Deriving causal relationships in resting-state functional connectivity using SSFO-based optogenetic fMRI

Abstract

Objective

The brain network has been extensively studied as a collection of brain regions that are functionally inter-connected. However, the study of the causal relationship in brain-wide functional connectivity, which is critical to the brain function, remains challenging. We aim to examine the feasibility of using stabilized step-function opsin (SSFO)-based optogenetic fMRI (opto-fMRI) to infer the causal relationship (i.e. directional information) in the brain network.

Approach

We combined SSFO-based optogenetics with fMRI in a resting-state rodent model to study how a local increase of excitability affects brain-wide neural activity and resting-state functional connectivity (RSFC). We incorporated Pearson’s correlation and partial correlation analyses in a graphic model to derive the directional information in connections exhibiting RSFC modulations.

Main results

When the dentate gyrus (DG) was sensitized by SSFO activation, we found significantly changed activity and connectivity in several brain regions associated with the DG, particularly in the medial prefrontal cortex (mPFC). Our causal inference result shows an 84–100% accuracy rate compared to the directional information based on anatomical tracing data.

Significance

This study establishes a system to investigate the relationship between local region activity and RSFC modulation, and provides a way to analyze the underlying causal relationship between brain regions.

Introduction

The mammalian brain is a highly organized network that continuously transmits and processes information across interconnected regions [1]. Systematical understanding of brain network organization requires large-scale and specific information in brain connectivity to be accurately measured. A particularly important goal in brain connectivity measurement is to determine the causal relationships between brain regions, as this directional information can be critical in deciphering how information is transferred and processed in a brain network [2]. However, this task remains challenging for most in vivo brain connectivity measurement methods such as resting-state functional magnetic resonance imaging (rsfMRI) [3,4,5].

To establish causality, one needs to find the direction in connectivity between two directly connected regions. Direction of causal effect in a brain network is informative as to whether the change of node A is an outcome of the change of another node B. Conventional rsfMRI analyzing method using Pearson’s correlation gives the strength of linear correlation between two brain regions’ BOLD time series, but it fails to elucidate direction.

The combination of the optogenetic methods and fMRI (i.e. opto-fMRI) potentially offers a solution to infer causal interactions by taking advantage of the relatively high spatial resolution and whole-brain coverage of fMRI and the temporal precision of optogenetic modulation [6,7]. Conventional opto-fMRI causality studies utilize channelrhodopsin-2 (ChR2) to stimulate neurons in a defined region, and measure the corresponding brain-wide responses using fMRI [6,8, 9,10,11]. The causal information can be determined from the optogenetically stimulated region (i.e. source) to brain regions exhibiting fMRI-detected activation (i.e. destination) [12,13]. However, ChR2 activation induces strong neuron depolarization [14] and can overdrive the neurons regardless of their normal activities in the brain. As a result, the conclusions from causal investigations using ChR2 may not reflect the real causal relationships in the resting-state brain.

A potential alternative solution is to use a variant of ChR2— stabilized step-function opsin (SSFO) [15], which does not cause neural firing instantly but induces a subthreshold depolarization of the membrane in a population of neurons, and thereby increases their excitability. Moreover, the active state of SSFO lasts for ~30 min after being elicited by a short pulse of blue light [15]. This feature allows rsfMRI data to be collected after the stimulation light is turned off, which eliminates the tissue heating-related artifacts in the blood-oxygen-level dependent (BOLD) signal [16]. In addition, the activated SSFO can be quickly switched off by a short pulse of yellow light [15], which can be given after the fMRI scan is completed. Taken together, SSFO in theory can provide an ideal tool to study causal relationships of spontaneous neural activities. However, this application of SSFO-based opto-fMRI in deriving causal relationships in resting-state functional connectivity (RSFC) has not been systematically explored [17,18].

In this study, we used SSFO to increase the excitability of the right dorsal dentate gyrus (DG) in awake rats. DG is a functional and structural hub region in the brain networks of both rodents [18,19,20] and humans [21,22]. DG is the first region in the hippocampus that receives sensory projections mainly from the entorhinal cortex [23,24], which gathers the input signal from the rest of the neocortex [19]. Inside the hippocampus, mossy fibers of DG send the encoded information to CA3 [25], which then projects to CA1 through Schaffer collaterals [26]. Dorsal CA1 further projects to the subiculum and deep layers of the entorhinal cortex [26], while ventral CA1 sends projections to the amygdala, nucleus accumbens, and medial prefrontal cortex (mPFC) [27]. These brain regions, in turn, are connected to multiple other regions. Therefore, perturbing DG activity can affect a variety of brain areas, and therefore provide a useful model to test the validity of inferring causal relationships in RSFC using SSFO-based optofMRI.

By locally increasing the excitability in the right dorsal DG (rdDG) with SSFO, we found altered spontaneous BOLD activity and RSFC in local and distal brain regions. Using these data, we further deduced the directional information in a set of brain connections. Our causal inference results in an 84–100% accuracy rate in reference to the directional information based on fiber tracing data [28].

Methods and Materials

Animals

Long Evans rats (250–350g, n = 17) were used in this study. The housing room was maintained at room temperature (22–24 °C) with a 12h:12h light: dark cycle. Standard rat chow and water were supplied ad libitum. Before any experiments, all rats were habituated to the animal facility for at least one week. All experimental procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of the Pennsylvania State University.

Surgery

An aseptic stereotaxic surgery was conducted to inject adreno-associated viral vectors (AAVs) and implant the optical fiber into the rat brain. The animal was first anesthetized with isoflurane (5%) in an induction chamber. A ketamine (40 mg/kg) and xylazine (12 mg/kg) mixture was intramuscularly injected to maintain anesthesia during surgery. Dexamethasone (0.5mg/kg) and Baytril (2.5mg/kg) were subcutaneously injected to reduce inflammation and prevent bacterial infection, respectively. The animal was intubated and connected to a ventilator (PhysioSuite, Kent Scientific Corporation) to support respiration. Body temperature was maintained at 37 °C by a warming pad. Blood oxygen level and heart rate were monitored by pulse oximetry (MouseSTAT^® Jr., Kent Scientific Corporation). For the opto-fMRI experiment (n = 14), AAVs expressing SSFO (AAV2/5-CaMKIIα-ChR2(C128S/D156A)-EYFP, 1 μL, titer ≥ 1× 10^13 vg/mL, Addgene) were injected into the rdDG (coordinates: AP: −3.5mm, ML: 2mm, DV: 3.6–4mm). For calcium signal recording (n=5), AAVs expressing GCaMP6s (AAV5.Syn.Flex.GCaMP6s.WPRE.SV40, 1 μL, titer ≥ 1×10^13 vg/mL, Addgene) were mixed with SSFO viral vectors before injection. Two animals were used in both opto-fMRI and calcium recording experiments.

A fiber-optic cannula (core = 400 μm, NA=0.39, Thorlabs) was implanted above the injection site. Dental cement was used to mount the optical fiber stably on the head. fMRI scanning and GCaMP recording were performed at least four weeks after the virus injection to ensure sufficient protein expression.

Acclimation

Before imaging, all animals were acclimated to the MRI scanning environment to minimize stress and motion during imaging [29,30]. Before acclimation, EMLA cream (2.5% lidocaine & 2.5% prilocaine) was applied to the head and body parts that had physical contact with the restrainer to reduce possible discomfort. After being briefly anesthetized with 3% isoflurane, the rat was placed into a head restrainer secured with a bite bar and two sidebars. Four limbs were restrained in four sockets in the body restrainer. The setup was placed in a black chamber where prerecorded MRI scanning noise was played. Each rat was acclimated for seven days with incremental increases in the acclimation duration (min): 15, 30, 45, 60, 60, 60, and 60. This acclimation procedure has been shown to sufficiently reduce stress and head motion in previous rsfMRI studies [31,32].

fMRI data acquisition

fMRI imaging was conducted on a 7T Bruker 70/30 BioSpec scanner running ParaVision 6.0.1(Bruker, Billerica, MA). Before imaging, the implanted fiber cannula was connected to a laser light source via a fiber-optic patch cord (core = 400 μm, NA = 0.48, 7 m long, Doric Lenses Inc). rsfMRI data were acquired by T2*-weighted gradient-echo images using the echo-planar-imaging (EPI) sequence with the following parameters: repetition time (TR) = 1000ms; echo time (TE) = 15ms; slice thickness = 1mm; in-plane resolution = 0.5 × 0.5mm; number of slices = 20; FOV = 32 × 32 mm²; image matrix = 64 × 64. Each scan lasted for 10 min.

In each imaging session, four consecutive rsfMRI scans were acquired, including a baseline scan at the beginning, followed by two scans after rdDG was activated, as well as the last scan after the activation was turned off (Fig. 1). To activate the SSFO in the rdDG, a blue light (7W/mm²) was turned on for 5 sec at the end of the first rsfMRI scan. At the end of the third scan, rdDG SSFO was deactivated by 5-second illumination of yellow light (14W/mm²).

Figure 1.

Schematic diagram of the experimental paradigm.

fMRI data preprocessing

Framewise displacement (FD) was calculated to estimate the motion level of each rsfMR frame. Volumes with FD larger than 0.25mm (half of the in-plane voxel size) were discarded together with their immediate adjacent volumes. The first ten volumes were also discarded to ensure that the MR signal had reached a steady state. Scans with more than 15% of volumes discarded were removed from further analysis. EPI images were manually aligned to a predefined structural template using an in-lab Matlab software [33]. The manually aligned images were then subjected to motion correction (SPM12). Distortions in EPI images were corrected using the ANTs software (Advanced Normalization Tools, http://picsl.upenn.edu/software/ants/) [34]. Data were then subjected to spatial smoothing (Gaussian kernel, full-width at half maximum = 1 mm) and nuisance regression of the signals from the white matter and ventricles, as well as motion parameters. Independent component analysis (ICA) cleaning was applied to remove non-neural artifacts potentially due to motion, physiological noise, or scanner instability. For correlation-based analysis (i.e. excluding fALFF), the EPI data were also bandpass (0.01–0.1Hz) filtered.

fALFF

fALFF is commonly used to quantify the amplitude of spontaneous BOLD activity, which is calculated as the ratio of the amplitude in a low frequency range (0.01–0.08 Hz) to the amplitude of the detectable frequency range (0–0.5Hz) [35]. For each voxel, the time series was converted to its power spectrum without any prior bandpass filtering. The amplitude spectrum was then produced by taking the square root of the power spectrum. The ratio of the amplitude across 0.01–0.08 Hz to that of the entire spectrum was calculated.

Causal inference

The goal of causal inference is to find the direction in connectivity between two directly connected regions (i.e. directness). Directness determines that the interaction between two nodes is not intermediated by any other node, which cannot be obtained using conventional rsfMRI analysis based on Pearson’s correlation. One approach that may overcome the limitations of Pearson’s correlation is partial correlation, which computes the linear correlation between two regions while controlling for the time series of all of the other regions [36,37].

In our study, a combination of Pearson correlation and partial correlation was used to infer the causal relationships between two regions of interest (ROIs). The brain was divided into 129 ROIs that have the whole brain coverage, including 62 left-side ROIs, 62 right-side ROIs and 5 midline ROIs according to Swanson’s brain atlas [38]. In addition, 2–3 voxels under the optic fiber tip were selected as the 130^th ROI. The time series assigned to each of the 130 ROIs was determined by averaging the time series of the voxels spatially comprising a given ROI. The Pearson’s correlation was then calculated across all pairwise combinations of the 130 ROIs.

Directness

To obtain directness between two ROIs, the ROI time series were also used to calculate the partial correlation across all pairwise ROIs. To calculate the partial correlation between ROI A and ROI B conditioning on C, where C is the collection of all the ROIs except for A and B, we first estimated the unique contribution of each component in C by multiple regression on A and B:

$${\beta }_A=pinv(\mathit{C})\cdot A$$

Where C is the time-series matrix of all the other ROIs, and ‘pinv ‘is a Matlab function to estimate the Moore-Penrose pseudoinverse matrix of C, C⁺ = (C^𝑻C)^−𝟏C^𝑻. The unique contribution of each component was then regressed from A and B:

$$A_{regressed}=A-\mathit{C}\cdot {\beta }_A$$

$$B_{regressed}=B-\mathit{C}\cdot {\beta }_B$$

The partial correlation between A and B was then calculated as the correlation between the regressed time series A_regressed and B_regressed. From the above description, one can see that partial correlation represents the remaining correlation between two ROI nodes after accounting for the influence of all other nodes. Thus, partial correlation can be used to deduce information about directness.

To determine the statistical significance, Pearson’s correlation and partial correlation coefficients were transformed to z values by Fisher’s z transformation, z = artanh(r). The z values were fitted to a linear mixed-effect model (within subject variability modeled as the random effect) to estimate the group-level Pearson’s correlations and partial correlations in t values. The directness between two nodes was determined by the existence of both Pearson’s and partial correlation. To increase the robustness of our results, a series of p-values (p = 0.05, 0.01, 0.005, FDR corrected) were used to threshold the Pearson’s and partial correlation matrices. The partial correlation matrix may have non-zero entries where the corresponding Pearson’s correlation is not significantly different from zero at a given threshold. This is caused by two ROIs conditioning on a collider (See Fig. 6). Thus, partial correlation between ROI A and ROI B is set to zero if the corresponding Pearson’s correlation is not significantly different from zero at a given threshold. This new matrix of updated partial correlation values is referred to as the ‘corrected partial correlation’ matrix, which is used to represent the direct functional connectivity between two nodes (Fig. 6). If a given entry is non-zero, the two ROIs associated with this corrected partial correlation are deemed to be directly connected. If equal to 0, the ROIs are deemed to be not directly connected. For simplicity of the model, the directness of the network is calculated using the ‘Baseline’ data.

Figure 6.

Causal inferences based on Pearson’s correlation and partial correlation. A) Naming of different scenarios regarding the existence of Pearson’s correlation and partial correlation. B) Matrices of Pearson’s correlation, uncorrected partial correlation and corrected partial correlation in ‘Baseline’. C) Illustration of the scenario of colliders. D) Illustration of the scenarios of confounders and chains. E) Illustration of inferring a chain or confounder by partial correlation changes. F) ROI-wise partial correlation comparisons (left: Active 1 vs Baseline; right: Active 2 vs Baseline). Lower triangles represent unthresholded t values, and upper triangles represent thresholded t values (p < 0.05, linear mixed model, FDR corrected).

Direction

Functional connectivity matrices of 130 ROIs covering the whole brain were compared between the baseline scan and the two SSFO activated scans. The Fisher z transformed RSFC values were fitted to a linear mixed-effect model (within subject variability modeled as the random effect) to estimate the group-level differences. The subset of ROIs that showed significant RSFC changes baseline and SSFO-activated states was selected for further analysis of direction.

Direction analysis is based on the basic configurations of collider, confounder, and chain (Fig. 6C&D, [39]). A confounder is characterized by two nodes sharing a common cause but are not themselves connected, such as A←C→B. All three nodes are functionally connected, but the partial correlation of A and B conditioning on C is zero, indicating they are not directly functionally connected (Fig. 6D). A chain configuration, such as A→C→B, is also functionally connected within all three nodes, but A and B are not directly functionally connected since the partial correlation conditioning on the intermediator C is zero (Fig. 6D).

In the case of a collider, such as A→C←B, node A and B are not functionally connected themselves, while conditioning on the collider C results in a pseudo partial correlation between A and B (Fig. 6C, [39]). The direction of a collider is directly inferred from baseline data: when two nodes A and B have a zero Pearson’s correlation and a non-zero partial correlation, it means at least one node among all the nodes regressed against A and B exhibits a collider configuration with A and B. We looked for the set of nodes Ci that has a direct functional connection to both A and B. To confirm that each node Ci is a collider for A and B, we further calculated the partial correlation between A and B conditioning on node Ci. If the partial correlation is significantly different from zero, we will conclude A→Ci←B is a collider. Since a collider already gives the direction, we can infer A →Ci and B→Ci.

In order to find the confounders and chains, we searched for all combinations of three nodes in the baseline scan that have three non-zero Pearson’s correlations and two non-zero corrected partial correlations, such as A-C-B, where A and B are not directly functionally connected. This means it has a possible arrangement of confounder: A←C→B, and has two possible arrangements of chain: A→C→B or B→C→A (Fig. 6E). To further separate those possibilities in confounders and chains, we looked to see if the partial correlation coefficients between A and C, as well as B and C are significantly different between the baseline scan and the SSFO activated scans.

The difference between two states was quantified by paired two-sample T-test of the Fisher Z transformed partial correlation values using linear mixed-effect model (subject variability modeled as random effect). If we observed significant difference of partial correlation in only one pair of nodes, e.g., in C-B, we inferred it is caused by perturbation through the middle node C to its downstream node B in a chain, such that A→C→B was deduced (Fig 6E). If both pairs of nodes, A-C and C-B, showed significant group level differences of partial correlation and their effective sizes are not significantly different (confidence interval overlaps), we inferred it is caused by perturbation through the confounder C to both of its downstream nodes, such that A←C→B was deduced (Fig 6E). When one pair of nodes belong to multiple three-node systems, it is possible that different directions between the two nodes were identified through different pathways. When that happens, we considered the connection between these two nodes as a bi-directional connection, given that they may act differently in different neural pathways.

Assessment of causal inference

In our causal inference, we focused on a subset of connections that showed significant RSFC changes from ‘Baseline’ to ‘Active’ states. To assess our results of causal inference, for each connection in this subset, the direction obtained with our model was compared with the direction of the same connection provided by the cerebral cortical projection database based on anatomical tracing data [28]. Our causal inference for the direction is considered correct if a projection of the same direction was found in the database (i.e. True Positive), and is incorrect if the projection of the same direction was not found in the database (i.e. False Positive). For bidirectional connections inferred by our model, the two directions were separately evaluated. For connections that were not included in the database, the assessment was not conducted. We reported the accuracy of our causal inference as follows:

$$\text{Accuracy}=\text{True Positives/(True Positives + False Positives)}$$

$$\text{False positive rate}=1-\text{Accuracy.}$$

GCaMP recording

A dual-wavelength optical system (Doris Lenses) was used to record the GCaMP calcium signal [40,41]. GCaMP and Ca2+ independent fluorescence signals were alternately excited by a 465 nm LED (Doris Lenses) and an isosbestic wavelength 405 nm LED (Doris Lenses), respectively. The alternating frequency was 400Hz with a duty cycle of 40% controlled by house-made LabView software and the stimulation was maintained at a low power of 10 μW to avoid activating SSFO unintentionally. The implanted optical fiber was connected to the LEDs through a low-autofluorescence optic cable (400 μm, NA = 0.5, Thorlabs). The recorded signal was sampled at 10 kHz with an NI-DAQ board (National Instruments).

GCaMP data analysis

The signal was first detrended by their best-fitting line calculated by airPLS [42]. The 405nm reference signal was then removed from the 465nm Ca²⁺ signal. The signal was low-pass filtered at 10 Hz. Peaks higher than four times standard deviation were selected as the GCaMP signal unless it coincides with a reference signal peak, which is likely due to system instability or animal motion.

Histology

Histology was carried out after the completion of all other experiments to confirm the expressions of SSFO and GCaMP. Before perfusion, the animal was anesthetized with ketamine (33 mg/kg) and xylazine (13 mg/kg). The animal was perfused first with saline and followed by 4% PFA solution. The brain was carefully taken out and fixed in 4% PFA with 30% sucrose at least 1 week before sectioning. The PFA fixed brain was cut into 60 μm thick coronal brain slices by cryosectioning, and the fluorescence signal was examined with a microscope.

Results

SSFO activation in the right dorsal DG increased local neural excitability

We first validated the sensitizing effect of SSFO using GCaMP-based fiber photometry. In this experiment, GCaMP6s and SSFO proteins were co-expressed in the rdDG (Fig. S1). The calcium signal was recorded at the fiber tip implanted in the rdDG in awake, restrained animals (n = 5). GCaMP signals were recorded in four consecutive conditions as shown in Fig. 1 with each condition lasting for 10 minutes. The first 10-minute captured data before SSFO was activated, and thereafter referred to as the ‘baseline’ state. After the initial 10 minutes, the SSFO was activated using a blue light pulse (5 sec in duration, 7W/mm²). Following this pulse, two back-to-back 10-minute recordings took place: respectively referred to as the ‘Active1’ and ‘Active2’ conditions. After the 20 minutes of SSFO-active recording, a yellow light pulse (5 sec in duration, 14W/mm²) was used to deactivate the rdDG SSFO. At this point, the 4th 10-minute measurement took place: referred to as the ‘Deactive’ condition.

The dynamics of the GCaMP signal recorded (Fig. 2A) was highly consistent with the literature report [40,43]. The number of calcium spikes in rdDG increased significantly after SSFO activation, and the activation effect lasted throughout the two ‘active’ recording sessions (Two-sample t-tests, ‘Active1’ vs. ‘Baseline’ p < 0.001; ‘Active2’ vs. ‘Baseline’ p < 0.01, Figs. 2B). After SSFO was deactivated by yellow light, the number of calcium spikes decreased significantly compared to ‘Active1’ (Two-sample t-test, p < 0.05, Fig. 2B).

Figure 2.

Validation of the sensitizing effect of SSFO. A) Top, averaged calcium peak at the tip of the optical fiber; Bottom, exemplar raw GCaMP signals at Baseline, Active 1, Active2, and Deactive conditions. B) Averaged number of spikes at the optical fiber tip in each condition (*: p<0.05, **: p<0.01, ***: p<0.001). C) fALFF at the optical fiber tip in each condition (*: p<0.05, ***: p<0.001).

The data demonstrated that SSFO activation by a short pulse of blue light significantly increased the excitability in the rdDG, and this effect lasts over the time course of the two ‘active’ recordings. Illumination of yellow light also effectively reduced the spiking rate in ‘Deactive’ state by switching off SSFO in our experiment.

Sensitizing the right dorsal DG increased BOLD activity in multiple brain regions

For fMRI sessions, a 10-minute ‘baseline’ resting state was acquired first. rdDG SSFO was then activated by the same blue light, which was immediately followed by two consecutive 10-minute ‘active’ scans. After SSFO deactivation by the yellow light, a fourth and final 10-minute scan took place. There were no differences in the motion level or respiration rate across the four conditions (Fig. S2). Voxel-wise fALFF, defined by low-frequency amplitude (0.01 – 0.08 Hz) normalized to full-spectrum amplitude of the BOLD signal, was calculated to estimate spontaneous BOLD activity across the brain [35]. In concordance with the GCaMP data, the fALFF at the optical fiber tip also increased significantly after SSFO activation (Two sample t-tests, ‘Active1’ vs. ‘Baseline’ p<0.001; ‘Active2’ vs. ‘Baseline’ p<0.01, Fig. 2C) and decreased after illumination of yellow light (Fig. 2C). In addition, SSFO-sensitization of the rdDG increased the brain activity of several other brain regions. The brain regions with increased fALFF in the ‘Active1’ state were mainly located ipsilaterally with the activated rdDG, including the right hippocampus (CA1, CA3, DG), right striatum, right mPFC including anterior cingulate cortex (ACC), prelimbic cortex (PL) and orbital cortex (ORB), and some parts of the sensory-motor cortex (Fig. 3, p < 0.05, linear mixed model, family-wise error corrected). All these brain regions are involved in the DG circuits [19,26]. When comparing the ‘Active2’ state with ‘Baseline’, although increased fALFF can still be appreciated, significant difference in fALFF was diminished under the threshold applied, likely due to reduced SSFO activation and/or increased variance over time. We found no difference in the fALFF maps when comparing the ‘Deactive’ state with ‘Baseline’.

Figure 3.

Sensitizing the right dorsal dentate gyrus increased BOLD activity across the brain. A) Fractional amplitude of low-frequency fluctuations (fALFF) in baseline, SSFO activated and deactivated states. B) Voxel-wise fALFF changes compared with baseline after SSFO activation and deactivation (p < 0.05, linear mixed model, family-wise error corrected).

SSFO activation in the rdDG prominently increased its functional connectivity with mPFC

To study the functional connectivity change in response to the SSFO activation of the rdDG, we selected the 2–3 voxels under the optical fiber tip as the seed region since they are the brain tissue directly sensitized by SSFO activation. Seedmaps of the optical fiber tip of the four conditions (Fig. 4B) show that the difference between ‘Baseline’ and the activated states is mainly located in the forebrain area. Fig. 4C shows the voxel-wise comparison between ‘Baseline’ and the other three states. Both ‘Active1’ and ‘Active2’ showed higher RSFC with the seed in the mPFC (ACC, PL, and infralimbic cortex (IL)) and ORB compared to ‘Baseline’, while after SSFO was switched off, the ‘Deactive’ scan showed no significant difference with the ‘Baseline’ scan (Fig. 4C, p < 0.05, linear mixed model, family-wise error corrected).

Figure 4.

SSFO activation in the rdDG increased its RSFC with the mPFC. A) Structural image showing the position of the seed region at the tip of the implanted optical fiber. B) Seedmaps of the optical fiber tip in different conditions. C) Voxel-wise RSFC changes compared with baseline after SSFO activation and deactivation (p < 0.05, linear mixed model, family-wise error corrected).

Causal inferences based on Pearson’s correlation and partial correlation

To comprehensively elucidate the RSFC changes after the SSFO activation of the rdDG, we compared the RSFC matrix of ‘Baseline’ with the other three conditions as shown in Fig. 5A (p < 0.05, linear mixed model, FDR corrected). 42 ROIs exhibited significant RSFC changes when in the ‘Active1’ state relative to ‘Baseline’ (Fig. 5A, upper panel), and a subset of 10 ROIs maintained altered RSFC in the ‘Active2’ state (Fig. 5A, middle panel). Although a lower number of ROIs exhibiting RSFC changes were detected in the ‘Active2’ condition, again likely due to reduced SSFO effects as shown in Fig. 3B, t-values involving all 42 ROIs in ‘Active1 - Baseline’ and ‘Active2 - Baseline’ comparisons were in fact highly consistent, evidenced by a high entry-to-entry correlation between the two t-value matrices (R = 0.75, Fig. 5B), indicating similar effects of SSFO in modulating RSFC between ‘Active1’ and ‘Active2’ states. Notably, the RSFC matrix of the ‘Deactive’ scan showed no significant change compared with ‘Baseline’ (Fig. 5A, lower panel), suggesting the brain RSFC returns to baseline when the rdDG is no longer sensitized.

Figure 5.

SSFO activation in the rdDG altered brain-wide RSFC. A) ROI-wise RSFC comparisons between different conditions (lower triangles: unthresholded t values; upper triangles: thresholded t values, p<0.05, linear mixed model, FDR corrected). Left column shows the t-value matrices of 130 ROIs covering the whole brain, and the right column shows the t-value matrices of 42 ROIs exhibiting significant RSFC changes in the ‘Active1’ state relative to Baseline. B) Entry-to-entry correlation between the two t-value matrices of the ‘Active1-Baseline’ and ‘Active2-Baseline’ comparisons. Both matrices include 42 ROIs exhibiting significant RSFC changes in the ‘Active1’ state.

The 42 ROIs mainly belong to the hippocampal region (CA1, CA3 DG), striatum, mPFC (ACC, PL and IL), and some parts of the sensory-motor cortex, which is highly consistent with the fALFF results (Fig. 3). However, unlike more ipsilaterally biased results in fALFF data, both right and left hemispheric ROIs were involved in RSFC changes. Since our method for causal inference relies on the changes of RSFC, we next used the 42 ROIs for further causal inferences.

To obtain the causal information (direction and directness) of a connection, we calculated the partial correlations of the 42 ROIs. Since the partial correlation between two ROIs regresses out the influence of all other ROIs, the residual correlation was interpreted as arising from direct connectivity [44]. The partial correlation could introduce a pseudo correlation in the regression step (Fig. 6A), and this was corrected by Pearson’s correlation that any partial correlation coefficient will be assigned to zero if the corresponding Pearson’s correlation is not significantly different than zero. Fig. 6B displays the p-value thresholded (p < 0.05) binary matrices of Pearson’s correlation, ‘uncorrected’ and ‘corrected’ partial correlation in the baseline, from which we searched for colliders, chains, and confounders. Fig. 6C shows the scenario of a collider, in which one of the partial correlations is a pseudo correlation produced during the regression step. Since a collider already gives the direction, we can infer A → C and B → C from our baseline data.

In the scenario depicted in Fig. 6D, the three-node system is fully connected in Pearson’s correlation, while the partial correlation detects two direct connections. Thus, the three nodes could form a confounder or a chain. To further separate these two scenarios, we compared partial correlations between ‘Baseline’ and the ‘Active1’ or ‘Active2’ state (Fig. 6F, p < 0.05, paired two-sample t-tests, linear mixed-effect model). As depicted in Fig. 6E, if only one partial correlation changed, we inferred that the SSFO-induced perturbation acted through the middle node to its downstream node in a chain. If both partial correlations changed and their effective sizes were not significantly different, we inferred that the perturbation acted through the middle node to both of its downstream nodes in a confounder.

Given that the binary matrices of Pearson’s correlation, ‘uncorrected’ and ‘corrected’ partial correlation in the baseline relies on the p-value threshold applied, here for a given connection we used different p-value thresholds (0.05 – 0.005) to determine whether its Pearson’s and partial correlation coefficients were significantly different from 0. At the p < 0.05 level, we detected the directions of 40 connections when comparing partial correlations between ‘Baseline’ and ‘Active1’ (p < 0.05, linear mixed model, FDR corrected, Fig. 7A, left), and directions of 32 connections when comparing between ‘Baseline’ and ‘Active2’ (p < 0.05, linear mixed model, FDR corrected, Fig. 7B, left). At higher p-value thresholds, directions of smaller subsets of connections were determined (Fig. 7, middle & right).

Figure 7.

Causal inferences. A) Determine the directions of connections based on partial correlation changes between ‘Active1’ and ‘Baseline’. B) Determine the directions of connections based on partial correlation changes between ‘Active2’ and ‘Baseline’. Different p-value thresholds (left, p = 0.05; middle, p = 0.01; right, p = 0.005. Same for A & B) were used to determine the existence of Pearson’s correlation and partial correlation for a given connection.

To validate the causal relationships inferenced, we compared the directional information we obtained to anatomical tracing data [28], which is considered as the gold standard of measuring the directions of fiber projections. Out of the 42 ROIs, we analyzed (Fig. 5), directional connectivity in 23 ROIs was found in the database of rat cerebral cortex projections [28]. Comparison to tracing data yielded the correction rate of our causal inferences in the range 84–100% depending on the p-value threshold applied, while a random guess would generate an accuracy rate of 36.8% (Fig. 8). Although the number of causal relationships we could infer was lower at a more stringent p-value threshold as shown in Fig. 7, the accuracy rate became higher (Fig. 8, middle & right), suggesting that more reliable connectivity measurement increases the robustness of causal inference. Taken together, these results indicate that SSFO-based opto-fMRI provides a valuable tool to retrieve causal relationships in brain-wide neural connections at the resting state.

Figure 8.

Comparing inferred causal relationships with anatomical tracing data. A) Causal relationships inferenced based on partial correlation changes between ‘Active1’ and ‘Baseline’. Red dashed lines in accuracy bars are the correction rate from a random guess. B) Causal relationship inferenced based on partial correlation changes between ‘Active2’ and ‘Baseline’. The same p-value thresholds were used as Fig. 7 (left, p = 0.05; middle, p = 0.01; right, p = 0.005. Same for A & B).

Discussion

By combining optogenetics and fMRI, opto-fMRI provides a tool to selectively manipulate a specific brain region with tissue specificity and record the corresponding changes in brain-wide activity/connectivity [45], making it possible to study causality in connections across distributed brain regions. The optogenetic stimulation induced by SSFO enables a circumscribed brain region to enter a steady state of excitation. This phenomenon could also be found in the normal brain, for example, elevated excitability in the cortex was found to be related to attention and sensorimotor gating [46]. In the present study, by utilizing SSFO-based opto-fMRI, we looked to infer causality in the spontaneous firing neurons, which better represents the resting-state brain than task-based experiments. We have shown that SSFO activation in the rdDG excitatory neurons increased local spiking rate. We also showed its impact on brain-wide activity and RSFC. Importantly, we showed that the modulated RSFC provides information that allows us to further analyze the causal interactions in the perturbed brain circuits.

Particularly, CaMKIIα:SSFO was used to specifically increase the excitability of granule cells in the rdDG. The dentate gyrus receives input primarily from the entorhinal cortex and controls the information flow through the hippocampus [26]. When the granule cells in the rdDG were more excited, we found increased brain activity in many ipsilateral brain regions, including the subregions within the hippocampus (CA3 CA1, and subiculum), as well as striatum, mPFC, and some cortical regions (Fig. 3). These results are consistent with the anatomical findings that DG projects to CA3, which then projects to CA1 [47]. Being the major output of the hippocampus, CA1 innervates a lot of brain regions. Anteriorly, CA1 strongly innervates the mPFC through fornix [48]. CA1 also sends projections to many cortical regions including visual, auditory, somatosensory, orbital and insular cortices, as well as olfactory and amygdala areas [49]. During the SSFO ‘Active’ conditions, the brain region with the most apparent change other than the hippocampus was mPFC, which exhibited significantly increased activity (Fig. 3) and RSFC with the excited region (Fig. 4). Previous studies have shown that the hippocampus-mPFC interaction plays a key role in learning and memory and its dysregulation has been associated with many brain disorders, such as schizophrenia, depression and PTSD [19,50,51]. Spellman et al. [52] found that the hippocampus to PFC projection is essential for memory encoding. The projection from the PFC to the dorsal hippocampus controls top-down memory retrieval [27]. Although the hippocampus receives a contralateral signal through the commissural pathway, our data show the activity change in the contralateral side of the brain is less pronounced, which is consistent with the finding that hippocampus projections are predominantly ipsilateral [53].

The elevated excitability of the rdDG was found to modulate the RSFC of a group of connections, involving regions in the hippocampus, mPFC, thalamus, striatum, and part of the cortex. Among these selected connections, we were able to further identify causal relationships by integrating Pearson’s correlation and partial correlation across different conditions. Importantly, unlike previous causal studies using ChR2 to stimulate brain regions in task-like conditions [12,13], SSFO-induced elevated excitability does not overdrive the normal neuronal function, which satisfies our goal to find the causal relationships in the resting state brain. Further, the current study provides a way to infer causal relationships within brain networks, as illustrated in Figs. 7&8. Since those relationships have both direction and directness, we think the causal relationship is best represented by direct axonal projection in anatomy. Therefore, we used the cerebral cortex projection database [28] to verify our results, and we found 84–100% of our findings could be supported by anatomical data.

The current method cannot cover all scenarios in three nodes. For example, the current model cannot resolve cyclic connections in three nodes. In addition, the current model also has difficulty in resolving bidirected connections, which are common in real brain anatomy. Also, the current method cannot remove all the false connections. In a model where node A and node B have a confounder C, in the meantime, they also collide to another node D. Pearson’s correlation of AB will be non-zero because of the confounder C, while partial correlation will also give a non-zero result because of the collider D. In this scenario, the current method will provide a false connection between node A and node B even if they are not directly connected. The current method depends on the discovery of the three simple scenarios (i.e., confounder, chain, and collider), so it cannot resolve more complicated scenarios where all three connections exist between the three nodes. The major goal of the present study is to examine the feasibility of using SSFO-based opto-fMRI to extract causal relationships at resting state by testing its sensitivity and reliability. A more sophisticated model that can extract more causal information warrants more modeling work in future studies.

Conclusion

The current study combines SSFO-based optogenetics with fMRI to investigate the relationship between local region activity and RSFC modulation in a resting-state rodent model. Data obtained in this study indicate that SSFO-based opto-fMRI provides a new way to analyze directional information in functional connectivity at the resting state.