An automated platform for Assessing Working Memory and prefrontal circuit function

Abstract

Working memory is a process for actively maintaining and updating task-relevant information, despite interference from competing inputs, and is supported in part by sustained activity in prefrontal cortical pyramidal neurons and coordinated interactions with inhibitory interneurons, which may serve to regulate interference. Chronic stress has potent effects on working memory performance, possibly by interfering with these interactions or by disrupting long-range inputs from key upstream brain regions. Still, the mechanisms by which chronic stress disrupts working memory are not well understood, due in part to a need for scalable, easy-to-implement behavioral assays that are compatible with two-photon calcium imaging and other tools for recording from large populations of neurons. Here, we describe the development and validation of a platform that was designed specifically for automated, high-throughput assessments of working memory and simultaneous two-photon imaging in chronic stress studies. This platform is relatively inexpensive and easy to build; fully automated and scalable such that one investigator can test relatively large cohorts of animals concurrently; fully compatible with two-photon imaging, yet also designed to mitigate head-fixation stress; and can be easily adapted for other behavioral paradigms. Our validation data confirm that mice could be trained to perform a delayed response working memory task with relatively high-fidelity over the course of ∼15 days. Two-photon imaging data validate the feasibility of recording from large populations of cells during working memory tasks performance and characterizing their functional properties. Activity patterns in >70% of medial prefrontal cortical neurons were modulated by at least one task feature, and a majority of cells were engaged by multiple task features. We conclude with a brief literature review of the circuit mechanisms supporting working memory and their disruption in chronic stress states—highlighting directions for future research enabled by this platform.

1. Introduction

Prefrontal cortical (PFC) circuit dysfunction has emerged as an important pathological finding in depression and other stress-related neuropsychiatric disorders, and has been linked to characteristic deficits in regulating attention, reward processing, and executive control (Davidson et al., 2002; Pizzagalli et al., 2014; Drysdale et al., 2017; Ressler and Mayberg, 2007; McTeague et al., 2017; Williams, 2016). All of these processes depend on working memory: the short-term storage and manipulation of information mediated in part by PFC pyramidal neurons, which exhibit sustained activity during active maintenance of a memory trace (Goldman-Rakic, 1995; Postle, 2006; Baddeley, 2003). This unusual capacity for sustained, internally generated activity is thought to emerge from precisely coordinated interactions between excitatory pyramidal cells and inhibitory interneurons (Goldman-Rakic, 1995; Lewis et al., 2005). Inhibitory interneurons may serve to regulate interference from competing, task-irrelevant information, but it has been technically challenging to test this hypothesis directly.

Converging evidence suggests that these interactions may be disrupted in chronic stress states and depression. Working memory deficits are a common feature of chronic stress states, depression, and other stress-related psychiatric disorders (Mizoguchi et al., 2000; Arnsten, 2009; Roozendaal et al., 2009; McEwen, 2007). These deficits can be highly disabling because working memory processes also support language, long-term memory, and the cognitive control of attention and emotion (Baddeley, 1992, 2000; Murray and Ranganath, 2007; Poldrack et al., 1999; Wagner et al., 2001; MacDonald et al., 2000; Miller and Cohen, 2001; Ochsner and Gross, 2005; Wager et al., 2008). Chronic stress also has potent effects on PFC microcircuits, causing a retraction of the apical dendrites and postsynaptic dendritic spines that receive long-range inputs to the PFC and disrupting functional connectivity in cognitive control networks (Radley et al., 2004, 2006; Holmes and Wellman, 2009; Liston et al., 2006, 2009; Shansky et al., 2009). Stress enhances glutamatergic neurotransmission acutely and suppresses interneuron function (Popoli et al., 2011; Reznikov et al., 2007; Yuen et al., 2009, 2010), and deficits in interneuron markers and GABA availability are consistent findings in post-mortem and clinical spectroscopy studies of depression and other stress-related disorders (Gilabert-Juan et al., 2013; Holm et al., 2011; Sibille et al., 2011; Tripp et al., 2012). Together, these studies suggest that chronic stress may disrupt working memory by interfering with precisely tuned interactions between excitatory pyramidal neurons and specific subtypes of inhibitory interneuron with PFC microcircuits.

Still, the mechanisms mediating chronic stress effects on working memory and related cognitive functions are not well understood, especially at the level of neural circuits. This is due partly to technical obstacles: while extracellular recording techniques in freely moving mice have been a mainstay of neurophysiology studies that have revealed fundamental mechanistic insights, it has been challenging to record from large populations of neurons and obtain definitive information about the identity of the neural sources. This is important, because PFC microcircuit function emerges from the coordinated activity of diverse neuronal subtypes with unique morphological, molecular, and cytoarchitectonic properties (Lewis et al., 2005; Bartos et al., 2007; Buzsáki and Draguhn, 2004; Cossart et al., 2003) and distinct responses to chronic stress (Shansky et al., 2009). Two-photon calcium imaging is an ideal tool for bridging this gap because it allows investigators to record from large cellular populations that can be genetically and topologically subtyped (Helmchen and Denk, 2005; Dombeck et al., 2007; Packer et al., 2014). Already, two-photon imaging and related physiology studies are rapidly advancing our understanding of the mechanisms by which prefrontal circuits support working memory and PFC-dependent cognition (Hanks et al., 2015; Kopec et al., 2015; Abbas et al., 2018; Spellman et al., 2015; Kamigaki and Dan, 2017). However, two-photon imaging experiments are technically challenging and for a variety of reasons detailed below, existing behavioral platforms for integrating two-photon imaging with behavior do not always lend themselves to investigations of how chronic stress affects circuit function in large cohorts of mice that are often required for this type of study.

Here, to bridge this gap, we describe the development and validation of a platform for automated, high-throughput assessments of working memory and simultaneous two-photon imaging, informed by six design goals. First, while many existing platforms have enabled important breakthroughs, they are not always easy to build or learn to use without specialized expertise. Thus, a major design goal was to build an accessible platform that would enable new lab users with relatively limited engineering or animal behavior experience to carry out experiments. A second aim was to be able to train and test a large number of mice rapidly and simultaneously, facilitating future investigations of chronic stress effects on circuit function and behavior, which often require large sample sizes. Third, head-fixation can be a potent stressor in itself, so we wanted to design a platform that would minimize the impact of head-fixation stress on task performance. Fourth, the behavioral platform should be compatible with two-photon imaging. Fifth, although we focus here on working memory, we aimed to design a platform with maximal flexibility for instantiating other tasks and behaviors. Finally, another important consideration was designing a robust system for data acquisition and storage.

Below, in this hybrid original data/review manuscript, we describe an imaging-compatible behavioral platform that achieves these goals, and we provide initial behavioral and circuit-level validation data. We conclude with a brief literature review of the circuit mechanisms supporting working memory and their disruption in chronic stress states—directions for future research enabled by this platform.

2. Materials and methods

2.1. Animals

Young adult (postnatal day 35–90) male and female C57BL6/J mice (Jackson Labs) were used for all experiments. Mice were group housed (2–5 mice per cage) under a reversed light cycle (12 h dark/12 h light, lights on 10:00pm). Mice had ad libitum access to food and water, except during training and imaging, when a restricted water schedule was imposed. During this time, a total of 1–1.5 mL of water was administered to each mouse daily during training and testing, and manually after each behavioral session. Training and imaging took place during the dark cycle to coincide with the mouse's natural activity schedule. All experiments and procedures were overseen by and adherent to the rules set forth by the Weill Cornell Medicine Institutional Animal Care and Use Committee.

2.2. Viral vectors

AAV1-Syn-WPRE40-GCaMP6s was obtained from the University of Pennsylvania Vector Core. The viral titer was typically ∼6 × 10 (Roozendaal et al., 2009) viral particles/mL. The virus was aliquoted into 5 μL aliquots and was kept at −80 °C until used.

2.3. Microprism implantation and virus infusion surgery

Mice were anesthetized using isoflurane (5% for induction, 1–2% for maintenance). We administered dexamethasone (1 mg/kg, i.p.) immediately prior to brain surgery to reduce brain swelling and inflammation and Metacam (1 mg/kg, i.p.) as a prophylactic analgesic. Mice were placed on a microwaveable heating pad to maintain body temperature, and sterile eye lubricant (Puralube, Fisher Scientific) was applied to prevent corneal drying during the surgery. Scalp fur was trimmed using scissors, and the area was disinfected using Betadine and alcohol swabs. The skull surface was exposed by a circular incision (approximately 1 cm in diameter) around the midline using sterile scissors (Fine Science Tools). The exposed surface was irrigated with sterile saline solution, and dried using sterile swabs. The periosteum was bluntly dissected away and bupivacaine (0.05 mL, 5 mg/kg) was topically applied, as a second prophylactic analgesic. The region to be imaged (medial prefrontal cortex) was identified using stereotaxic coordinates (an area centered on AP +1.7 mm, ML ±0.4 mm, DV –1.2 mm). A custom-made circular titanium headplate was attached to the skull using dental cement (C&B Metabond, Parkell Inc). The titanium head plate was then screwed into a custom-built fork fixed to a solid metal base.

An approximately 4-mm circular craniotomy centered over the midline (AP +1.7 mm, ML0mm) was opened using a 0.5 mm burr (Fine Science Tools) and a high-speed hand dental drill (Osada), taking great care not to compress brain tissue or damage the sagittal venous sinus. Gentle irrigation with phosphate-buffered saline (137 mM NaCl, 27 mM KCl, 10 mM phosphate buffer, VWR) was used to clear debris at regular intervals. The dura beneath the craniotomy was delicately removed using fine forceps (Fine Science Tools). Sugi swabs (John Weiss & Son, Ltd) and gentle vacuum suction were used to absorb trace bleeding.

A virus for driving GCaMP6s expression (AAV1-Syn-WPRE40-GCaMP6s) was then injected at a rate of 50–100 nL/min using a Nanofil syringe (World Precision Instruments) with a 33G beveled needle (World Precision Instruments) and pump (World Precision Instruments) mounted onto the stereotactic frame. 500 nL of AAV1-Syn-GCaMP6s was injected for two-photon calcium imaging experiments. After each injection, the needle was kept in place for 2 min to allow time for diffusion of the virus prior to removing the needle from the brain slowly over a 2-min period. Prefrontal infusion coordinates for two-photon imaging through a microprism were (from bregma): anterior-posterior (AP) +1.7 mm; mediolateral (ML) ±0.4 mm; and dorsal-ventral (DV) −1.2 mm.

The microprism to be implanted was a square (1.5 mm), right-angle prism made of borosilicate glass with a reflective aluminum coating on the hypotenuse with a silicon dioxide protective coating (Optosigma). We secured the microprism to a 3 mm glass coverslip (Warner Instruments) using ultraviolet light-activated optical adhesive (Norland, Thorlabs), positioning the prism face at the center of the coverslip and taking care to leave the prism face free of adhesive material. The prism was slowly lowered into the brain using a digital micromanipulator over the course of ∼10 min, until the prism face was flush with the midline and the prism was submerged in the brain. To aid in this process, and to ensure micrometer-precision of the position of the prism, we used a stereotaxic micromanipulator (Kopf) with the prism attached to a central vacuum line via an 18G needle. Veterinary adhesive (Vetbond, Fisher Scientific) was used to form a seal between the coverslip and the skull. A layer of Metabond was then applied for added durability. Metacam (1 mg/kg, i.p.) was administered as an analgesic 24 h after surgery, and as needed thereafter. Weand others have shown that this preparation is well tolerated (Andermann et al., 2013; Low et al., 2014; Pattwell et al., 2016; Moda-Sava et al., 2019). In previous studies (Pattwell et al., 2016), we found there is minimal reactive gliosis beyond 50 μm from the prism face; therefore, we limited our analyses to images acquired from a position >50 μm from the prism.

2.4. Imaging-compatible working memory behavioral platform

2.4.1. Working memory rig and system architecture

Details on the design of the working memory rig and the rationale for important design decisions are described in detail in the following section (“Results”). Briefly, each rig included a platform for head fixation; a freely moving, circular disk that served as a treadmill to reduce head immobilization stress; brushes for delivering different types of whisker stimulation; and lickports for recording the animals’ responses and delivering rewards. The rigs were contained within light- and sound-proofed boxes so they could be simultaneously operated by a single user in a common space. As described in more detail below, to ensure the rigs are easy to build and operate by new users, they were designed to be constructed from commonly available physical hardware components, and they rely on a software framework that is run on a Raspberry Pi controller. Because most two-photon imaging systems include auxiliary hardware for external hardware synchronization via Transistor-Transistor Logic (TTL), we made it a priority to build a system that would utilize the TTL interface to automatically synchronize the two datasets (behavioral data and imaging data) in real time. Additional details are described in the following section. A detailed component list, instructions for assembly, and python code for data running the rigs and acquiring data are available from the corresponding author upon request and will be shared with the community via GitHub prior to publication.

2.4.2. Training protocol

We designed a protocol to train mice to perform a delayed response task in a two-alternative forced choice paradigm. In this paradigm, the mouse was trained to respond to two whisker stimuli (cue A, cue B) by licking a corresponding lickport within a certain time window (2 s). The whisker stimuli were generated by running a soft brush against the mouse's whiskers on one side of its snout (left or right). Mice were trained to respond by licking the corresponding lickport. If the mouse responds correctly, a reward was dispensed through the lickport (a ∼4 μL drop of Kool-Aid with sucrose solution). If the mouse responds incorrectly, a 4-s timeout period was imposed. A new trial commenced immediately after the end of either the reward or timeout period. In order to train mice to perform this task, we designed a protocol involving multiple stages of habituation, shaping, and learning, described below.

2.4.3. Water restriction and habituation

Mice were placed on a restricted water schedule for 3–5 days prior to the training period. Each mouse was monitored daily for any signs of distress. Mice received up to 1 mL water per training session, depending on performance. At the end of the training session this amount was supplemented to a total of 1.5 mL a day. Each mouse's health and weight were monitored daily. Mice that showed distress signs or lost more than 20% of their initial body weight were taken off the restricted water schedule. During the first 3–5 days of the water schedule, the mice were habituated to handling by their experimenter and were placed on the training apparatus without head fixation, to allow them to explore and habituate to the apparatus. Using a 1 mL syringe with a metal feeding tube, reward drops were manually dispensed leading the mice to the lickports, so as to condition the mice to seek rewards around the metal lickports. Two days before the beginning of training, mice were head-fixed, while reward drops were manually dispensed from either one of the lickports, to habituate them to the experience of being head-fixed. If the mice did not retrieve the reward (potentially due to head fixation stress), we let them sit for 2–10 min without disturbing them until they became responsive to the reward. A small proportion of mice (<5%) stopped participating during this training phase and were excluded from further training.

2.4.4. Training phase 1: Cue-response association

The goal of the first phase of training was to have the mice learn to associate the whisker stimulus (“cue”) with licking a specific lickport (“response”) where the reward will emerge. Each trial begins with a 5-s baseline period that also serves as the intertrial interval, which separates consecutive trials. After the baseline period, the cue epoch commences: the whiskers are stimulated by positioning a soft brush adjacent to the left or right side of the mouse's snout, so it could whisk and sense the presence of the brush. After 2 s, the brush is removed, signaling the end of the cue epoch. Because we were not primarily concerned with the encoding of the whisker stimuli in the primary somatosensory cortex, we did not trim the whiskers, nor did we target any specific single whisker or row of whiskers.

During this training phase, there was no delay between the cue presentation epoch and the response epoch. Thus, immediately after the cue offset, both lickports became available, signaling the beginning of the response epoch and indicating that a response is expected. During this training phase, the reward was dispensed through the corresponding lickport, (left whisker cue => lick left lickport, right whisker cue => lick right lickport). However, the mouse was free to lick both lickports and to lick either lickport first, without penalty. To track its learning over time, we recorded the first choice of the mouse and the response time. Thus, trials were marked as correct when the mouse licked the appropriate lickport first. Collecting these data enabled us to calculate the performance of the mouse during the session. (Trials during which the mouse did not retrieve the reward were classified as idle trials and were excluded in this analysis.) Each mouse progressed to the next phase of training when it met a performance criterion of >80% correct trials within a 20-trial window, and >70% overall correct trials during a given session. This first training phase typically lasted for 1–5 days.

2.4.5. Training phase 2: Active cue discrimination

The goal of the next phase of training was to have the mouse learn to actively discriminate between the left and right whisker cues by licking only the appropriate lickport and not licking the incorrect lickport. Error trials occurred when the mouse either a) licked the incorrect lickport first; b) licked both lickports simultaneously; c) neglected to lick either lickport; or d) licked before the response period commenced. In these cases, both lickports were removed (become unavailable), and a 4-s timeout period commenced. At the end of the timeout period, a new trial commenced. Correct trials occurred when the mouse licked the appropriate lickport first. In these cases, the contralateral lickport was removed and a reward was dispensed through the correct (ipsilateral) lickport, to reinforce the cue-response association, and to enable the experimenter to track the trial outcome via the behavioral camera. The mouse's performance was monitored as above, and when it met the above-mentioned performance criterion (>80% correct trials in a 20-trial window, >70% overall correct trials during a given session), it progressed to the next training phase.

2.4.6. Training phase 3: Delayed-response task

The goal of the final phase of training was to have the mouse learn to tolerate a delay period between the cue presentation and response and to maintain a representation of the cue during the delay. Thus, in this phase, a short delay was introduced between the cue and the response epochs. The delay duration was randomly selected from a geometric distribution of delay lengths with increments of 500 ms steps, with an initial maximal delay period of 1 s. No-delay trials were also interleaved, in order to keep the mouse motivated. Mice that met a performance criterion of >80% correct trials within a 20-trial window moved on to longer delays, with increments of 500 ms up to a maximum of 4s. For example, a mouse that performed at 80% correct during a 20-trial window moved on to a maximal delay of 1.5 s within the same training session. When the mouse reached the target delay length and performed >70% correct trials (total session performance) and had at least one interval of >80% correct trials within a 20-trial window, the third and final training phase was complete, and it was ready for imaging.

2.4.7. Behavior data analysis

Performance was calculated as percent of correct trials out of all trials during which the mouse responded (i.e. trials where mice did not lick were excluded from the analysis).

2.4.8. Two-photon calcium imaging

All images were acquired using a commercial two-photon laser-scanning microscope (Olympus RS) equipped with a scanning galvanometer and a Spectra-Physics Mai Tai DeepSee laser tuned to 920 nm. To allow for high-resolution imaging through a 1.5 mm microprism, we used a 10X, 0.6 numerical aperture water immersion objective with an 8-mm working distance (Olympus XLPLN10XSVMP). Fluorescence was recorded through gallium arsenide phosphide (GaAsP) detectors using the Fluoview acquisition software (Olympus) using a green light emission bandpass filter (Semrock). All imaging experiments began by obtaining a low-magnification z stack to locate cell-rich regions for imaging. For calcium imaging experiments, we acquired time-lapse images (image size varies with sample, 30–64Hz) spanning an area of mPFC measuring approximately 1273 μm by 1273 μm (for full frame, 512 x 512 pixels).

2.4.9. Plasma corticosterone assays

To evaluate the degree to which our procedure for habituating and training mice on a treadmill attenuated the neuroendocrine response to head-fixation stress, we obtained plasma corticosterone measurements from three groups of mice. For mice in Group 1, we collected trunk blood samples approximately 15 min after a 30-min task session in habituated well-trained mice, i.e. mice that successfully completed “Training Phase 3” as described above. For mice in Group 2, we obtained trunk blood samples approximately 15 min after a 30-min head-fixation session on the training rig without any habituation and with the treadmill locked. For mice in Group 3, we obtained trunk blood samples approximately 15 min after a brief (∼5-min) session of gentle handling. Plasma corticosterone was assessed using a commercially available enzyme-linked immunoassay (ELISA) kit. Group differences were assessed using Kruskal Wallis ANOVA and post-hoc rank-sum tests.

2.5. Imaging data analysis

2.5.1. Preprocessing

Image time series were segmented into individual cells using custom MATLAB scripts based on an established sorting algorithm combining independent components analysis and image segmentation based on threshold intensity, variance, and skewness in the x-y motion corrected data set, and upper and lower limits for cell size (Mukamel et al., 2009). We manually curated the segmented data to include only somas (some of which also included dendrites emanating from the cell body). Here, we refer to these segments as “neurons''. The final data set contained traces from 4423 neurons recorded from 4 mice. Fluorescence signal time series (ΔF/F change in fluorescence divided by baseline fluorescence) were calculated for each individual neuron. Baseline fluorescence was calculated by finding the 20th percentile in a moving 10-s window. We calculated the fluorescence signal time series in the same manner for a ∼7.5 μm (3 pixels) annulus surrounding each soma to estimate neuropil fluorescence contamination, and subtracted this value from the cell fluorescence time series, as described in previously published work (Andermann et al., 2013; Low et al., 2014). We denoised and deconvolved the normalized traces using a previously described denoising and deconvolution algorithm (Pnevmatikakis et al., 2016).

2.5.2. Analysis

To characterize the functional properties of medial prefrontal cortical neurons in the context of this task, we began by detecting cells that were responsive to different task parameters using a multiple linear regression model, similar to models previously described approaches (Engelhard et al., 2019). Here, we used multiple linear regression to characterize and quantify each task parameter's contribution to the neural activity. The fluorescence trace of each cell, ΔF/F, was the dependent variable, and the predictors were derived from the behavioral events (also described as task epochs below: cue, delay, reward) and the stimuli (left whisker stimulus requiring licking the left lickport, and right whisker stimulus, requiring licking the right lickport). Thus, we had six predictors in our model: three epochs x two stimuli. Below, we refer to each predictor by its respective epoch and stimulus, for instance: “Cue Left”. We used the full-session ΔF/F trace of each neuron, and generated vectors of the same length for each predictor. To generate the predictor we convolved a binary vector with a ‘1’ value in all time points (frames) in which a task event occurred. For example, for the “Cue Left” predictor the binary vector would have ‘1’ values in all time points in which the tested stimulus was left and when the stimulus was present. We then convolved each predictor's binary vector with an impulse response that we generated with GCaMP6s rising and falling kinetics as described elsewhere (Chen et al., 2013). The model we used was:

$$F={\sum }_{j=1}^{N_T}{\sum }_{k=1}^K{\beta }_k{e}_j^k+\epsilon$$

Where F is the cell's ΔF/F fluorescence, $N_T$ is the number of time points, K is the number of predictors, $e_j^k$ the kth predictor convolved function at time point j. We calculated the $\beta$ values for each predictor and cell using a Python package (statsmodels.OLS).

3. Results

3.1. Platform design and software architecture

The development of highly sensitive genetically encoded calcium sensors (Chen et al., 2013), advances in two-photon imaging (Helmchen and Denk, 2005; Dombeck et al., 2007; Packer et al., 2014), and the adaptation of a variety of behavioral paradigms for use in head-fixed mice (Hanks et al., 2015; Kopec et al., 2015; Kamigaki and Dan, 2017) have enabled breakthroughs in our understanding of how circuits and networks support a variety of cognitive functions and complex behaviors. Building on this foundation, in this project we set out to design a new platform to support circuit-level investigations of working memory mechanisms specifically in the context of chronic stress paradigms, and focused on achieving several design goals. These goals were described in detail in Section 1. Briefly, they include 1) accessibility for users with relatively limited engineering or animal behavior experience; 2) automated and scalable for facilitating relatively high-throughput future investigations of chronic stress effects on circuit function and behavior; 3) mitigating head-immobilization stress; 4) compatibility with two-photon imaging; 5) maximal flexibility for instantiating other tasks and behaviors; and 6) a robust software architecture for data acquisition and storage.

Detailed component lists, python code, and setup instructions for the resulting system are available from the corresponding author and will be made available to the community via GitHub prior to publication. Here, we highlight key design features that supported these goals. As noted above, the ability to train a large number of mice rapidly and simultaneously was a key priority. Thus, we designed our experimental platform to be inexpensive and easily scalable. The scalability is achieved by an autonomous software design, in which each behavioral rig operates independently, and occupies a small footprint (Fig. 1A–C). This in turn allows multiple rigs to be controlled via a single computer (Fig. 1D) and operated simultaneously by a single user. The small footprint also allows multiple rigs to be operated in the same workspace. Using light- and sound-proofed enclosures enabled us to simultaneously train multiple mice in a high-throughput fashion, and the rigs could in principle be operated on a benchtop in an open laboratory environment occupied concurrently by other investigators.

Fig. 1.

An Automated 2-Photon Imaging-Compatible Platform for Assessing Working Memory. A) Imaging rig mounted under an Olympus RS-2P system. B) Closeup view of the training rig hardware: two servo motors operating the whisker stimulation brushes, two servo motors controlling the lickports (left and right) angles, standard headphones, a behavioral camera, and ventilation fans. C) Full training rig unit with a Raspberry Pi controller and double syringe pump to dispense rewards. D) Behavioral rig scalable system architecture. A single graphical user interface can control multiple experimental rigs that are connected to the same network. Some of the rigs can be connected to imaging equipment, to collect behavioral data and time-locked imaging data. All behavioral data are uploaded to a central database located remotely. The graphical interface is run on a single desktop or a laptop.

To ensure the rigs are easy to build and operate by new users, they were designed to be constructed from commonly available physical hardware components, and they rely on a software framework that is run on a Raspberry Pi controller. All hardware components including the Raspberry Pi controllers and additional electronic modules are fairly cheap and easy to purchase. Because each rig was controlled by one independent Raspberry Pi controller, the training programs for individual mice could be independently tuned, i.e. the training protocol could be tailored to each mouse based on its progress.

Another important consideration was designing a robust system for data acquisition and storage. Complicated tasks often require long training programs (Guo et al., 2014; Harvey et al., 2012), which yields a considerable amount of behavioral data. Using consistent data structures to store a large amount of data facilitates data organization and backup, data sharing, and analysis. The behavioral data were also collected, uploaded, and stored in a central location in real time so they were available for both online and offline analysis (Fig. 1D).

Several design features were important for enabling two-photon imaging experiments to occur concurrently with behavioral testing. First, the rig was designed to be portable and have a low profile, so that it could be easily installed on a microscope stage and fit beneath even large, high-NA, long-working distance microscope objectives. Second, the design needed to allow for easy temporal synchronization between the behavioral data and the two-photon imaging data collected on a separate system. Because most two-photon imaging systems include auxiliary hardware for external hardware synchronization via Transistor-Transistor Logic (TTL), we made it a priority to build a system that would utilize the TTL interface to automatically synchronize the two datasets (behavioral data and imaging data) in real time.

As noted above, to mitigate the impact of head-fixation stress, each behavioral rig was designed around a conductive rotating disc that functioned like a mouse-initiated treadmill (Fig. 1B). The disc was wrapped in aluminum foil for easy cleaning, and was changed between animals. The conductive disc also served as a part of the lick registration sensor: each lickport was made of a metal feeding tube connected to an input pin, allowing us to record each contact that was made by the mouse. The disc itself conducted electrical current via its axle which was made of conductive metal. The axel was a part of a commutator that maintained electrical contact with the current source. The mouse acted as the switch of this circuit: with each touch to the lickport, the circuit would close, and current would run from the current source to the input pin, triggering an interrupt event that was handled by the software.

Training mice on a rotating disc also allowed the head-fixed mouse to walk over the disc when desired, reducing the sensation of being immobilized. To evaluate the degree to which this procedure attenuated the neuroendocrine response to head-fixation stress, we obtained plasma corticosterone measurements from 1) a group of habituated, well-trained mice (i.e. completed “Training Phase 3” as described above in the Methods Section), in blood samples obtained ∼15 min after a 30-min task session, and we compared them to 2) a group of mice exposed to a 30-min head-fixation session on the training rig without any habituation and with the treadmill locked, and 3) a group of gently handled control mice (see Methods Section for details). This analysis confirmed that our procedure for habituating the mice and testing them on a treadmill significantly attenuated the neuroendocrine response to head fixation stress (Kruskal Wallis ANOVA Chi (Pizzagalli et al., 2014) = 8.29, P = 0.016). As expected, plasma corticosterone levels were increased in mice exposed to head-fixation without habituation or a freely moving treadmill (mean plasma corticosterone [ ± SD] = 256.7 ± 95.2 ng/mL, N = 8), compared to gently handled control mice (mean plasma corticosterone [ ± SD] = 132.3 ± 49.0 ng/mL, N = 7; P = 0.0093 for rank-sum test of Group 2 vs. 3). This effect was significantly attenuated in habituated mice tested on a treadmill (mean plasma corticosterone [ ± SD] = 168.4 ± 48.7 ng/mL, N = 8), such that plasma corticostereone was significantly decreased compared to the head-fixed Group 2 (P = 0.049 for rank-sum test of Group 1 vs. 2) and was not significantly different from the gently handled control mice (P = 0.189 for rank-sum test of Group 1 vs. 3). Although these data do not rule out the possibility that our testing on our task platform may engage the neuroendocrine response or other aspects of the stress response, they do confirm that our procedure for habituation and testing was successful in mitigating the corticosterone response to head-fixation.

Finally, although these rigs were designed to enable us to test head-fixed mice on a delayed response working memory task, it should also be emphasized that they were designed to be highly flexible and easily adapted for use in other tasks. For this project, we used four servo motors to control the angle of the whisker stimulus (left and right), and the angle of the lick spouts (lickports, left and right) through which we delivered the reward. We also used two stepper motors to dispense the reward through each lickport separately. The response was recorded by a lickometer, which was a custom-made electronic circuit for electrical switch operation: mice that made contact with the lickport closed a circuit, which in turn evoked a hardware event monitored by a software interrupt function. However, the software was designed to operate a variety of hardware components and sensors, such as servo motors, stepper motors, LEDs, headphones, speakers, vibration sensors, switches, and beam-breakers, among others, so it could be easily adapted for other purposes. Notably, the rigs are also not limited to head-fixed configurations, and could also be adapted for use in freely moving mice.

3.2. Behavioral validation data

To validate the platform, we first used the behavioral rig described above to train a cohort of mice (N = 10 total) to perform a delayed-response working memory task with varying delay periods (0–5 s, timeline illustrated in Fig. 2A). In the first phase of training, mice learned to associate a given cue with a particular response (e.g. left whisker stimulation => lick left lickport). All 10 mice achieved our training performance criterion (>80% correct trials within a 20-trial window) over a period of 1–5 days. In the next phase of training, gradually increasing delays were introduced between the termination of the whisker stimulation and the presentation of the lickport (Fig. 2A; for details, see Section 2: Materials and Methods). During this phase of training lasting an additional 15 days, the cohort maintained a mean performance rate >60%, with mean performance exceeding 70% for most training sessions, while tolerating a gradual increase in mean delay length over time (Fig. 2B–D). Fig. 2B depicts task performance across all 10 mice. Interestingly, within single training sessions, performance tended to fluctuate over the course of the session, perhaps reflecting fluctuating levels of vigilance, attention, or arousal, as depicted in a representative example in Fig. 2C. Together, these data demonstrate that mice could be trained to perform a head-fixed, two-photon imaging-compatible delayed response working memory task using the automated behavioral platform and protocol that we designed.

Fig. 2.

Behavioral Validation Data. A) Task timeline schematic. B) Behavioral performance during training in a cohort of N = 10 mice. Blue trace: individual mean per-session performance (% of active trials where the mouse has responded), with 95% confidence intervals. Chance performance is represented by the red dashed line. Orange trace: individual mean per-session delay period length in milliseconds, with 95% confidence intervals. Sessions in the shaded region are training phase 1. C) Behavioral performance in a representative mouse during a single, representative session. Tick colors denote cue: purple = left cue, green = right cue. Chance performance is represented by the red dashed line. Performance is quantified as the rolling mean % correct in a 3-trial moving window. D) Performance summary at the end of training for delay periods up to 3000 ms (Sessions 15–20 in Panel B): overall performance, mean delay period, and performance as a function of delay, which was not significantly different for long-vs. short-delay trials. Boxplots denote the median and interquartile range and the whiskers denote the full range of the data. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

3.3. Two-photon calcium imaging validation data

Next, we sought to validate the compatibility of the platform with two-photon calcium imaging. This experiment focused on the medial prefrontal cortex, which has been implicated in supporting working memory in mice and is sensitive to chronic stress (Mizoguchi et al., 2000; Holmes and Wellman, 2009; Liston et al., 2006; Kopec et al., 2015; Abbas et al., 2018; Spellman et al., 2015). To gain optical access to the medial wall of the prefrontal cortex and visualize calcium signals in medial prefrontal cortical neurons, we surgically implanted a microprism in each mouse as described elsewhere (Andermann et al., 2013; Low et al., 2014; Pattwell et al., 2016; Moda-Sava et al., 2019) and injected the medial prefrontal cortex with AAV1-Syn-GCaMP6s, driving the expression of GCaMP6s pan-neuronally (Fig. 3A).

Fig. 3.

Calcium Imaging Validation Data. A) Microprism implantation schematic. The microprism was implanted along the midline, opening optical access to the medial wall of the prefrontal cortex. B) Representative image of a population of cells obtained via two-photon imaging through the microprism during task performance, depicted here as the mean-projection over a short image time series. C) Proportion of cells with activity patterns that were significantly modulated by each of the six task parameters, quantified as the proportion of the total recorded population (mean ± 1 standard deviation).

After N = 4 mice were trained to criterion over a period of ∼15 days as described above, we performed two-photon calcium imaging in 1–2 sessions per mouse, recording from an average of 737 cells per session (range: 391–1059 cells per session; Fig. 3B). Each mouse completed an average of 98 trials per session (range: 35–167 trials per imaging session), meeting the performance criterion that was described above. In total, the resulting dataset included calcium traces collected from a total of 4423 cells and a total of 588 trials in N = 4 mice.

To characterize the functional properties of medial prefrontal cortical neurons in the context of this task, we began by detecting cells that were responsive to different task parameters using a multiple linear regression model, similar to models used in previously published work (Engelhard et al., 2019) and described in detail in Section 2 above. The model consisted of six predictors derived from the task, each corresponding to a task epoch (cue, delay, reward) and the relevant stimulus in that trial (left, right). Because we were interested in reward encoding and its association with cells encoding task features during the delay period, we included only correct trials in this analysis, when reward was available. We generated each predictor by convolving the time series of the epoch, in trials with a specific stimulus, with a GCaMP6s response curve that we generated using the fluorophore's reported kinetics (Chen et al., 2013).

We found that among 4423 cells, activity patterns in 3362 (76.01%) cells were significantly modulated by at least one task parameter (Fig. 3C). ∼30% of medial prefrontal neurons were modulated by the reward epoch, constituting a significantly larger proportion than those modulated by either the cue (∼7%) or the delay epochs (∼2–5%). Similar numbers of cells were responsive to the left (6.76 ± 3.43%) and right whisker stimuli (6.42 ± 4.56%).

To further understand the relationship between a cell's functional properties across task epochs, we examined the fluorescent activity traces of each of the six parameter-modulated populations throughout the trials, focusing on cells with the strongest positive responses during each task epoch and examining its activity patterns across other task epochs. Fig. 4 depicts the activity of each of the six populations throughout the different epochs of the task. Interestingly, most cells with cue-responses time-locked to stimulus presentation were not active during the delay period (Fig. 4A and B). However, ∼⅔ of these cells were subsequently active during the reward epoch, potentially providing a mechanism at the single cell level for linking the presentation of a particular whisker cue with the later receipt of a reward. Similarly, we found that on average, cells that were active during the delay period also tended to exhibit sustained activity during the subsequent reward period (Fig. 4C and D), but most were not active during the cue epoch (i.e. during whisker stimulation). Interestingly, delay-modulated cells also exhibited increasing activity slightly before the delay epoch commenced. Finally, we found that most reward-responsive cells also exhibited increased activity during either the cue or delay period (Fig. 4E and F). Together, these data validate the feasibility of using the platform described above to record from large populations of cells during a working memory task and characterize their functional properties. They also point to multiple hypotheses about the circuit-level mechanisms that support working memory in the prefrontal cortex that could be tested in future studies.

Fig. 4.

Cell Activity Analysis. Task related activity is plotted for subpopulations of cells that were significantly modulated by each of six task epochs. Each upper panel contains a mean trace ( ± SEM) for the calcium signal time locked to cue onset (left) or reward receipt (right). Each lower panel contains a heatmap depicting individual cells over the task timeline, with each normalized to its maximum activity and sorted by the timing of its peak activity. Delay periods varied across trials, thus we present only the first second of delay epoch in baseline-cue-delay panels, and the last second of the delay period in the delay-reward panels. A–B) Cue-responsive cells (left: N = 281 cells, right: N = 261 cells). C–D) Delay-responsive cells (left: N = 69 cells, right: N = 78 cells). E–F) Reward-responsive cells (left: N = 1341 cells, right: N = 1332 cells).

4. Discussion

The materials, methods, and data presented above describe the development and validation of a scalable, two-photon imaging-compatible behavioral platform for studying the circuit mechanisms underlying working memory and its disruption in chronic stress states. Our platform was designed to be inexpensive and easy to build, relying on readily available and widely used parts and components. Because it is scalable and automated, a single user can operate many rigs simultaneously, enabling relatively high-throughput studies involving large cohorts of mice—typically a critical prerequisite for chronic stress paradigms. It was also designed to reduce the impact of head-fixation stress, and to be maximally flexible for instantiating other tasks and behaviors. Our behavioral validation data confirm that mice could be trained to perform a delayed response working memory task with relatively high-fidelity over the course of ∼15 days.

Finally, our two-photon imaging data validate the feasibility of using the platform described above to record from large populations of cells during a working memory task and characterize their functional properties. Activity patterns in >70% of medial prefrontal cortical neurons were modulated by at least one task feature—most commonly receiving a reward. Unexpectedly, most strongly cue-responsive cells were not active during the delay period or vice versa, suggesting that distinct populations of neurons may register task-relevant stimuli versus maintain an internally generated working memory of the stimuli during the delay period. However, many cue- and delay-responsive cells were also re-activated upon receiving a reward, suggesting a potential mechanism at the single cell level for linking external reward-predictive stimuli, internally generated representations of those stimuli, and subsequent rewards. These and other hypotheses could be tested in future studies using the platform described above.

Importantly, while this manuscript is focused primarily on reporting how we developed and validated a new platform for concurrent imaging and working memory assessments, it will enable relatively high-throughput assessments that will allow investigators not only to study the basic circuit-level mechanisms that support working memory but also to investigate how they are disrupted in chronic stress states. With this goal in mind, we conclude with a brief review on this subject and directions for future study.

4.1. Circuit mechanisms supporting working memory

Working memory is a process for actively maintaining and updating task-relevant information, despite interference from competing inputs (Goldman-Rakic, 1995; Postle, 2006; Baddeley, 2003). Pioneering neurophysiology experiments and fMRI studies established a central role for the prefrontal cortex (PFC) in working memory (Goldman-Rakic, 1995; Cohen et al., 1997; Paulesu et al., 1993; D'Esposito et al., 1995; Jonides et al., 1993; Courtney et al., 1997; Sawaguchi and Goldman-Rakic, 1991; Wang et al., 2007; Vijayraghavan et al., 2007). While recent studies suggest that activity-silent mechanisms (e.g. those involving short-term synaptic plasticity) might also be important (Wimmer et al., 2014; Wolff et al., 2017; Liu et al., 2014; Rose et al., 2016), a large body of work indicates that PFC microcircuits support working memory in part by generating sustained activity in pyramidal cells during active maintenance of a memory trace. (Cohen et al., 1997; D'Esposito et al., 1995; Courtney et al., 1997; Constantinidis et al., 2002), supported by inputs from the hippocampus, thalamus, posterior parietal cortex, and other regions (Spellman et al., 2015; Bolkan et al., 2017; Kamiński et al., 2017; Jacob and Nieder, 2014). Still, the microcircuit mechanisms underlying this relatively unique capacity for sustained, internally generated activity are incompletely understood. It has been proposed that this active maintenance process may be supported by recurrent and local connectivity between pyramidal cells (Goldman-Rakic, 1995). Such activity also recruits fast-spiking interneurons and is disrupted by GABA antagonists (Sawaguchi et al., 1989; Wilson et al., 1994), indicating that inhibitory interneurons are also critical. Some studies suggest that specific interneuron subtypes may serve to regulate interference from competing, task-irrelevant information (Goldman-Rakic, 1995; Lewis et al., 2005), although it should also be noted that recent studies suggest that distracting stimuli can be bypassed by other mechanisms as well (Jacob and Nieder, 2014).

Proposed models of PFC microcircuits in delayed response tasks posit distinct clusters of pyramidal cells that encode conflicting representations of task-related information (e.g. Cue A vs. Cue B) (Goldman-Rakic, 1995). Recurrent and local excitatory connectivity between pyramidal cells tends to drive correlated and sustained activity within clusters, but they are also susceptible to interference from competing clusters. It has been proposed that interneurons may support this active maintenance process by regulating interference from competing, task-irrelevant information (Sawaguchi et al., 1989; Wilson et al., 1994).

Complex interactions between many cell types are undoubtedly involved, but converging evidence points to important roles for at least two interneuron subtypes—parvalbumin (PV)- and somatostatin (SST)-expressing interneurons—which may act in concert to support sustained, memory-related activity in PFC pyramidal cells during active maintenance of a memory trace. These cell types are exemplars of two categories of interneuron. SST-interneurons target apical dendrites (Lewis et al., 2005), and may serve to regulate synaptic inputs to pyramidal cells, selectively inhibiting dendritic responses to excitatory inputs (Pi et al., 2013; Kvitsiani et al., 2013; Murayama et al., 2009; Gentet et al., 2012). Optogenetic inhibition of SST interneurons in the PFC during the encoding phase of a working memory tasks disrupts performance and disrupts hippocampal-prefrontal synchrony (Abbas et al., 2018), further reinforcing the importance of this cell type. Together, these studies suggest that SST-interneurons may facilitate working memory by supporting synchronous activity in task-relevant cells and circuits, and suppressing activity in cells representing conflicting, interfering, or task-irrelevant information.

PV-interneurons target the perisomatic region of pyramidal cells, including the axon initial segment (Lewis et al., 2005). They play multiple roles in cortical microcircuits, including facilitating long-range synchrony across neuroanatomically distributed networks (Sohal et al., 2009; Stark et al., 2013; Courtin et al., 2013), regulating experience-dependent synaptic plasticity, (Donato et al., 2013; Kuhlman et al., 2011), balancing excitation and inhibition (Kuhlman et al., 2011; Campanac et al., 2013), and regulating gamma burst activity in the PFC, which in turn gates access to working memory (Lundqvist et al., 2016). Importantly, it has been proposed that PV-interneurons may promote correlated spiking in pyramidal cells clusters by synchronizing spike timing (Pi et al., 2013; Kvitsiani et al., 2013; Lovett-Barron et al., 2012; Atallah et al., 2012), and by disynaptic disinhibition mediated by projections to SST- and other dendrite-targeting interneurons (Kvitsiani et al., 2013). Other studies indicate that working memory maintenance depends, in turn, on sustained activity in clusters of functionally connected pyramidal cells that encode task-relevant information (Wang et al., 2007; Vijayraghavan et al., 2007; Funahashi et al., 1989, 1993). Together, these findings suggest that PV-interneurons may support working memory by enhancing correlated, task-relevant activity (memory-related signal) in pyramidal cell clusters during the delay period.

4.2. Stress effects on working memory and their neurobiological substrates

Converging evidence indicates that a variety of chronic stress models reliably disrupt working memory (Mizoguchi et al., 2000; Hinwood et al., 2012; Yu et al., 2011; Barsegyan et al., 2010; Coburn-Litvak et al., 2003; Roozendaal et al., 2004) and alter multiple structural and functional features of both cortical pyramidal cells and interneurons (McEwen, 2007; Radley et al., 2004; Holmes and Wellman, 2009; Li et al., 2011; McEwen et al., 2015; Dias-Ferreira et al., 2009; Cook and Wellman, 2004). Acutely, glucocorticoid stress hormones increase the excitability of PFC pyramidal cells, enhancing glutamatergic synaptic transmission by increasing presynaptic release and postsynaptic AMPA and NMDA receptor trafficking (Popoli et al., 2011; Reznikov et al., 2007; Yuen et al., 2009, 2010; Musazzi et al., 2010; Venero and Borrell, 1999). Possibly to compensate for this increase in excitability, repeated restraint stress and other chronic strsesors cause postsynaptic spine loss and apical dendritic atrophy in layer 2/3 pyramidal cells (Radley et al., 2004, 2006; Shansky et al., 2009; Cook and Wellman, 2004; Wellman, 2001; Izquierdo et al., 2006), as well as correlated deficits in the cognitive control of attention (Liston et al., 2006, 2009), which is thought to engage working memory-like processes (Miller and Cohen, 2001). We have shown that these effects may be due in part to excessive glucocorticoid exposure, which interferes with spine stabilization, leading to widespread loss of apical dendritic spines (Liston and Gan, 2011; Liston et al., 2013). Since these dendrites are the principal targets of long-range excitatory corticocortical projections, these structural changes may compensate for local glutamate release by reducing excitatory drive to PFC pyramidal cells. In this way, synapse loss on apical dendrites may disrupt working memory either through effects on activity in PFC pyramidal cells, or by disrupting long-range inputs from the hippocampus and thalamus, which are critical for working memory (Spellman et al., 2015; Bolkan et al., 2017; Kamiński et al., 2017; Jacob and Nieder, 2014) and sensitive to stress (Gould et al., 1997, 1998; Vyas et al., 2002; de Kloet et al., 2005). Loss of excitatory postsynaptic spines also disrupts local connectivity between PFC pyramidal cells (Moda-Sava et al., 2019).

Relatively few studies have tested for stress effects on interneurons, but several suggest that chronic stress suppresses interneuron function. Chronic mild stress reduces the frequency of spontaneous postsynaptic inhibitory currents in pyramidal cells (Holm et al., 2011), and repeated restraint stress causes a loss of synaptic proteins in dendrite-targeting interneurons (Gilabert-Juan et al., 2013). Notably, postmortem analyses PFC tissue derived from patients with depression have revealed parallel reductions in the expression of SST and other dendrite-targeting interneuron markers (Sibille et al., 2011; Tripp et al., 2012). While PV-expressing interneurons have been studied extensively in the context of schizophrenia, maternal stress, and other perinatal insults (Lewis et al., 2005; Belforte et al., 2010; Helmeke et al., 2008; Cabungcal et al., 2013; Uhlhaas and Singer, 2010), relatively few studies have investigated how chronic stress affects PV interneurons after the perinatal period. Those that have suggest that chronic stress has no detectable effects on the density of PV-expressing interneurons or PV-positive dendritic material (Zadrożna et al., 2011), and PV-interneurons appear not to be involved in stress effects on depression- and anxiety-like behavior (Pozzi et al., 2014). Similarly, in postmortem analyses of depressed patients, PV expression is unaffected in the lateral prefrontal cortical areas that mediate working memory in humans (Sibille et al., 2011). Together, these findings suggest the hypothesis that chronic stress may impair working memory and disrupt PFC microcircuit function not only by disrupting recurrent and local connectivity between pyramidal cells but also altering dynamic interactions with SST-interneurons, reducing the influence of inhibitory inputs to pyramidal cell dendrites. If so, then interventions aimed at restoring SST-interneuron function may be useful for rescuing working memory and PFC microcircuit dysfunction in chronic stress and depression.

4.3. Conclusions and future directions

Here, we have described the development and validation of a physiology-compatible behavioral platform designed specifically for studying the circuit mechanisms underlying working memory and related cognitive functions in chronic stress states. Mice can be trained to perform the task reliably in an automated fashion in ∼15 days, facilitating relatively high-throughput assessments, which could be critical for studying chronic stress effects on circuit function and behavior, especially for investigators interested in resilience, susceptibility and individual differences. Of note, the same platform could be easily paired with techniques for larger-scale calcium imaging such as light beads microscopy (Demas et al., 2021), which was recently shown to enable mesoscopic, volumetric imaging at multiple scales (including single-cell), enabling simultaneous recording from >200,000 neurons—a potentially critical advantage for dissecting complex, high-dimensional circuit-level coding mechanisms in higher-order association cortex. It is also a convenient complement—but not a replacement—for calcium imaging tools in freely moving mice enabled by advances in one-photon microendoscopy (Ghosh et al., 2011; Cai et al., 2016; Aharoni et al., 2019) and recently extended to include two-photon (MINI2P) imaging (Zong et al., 2022). Imaging in freely moving mice affords a variety of advantages over head-fixed preparations, including the ability to study more complex, naturalistic behaviors that are difficult to model in a head-fixed state.

Our calcium imaging validation data demonstrate the feasibility of this approach and suggest hypotheses that could be tested in future studies concerning the mechanisms by which prefrontal circuits encode and maintain working memory-related information. A brief literature review suggests the hypothesis that PV- and SST-interneurons may support working memory by facilitating sustained, memory-related activity in prefrontal pyramidal neurons and suppressing activity in cells representing interfering, task-irrelevant information. Chronic stress, in turn, may disrupt working memory by disrupting long-range inputs from the hippocampus, thalamus, posterior parietal cortex, and other key regions, and by interfering with interactions between SST interneurons and pyramidal cells within PFC microcircuits.

CRediT authorship contribution statement

Jonathan Witztum: developed the concept for the experimental platform, empirical data, and literature review, collected and analyzed all of the data and generated the figures, conducted the literature review and synthesized the findings. wrote the paper. Ashna Singh: conducted the literature review and synthesized the findings, with input. Rebecca Zhang: conducted the literature review and synthesized the findings, with input. Megan Johnson: conducted the literature review and synthesized the findings, with input. Conor Liston: developed the concept for the experimental platform, empirical data, and literature review, conducted the literature review and synthesized the findings. wrote the paper, All authors reviewed and edited the manuscript.

Declaration of competing interest

C.L. is listed as an inventor for Cornell University patent applications on neuroimaging biomarkers for depression that are pending or in preparation. C.L. serves on the Scientific Advisory Board of Delix Therapeutics, Magnus Medical, and Brainify.AI, and has formerly served as a consultant to Compass, P.L.C. The authors declare no other competing interests.

Handling Editor: Prof R Lawrence Reagan