicated (often expensive or invasive) hardware, such as radio frequency identification (RFID) or color hair dyes (Shemesh et al., 2013, de Chaumont et al., 2012).
Many of these issues were addressed in recent years by the development of neural network models that work on image data directly, without the need for physical markers.For example, tools such as DeepLabCut (DLC) (Mathis et al., 2018), Social Leap Estimates Animal Poses (SLEAP) (Pereira et al., 2022), and SIPEC (Marks et al., 2022), have made it possible to gather enormous amounts of time series data on multiple body parts with human-level accuracy (Sturman et al., 2020).In addition, some of these models are now capable of retaining individual identification in social settings, without the need for dedicated hardware (Lauer et al., 2022).A concept we believe is worth mentioning here is transfer learning.That is, leveraging of previously trained models to classify gigantic datasets of unrelated images, which can lead to very good tracking with little (or no) labeling (known as few-shot learning) (Lauer et al., 2022, Ye et al., 2022).Furthermore, this has been shown to work well both in lab environments as in the wild, enabling its use for ethological studies.While marker-less animal tracking is in itself an accomplishment worth mentioning, many tools have become available that can take this one step further, and identify behavioral patterns in motion tracking data in both supervised (Nilsson, 2022;Segalin et al., 2021) and unsupervised (Bordes & Miranda, 2022;Hsu & Yttri, 2021;Luxem et al., 2022) ways, paving the way for automated behavioral screenings, that are both less laborious and more robust than more classical methods.Along these lines, we recently developed and introduced an open-source tool called DeepOF (Bordes & Miranda, 2022), which is capable of reporting interpretable patterns in open field motion tracking data in both supervised and unsupervised ways.The study emphasizes the importance of such analyses for stress research, as we showed how DeepOF can be used to detect distinct stressinduced behavioral patterns following chronic social defeat stress.In particular, we see (in a fully unsupervised way) an increase in huddling and escaping behaviors in chronically stressed mice, and an enrichment in exploratory patterns in controls.Moreover, DeepOF can detect habituation to nonhostile environments, reporting how behavioral differences between stressed and non-stressed mice fade over time.In a recent publication, Shemesh and Chen review different novel systems fit to investigate the behavior of rodents and discuss what they deemed as a paradigm shift in translational psychiatry through rodent neuroethology (Shemesh & Chen, 2023).The authors suggest that these new methods possess the best out of classical ethology and the reductive behaviorist approach, which may provide a breakthrough in discovering new efficient therapies for mental illnesses.
All these developments can have large implications for stress research.First, by measuring the position over time of one or more markers in less restricted environments, scientists can increase throughput, since extracting information from freely moving animals makes it easier to replace expensive and time-consuming batteries of univariate tests, while significantly reducing the large numbers of animals needed to accomplish the task.In addition, stress research is a field in which the leap between human conditions and their animal equivalents is often significant and questioned.Since these tools allow for more complex data-driven definitions of the outcomes we intend to measure, they have the potential to increase construct and face validity.For example, earlier this year we, as part of a larger group, proposed an algorithm to identify a 'depression-like syndrome' in mice based on mappings from both DSM-V and ICD-11 (von M€ ucke-Heim et al., 2022).While rodent behavior will still remain a distant proxy of their human counterpart, given that factors such as social, economic, and inferential features are hard (if not impossible) to model, we believe efforts like this, which yield clearer, standardized definitions of preclinical phenotypes, will be extremely important for the field moving forward.

Translational tools: novel methods to improve translatability in stress research
All in all, both omics and motion tracking examples illustrate well how, in our view, having more data can lead to increased resolution and, in turn, accelerated discovery.However, to date, they are mostly applicable to animal models.Given that the focus of stress research is, at the end of the day, understanding and improving the life quality of people, translation and research in humans are of course crucial.In this regard, developments in understanding human behavior using virtual reality (VR) are worth mentioning.
VR currently allows researchers to track movement with precision in carefully created environments, making it possible to translate paradigms such as fear conditioning to human subjects in a noninvasive way (Binder et al., 2022, Binder & Spoormaker, 2020).Furthermore, imaging techniques like structural and functional magnetic resonance imaging (MRI), promise to accelerate translation by enabling data generation directly from human brains.While low test-retest reliability and potential construct validity issues coming from the heterogeneity of the labels that researchers use to study stress-related psychopathologies (Miranda et al., 2021), we believe that data-driven initiatives such as the research domain criteria (RDoC) (Insel et al., 2010, Morris et al., 2022), together with large scale multi-site data collection efforts (such as PRONIA) have enormous potential on bringing these promises closer to clinical reality (Haidl et al., 2023;Luutonen et al., 2013;Popovicet al., 2020).Furthermore, specific tools such as Neurominer (Koutsouleris, 2022), provide state of the art machine learning tools for brain imaging data with littleto-no code, which can be helpful in bringing this kind of expertise closer to doctors, in search for multivariate patterns that may aid diagnosis, prognosis prediction, and treatment optimization of stress-related disorders.Finally, while these developments have led on their own right to exciting research in the field of stress, they are limited to extending single data modalities.Understanding the stress response system goes far beyond understanding single cells, neural activity or behavior alone, and we believe that the key in the near future will rely on data modality integration.

Reaching broader: gaining integrated knowledge by combining multimodal data
A living system is far more than the sum of its parts, with different biological levels interacting and regulating one another constantly in complex ways.From genetics, transcriptomics, epigenomics, and proteomics, to neural signaling, behavior, and environmental factors, being able to merge information acquired at different biological levels in clever ways can be key to understanding any phenotype (Stahlschmidt et al., 2022).This can help exploit the available data more efficiently, and lead to more holistic research questions.Moreover, a healthy response to stress depends on the interplay of many regulatory factors acting at several interdependent levels, which result in the allocation of energy resources to resolve the stressor situation (Russell & Lightman, 2019).Efforts in both describing this response, and understanding how it's altered in stress-related disorders in a multimodal way can help disentangling individual differences between, for example, susceptibility and resilience toward stress exposure or response and non-response to antidepressant treatments.
At a basic level, multimodal integration requires researchers to draw conclusions of experiments describing multiple (complementary) axes of the same problem, and drawing conclusions explaining all observed patterns.von Ziegler et al. ( 2022), for example, used a combination of proteomics, phospho-proteomics, bulk and single-nucleus RNAseq, and TRAP sequencing, to describe the temporal response in the mouse hippocampus to acute stress induced by forced swimming.By exploring all data types independently and taking prior knowledge into account, they provided a comprehensive analysis of the temporal dynamics involved.While undoubtedly useful, this approach may not scale to larger and more complex datasets, as researchers would have to learn their joint properties by hand.Furthermore, experiments may have different sensitivities, time scales, and intrinsic artifactual limitations, which in turn highlights the need for technologies capable of storing, handling, and automatically reporting joint features from multimodal data.Along these lines, several extremely relevant subfields for stress research, such as omics, are flourishing with options for researchers to benefit from.The recently published MUON package (Bredikhin et al., 2022), for example, aims at providing accessible and scalable storage and manipulation of multiple omics layers, where different modalities can be organized and analyzed with ease.The also recent tool MEFISTO (Velten et al., 2022), for instance, can then be used to map all modalities to a shared embedding space, using latent factor analysis.Interestingly, these tools are even capable of leveraging spatial and temporal dimensions, when available.
Another key point where advances are promising is the integration of behavioral and neural data.This remains key for studies going from basic neuroscience to psychiatric research (including stress), as finding neural correlates of adaptive behavioral patterns can pave the way to gain mechanistic insights into the mechanisms causing pathology or drug action.Along these lines, the recently presented software CEBRA (Schneider et al., 2022) promises to be of great help.Using a representation-learning approach, the package is able to report non-linear neural correlates of behavior, directly enabling questions regarding how one affects the other in complex ways that may be difficult to detect without computer assistance.

Concluding remarks
Here, we have discussed the strengths and limitations of state-of-the-art molecular and computational advances employed in stress neurobiology research, with a focus on the ever-increasing spatiotemporal resolution.Overall, we expect these types of molecular techniques and computational tools to encompass more combinations of modalities as the field matures, and increasing high-quality data becomes available.We want to highlight that with many technological advancements making the integration of these datasets possible, especially those involving complex, blackbox models, explainability and interpretability of results is key to avoid reporting non-generalizable results that may ultimately be prejudicial to the field.While, in some cases, research can arguably directly inspect results visually (as it is the case for motion tracking), anything involving making predictions that rely on biological data should be thoroughly tested, especially if the underlying dataset is small or too specific, to make sure that our models are not learning from noise, or undetected confounders.Fortunately, the research community is becoming more aware of this issue, and both tools and best practices guidelines (Goodwin et al., 2022;Luecken & Theis, 2019) are being published to sort it out.As it is already the case with artificial intelligence in healthcare as a whole, we expect this topic to be on the spotlight of stress research as available tools become more complex.Finally, while the real impact on stress research remains to be explored, we strongly believe that integrating multimodal and complementary datasets will shed light over patterns too complex for humans to interpret directly, but relevant to ultimately understand and treat such a complex phenomenon.From single cells to social behavior, the dream of jointly mapping stress response as a whole is closer than ever.

Figure 1 .
Figure1.From macro to micro: increasing resolution in stress neurobiology.State-of-the-art molecular and computational advances employed in stress neurobiology research, with a focus on the ever-increasing spatiotemporal resolution in cell biology and behavioral science.The response to stress can be explored at different levels.For example, the type of stressor (physical [Phys] vs. psychological [Psych]), duration of the stressor (acute vs. chronic), developmental stage (early life, adolescence, adulthood, old age), or sex (male vs. female).Created with BioRender.com.