Mapping brain-wide excitatory projectome of primate prefrontal cortex at submicron resolution: relevance to diffusion tractography

Resolving trajectories of axonal pathways in the primate prefrontal cortex remains crucial to gain insights into higher-order processes of cognition and emotion, which requires a comprehensive map of axonal projections linking demarcated subdivisions of prefrontal cortex and the rest of brain. Here we report a mesoscale excitatory projectome issued from the ventrolateral prefrontal cortex (vlPFC) to the entire macaque brain by using viral-based genetic axonal tracing in tandem with high-throughput serial two-photon tomography, which demonstrated prominent monosynaptic projections to other prefrontal areas, temporal, limbic and subcortical areas, relatively weak projections to parietal and insular cortices but no projections directly to the occipital lobe. In a common 3D space, we quantitatively validated an atlas of diffusion tractography-derived vlPFC connections with correlative enhanced green fluorescent protein-labelled axonal tracing, and observed generally good agreement except a major difference in the posterior projections of inferior fronto-occipital fasciculus. These findings raise an intriguing question as to how neural information passes along long-range association fiber bundles in macaque brains, and call for the caution of using diffusion tractography to map the wiring diagram of brain circuits.

Higher-order processes of cognition and emotion regulation that depend on the prefrontal 2 cortex are all based on multiple, long-range connections between neurons 1, 2, 3 . Axons 3 connecting local and distant neurons form a fundamental skeleton of the brain circuitry, 4 which is of paramount importance to fathom the organization of in-/output pathways that 5 enable those vital functions 4, 5 . Given the complexity and heterogeneity of the primate 6 prefrontal cortex 2 , understanding the working mechanisms of the prefrontal cortex requires 7 a comprehensive map of axonal projections linking its demarcated subdivisions and the 8 rest of brain. A subdivision of the prefrontal cortex -the ventrolateral section (vlPFC), 9 which mainly spans Brodmann's Areas 44, 45a/b, 46v/f and 12r/l 6 , is central to a variety 10 of functions including language, objective memory and decision making 7, 8 . Emerging 11 evidence further demonstrates abnormalities of vlPFC in tight association with deficits in 12 cognitive flexibility 1, 9, 10 , suggesting that an elaborate delineation of its hard wiring would 13 shed light on the underlying neuropathology of psychiatric disorders 11 . 14 Such neuroanatomical inter-areal connectivity has been probed using invasive bulk 15 injections of tracers and noninvasive imaging methods with millimeter-scale spatial 16 resolution 12, 13, 14 . Histological neural tracing has been historically utilized for 17 circuit/pathway mapping and continues to be the most reliable way of survey all myelinated 18 axons in mammalian brains 12, 20, 21 , which has also been used as a gold standard to validate 19 other modalities like diffusion tractography 18,22,23,24,25,26 . Diffusion tractography, which 20 has been developed in 1990s to estimate the tissue microstructure by means of spatial 21 encoding of water molecule movements 15 , represents the only methodology capable of 22 inferring the ensemble of anatomical connections in the living animal or human brain 16,17 . 23 But this technique is an indirect observation with limited resolution and accuracy, and its 24 reliability of false negative and false positive findings remains to be fully validated in a 3D 25 space 18,19 . Notably, some classic tract tracing methods are not sensitive to specific 26 neuronal types or axonal trajectories. They do not report the traveling course in a 3D space 27 through which the axons travel for a remarkably long distance (i.e., over centimeter length). 28 The pursuit of long-range axonal fiber tracing across the entire monkey brain has become 29 feasible thanks to rapid advance in viral and genetic tools in the primate species, tissue 30 labeling, large-scale microscopy and computational image analysis 27,28,29,30 . Moreover, 31 viral-based techniques for targeting specific neuronal types in macaque brain have 1 achieved remarkable success 27, 31 , which may furnish the requisite biological detail 2 including excitatory and inhibitory in-/output to enrich structural network reconstructions 3 for improved prediction of brain function 32 . However, it remains unclear thus far what type 4 of viral vector is suitable for long-range axonal fiber tracing 12,20 . 5 In the present study, we aim to establish a comprehensive brain-wide excitatory 6 projectome of the vlPFC in macaque monkeys using viral-based genetic tracing in tandem 7 with serial two-photon (STP) tomography, a technique that has successfully achieved high- 8 throughput fluorescence imaging of the entire mouse brain by integrating two-photon 9 microscopy and tissue sectioning 33 . In addition, in a common 3D space reconstructed with 10 STP tomography, we intended to make a direct comparison of this mesoscale projectome 11 to that derived from ultra-high field diffusion tractography (Fig. 1). Note that cross-12 comparison of the fiber details generated by two modalities with spatial scale differences 13 in order of magnitude is technically demanding as many cellular structures or fiber 14 pathways of biological interest are rather small relative to the voxel size of most diffusion 15 MRI data 17 . One of challenging undertakings is to image long-range axonal fibers of many 16 neurons with sufficiently high resolution to enable tracking axonal trajectories across the 17 entire brain 33, 34 , which has stirred some debates such as right-angle fiber crossings 16, 24 and 18 the existence of the inferior fronto-occipital fasciculus in the primate brain 35 . by STP tomography throughout the brain, which enables a close-up view and quantitative analysis of 6 any region-of-interest (middle panel). A supervised machine learning approach was used for 7 segmentation of GFP-positive signal and removal of autofluorescence in STP data. The serial segmented 8 GFP images were down-sampled to compute the total signal intensity for each 200 μm × 200 μm grid 9 by summing the number of signal-positive pixels in that grid and to generate the axonal density map 10 (bottom panel). (B) An MRI atlas of cynomolgus macaques was used to construct a common 3D space. 11 (C) Ex-vivo dMRI of macaque brain were acquired with using an 11.7T MRI scanner, illustrated as 12 representative B0 (left) and direction-encoded color FA maps (right). Using the injection site identified 13 from the STP data as seed regions, the target fiber tracts can be derived from diffusion tractography. (D) 14 Integration of STP and dMRI data was implemented in a common 3D space, which allows quantitative 15 analyses including whole-brain analysis of axonal projectome (left), comparison of vlPFC projectome 16 by STP and dMRI (middle), and cross-validation of fiber tracking in both STP and dMRI (right).

20
Determination of viral vectors for long-range anterograde tracing 21 We tested whether VSV, lentivirus, and AAV vectors with demonstrated success in rodents 1 worked in the macaque brain and which vector was best suitable for long-range axonal 2 fiber tracing. Five days after infection with VSV-△G, the neuronal cell bodies in the 3 cerebral cortex ( Fig. 2A and B) and thalamus (Fig. S1A) were clearly labelled with GFP, 4 although only proximal neurites were labeled with no long-range axonal fibers detected 5 (Fig. S2A). When the infection time was extended to about a month, we observed 6 widespread axon loss and neuronal cell death (Fig. S1B-G). The infected neurons 7 underwent morphological changes such as membrane blebbing ( Fig. S1B and C), a key  Lenti-Ubic-GFP exhibited stable expression in the cell soma even after 9 months (Fig.

19
To acquire a detailed account of the brain-wide vlPFC projectome, we analyzed its 20 connectivity profile with other 173 parcellated regions in the monkey brain atlas using the 21 STP tomography data ( Fig. 3 and 4). The GFP-labelled projecting axons largely 22 encompassed the anterior part of the brain including the frontal lobe, temporal lobe, limbic 23 lobe, insular, and some subcortical regions, but no labeled axons were found in the occipital from vlPFC relative to other OFC subregions (Fig. 4E). Laterally, axonal projections were 28 found in the FEF including 8Av (Fig. S5A, B and C) and 8Ad ( Fig. S5A and D). Dorsally, 29 there were dense axonal fibers in the dorsal prefrontal cortex, including area 8Bd (Fig. S5E,  The axons with the premotor cortex exhibited a gradient pattern with the largest axon 3 distribution along the anterior part (Fig. S5L). In addition, axons were noted in the 4 precentral opercular area (PrCO) and medial prefrontal area (mainly in 10mr) (Fig. 3C).

5
Interestingly, the projections anchored in the prefrontal cortex of these axonal fibers formed 6 isolated clusters ( Fig. 4F and H, indicated by arrows). The z-axis extent of axonal clusters 7 was ranging from 1.2 mm to 3.8 mm (2.24 ± 0.80 mm) (Fig. S6). 8 Beyond the frontal lobe, rich connections were observed in the temporal lobe (Fig. 3),   Comparison of vlPFC axonal projections by dMRI and STP 11 We further introduced a quantitative comparison of vlPFC connectivity profile obtained by 1 dMRI-based tractography and STP data. Typical T2-weighted and dMRI images of the 2 macaque brain acquired from an ultra-high field MRI scanner were shown in Fig. 5A-D.

3
To carry out a proof-of-principle investigation, we focused on the vlPFC-CC-contralateral 4 tract that was reconstructed in 3D space by using STP and dMRI data, respectively (  Szymkiewicz-Simpson overlap coefficients between 2D coronal brain slices of the dMRI-derived IFOF 8 tract and vlPFC projections were plotted along the anterior-posterior axis of the macaque brain. Four 9 cross-sectional slices (D-G) along the IFOF tracts were arbitrarily chosen to demonstrate the spatial 1 correspondence between the diffusion tractography and axonal tracing of STP images. (D-G) The 2 detected GFP signals (green) of vlPFC projectome and the IFOF tracts (red) obtained by diffusion 3 tractography were overlaid on anatomical MRI images, with a magnified view of the box area. Evidently 4 there was no fluorescent signal detected in the superior temporal area where the dMRI-derived IFOF 5 tract passes through (G). 6 7 Discussion 8 Brain-wide excitatory projectome of vlPFC in macaques 9 We customized STP tomography for whole-brain imaging of the macaque monkey at 10 submicron resolution and accomplished brain-wide 3D reconstruction of axonal 11 connectome, thanks to prominent characteristics of STP tomography including free of 12 tissue distortions, no need for section-to-section alignment, and high-resolution image sets 13 readily warped in 3D space 36 . Importantly, we coupled STP tomography with genetic  Extrinsic connections beyond the PFC, vlPFC is connected mainly to the dysgranular insula, 27 frontal operculum, somatosensory-related areas in the parietal operculum and inferior 28 parietal cortex, visual-related areas in the inferior temporal cortex, and anterior cingulate 29 areas 6, 37 . We found the excitatory projection of vlPFC to the rest of brain was compatible 30 with previous reports using chemical tracers 3, 37, 38, 39 . Furthermore, we compared the 31 current vlPFC projectome data with the well-known macaque connectivity database 32 CoCoMac 40 , which includes the results of several hundred published axonal tract-tracing 1 studies in the macaque monkey brain 41 . Essentially the vlPFC connectivity profile shown 2 here was markedly similar to that of CoCoMac, except that vlPFC projections to PFCol, 3 PFCdm, PFCdl, PFCoi, PFCm, PMCvl, amygdala, and SII have not been reported in 4 CoCoMac database or reported merely with unspecified strength. In addition, we compared 5 vlPFC projections with one recent report 37 , showing that the brain regions projected from 6 area 45 were clearly observed in the present vlPFC projection data. Note that we used the 7 projection volume index instead of the fluorescence intensity, which has been demonstrated 8 reliably to quantify axonal connectivity strength 4 . Although the passing fiber and terminal 9 were not readily distinguishable, the results of terminal labelling that compared 10 synaptophysin-EGFP-expressing AAV with the cytoplasmic EGFP AAV have shown high 11 correspondence in target areas 4 . 12 13 AAV2/9 is suitable for long-range axonal tracing in the macaque brain 14 Methods for tissue labeling have been continuously evolving from silver impregnation of 15 degenerating fibers to ex-vivo visualization of axonally-transported tracers injected at 16 single brain nuclei, and finally to an integrated style which coupled high-resolution whole-17 brain imaging technologies with viral and genetic tracers 28 . Among four viral vectors tested 18 here, we found that AAV2/9 demonstrated the highest efficiency of long-range axonal 19 tracing in the macaque brain. VSV was initially utilized as a transsynaptic tracer in a 20 previous study since VSV encodes five genes, including G protein which promotes 21 anterograde transsynaptic spread among neurons 42 . In our study, we used VSV with G 22 deletions to trace axonal projection without trans-synaptic labeling, which enabled robust 23 gene expression at remarkably higher level relative to other vectors in a very short time 24 (less than a week). But we found that a shorter expression time of VSV-△G was 25 insufficient to label axons traveling long distance whereas a longer expression time of 26 VSV-△G caused cell death, consistent with a prior finding that VSV-G failed to label 27 transsynaptic cells at distant areas 43 . The advantage of lentivirus, which is derived from 28 human immunodeficiency virus type 1 (HIV-1) 44 , is that it has a large genetic capacity of 29 approximately 10 Kb which allows for the expression of multiple gene and usage of more 30 than one promoter or regulatory elements. And we found GFP expression induced by 31 lentivirus remarkably stable after 9 months in macaque monkeys, even though the labeled 1 level was mild 45 and the labeled scope was limited.

2
As an effective carrier for gene delivery into the brain, AAV has a number of 3 established advantages including minimal toxicity, weak host immune response, stable 4 gene expression in neurons with extraordinarily high transfection efficiency (titers up to 5 10 12 -10 13 genome copies per mL) 30 . One major drawback of AAV vectors is the limited 6 packaging capacity. AAVs usually deliver gene cassettes of approximately 4.8 Kb (i.e., one 7 or two small genes) 28 , which has motivated us in pursuit of biocompatible nano-based 8 carriers 46 . It is well known that different AAV serotypes have their own sequences in the 9 inverted terminal repeats such that they have distinct transfection bases towards various 10 cell types in the brain. The recombinant virus we used was AAV2/9 which contains the 11 inverted terminal repeats from AAV serotype 2 and the capsid proteins from AAV serotype  HRP tracer in a monkey brain, but suffered certain limitations for regions at remote 30 locations from seeds 54 . Moreover, the structural connectivity analyses based on the 31 histological dataset provided varying correlative evidence between these two 1 measurements (like r = 0.21 49 using the CoCoMac tracer data 40 and r = 0.59 18 using the 2 tracer connectivity matrix from 55 ). Note that such structural connectivity analysis does not 3 describe a 3D correspondence of the axonal fiber trajectory, but an "end-to-end" match.

4
STP tomography effectively transformed a series of histological slice images into a 3D 5 space with which dMRI-derived tracts were co-registered, thus enabling a direct, 6 quantitative comparison of the high-throughput data from these two modalities. This is 7 technically challenging due to a giant difference in scale between the axonal fibers and 8 image resolution of dMRI 13 . We have taken meticulous steps to maximize the signal-to-9 noise like using Gd-DTPA as an enhanced contrast agent 56 and to minimize the image 10 artifacts in an ultrahigh field scanner for achieving a reasonably high spatial resolution. We interesting questions about the long-range brain organization and the functional role of 10 superior temporal area in primates which definitely merits future examination.

11
In summary, we present a detailed excitatory connectivity projection map from vlPFC 12 to the entire macaque brain, and demonstrate a broadly applicable roadmap of integrating 13 3D STP tomography labeled with antero-/retro-grade tracer and diffusion tractography for 14 the mesoscopic mapping of brain circuits in the primate species.   Fluorescence signals of AAV labeled areas were detected and recorded using a customized 4 STP tomography (Fig. S8, Supplementary materials). High x-y resolution (0.95 μm/pixel) 5 serial 2D images were acquired at a z-interval of 200 μm across the entire macaque brain, 6 as resulted in a continuous ~1 month scanning and ~5 TB STP tomography data for one 7 monkey brain (Fig. S9). Once finished scanning, all sections were retrieved from the 8 cutting bath and stored in cryo protection solution (containing 30% glycol, 30% sucrose in 9 PBS) at -20℃ for further histological examination.

19
We therefore implemented and compared the following three methods for background machine learning for autofluorescence exclusion (Fig. S10J).

25
The first method involved staining the brain tissue with anti-GFP antibody and Alexa

26
Fluor 405 conjugated secondary antibody to transform the GFP signal from green channel 27 to blue channel. Unlike the green and red channels, the transferred blue channel (Fig. S10M) 28 did not contain high intensity autofluorescence puncta. Although this post-hoc thick-29 section immunofluorescent method successfully reduced autofluorescence, it was 30 incompatible with the block face imaging method. The second one was to subtract the 31 normalized red channel from the green channel using the broad emission spectrum 1 characteristic of autofluorescence puncta, which was able to remove high intensity 2 background signal (Fig. S10F). The third was based on a supervised machine learning 3 plugin for ImageJ, trainable WEKA segmentation 69 , which classifies and binarizes GFP 4 and autofluorescence background signal for background exclusion (Fig. S10J). Both 5 subtraction and machine learning methods were used for better visualization of 6 fluorescence images when necessary, whereas only the supervised machine learning 7 approach was used for quantitative analysis of STP data 68 .

24
Red channel volume was used to perform registration to the monkey brain template, as red 25 channel images contain visible anatomical information of brain structures 71 . The brain 26 template of cynomolgus macaque was adopted from an MRI-based atlas generated from 27 162 cynomolgus monkeys 66 . We warped the red channel volume to the template space by 28 using a symmetric normalization (SyN) algorithm in ANTs (Fig. S11). The cortical label 29 was adopted from the D99 parcellation map 72 , and subcortical label was adopted from 30 INIA19 parcellation map 73 . Also the segmented GFP volume and injection site volume 31 were co-registered onto the same template. Density of GFP signal and total GFP volume in 1 each parcellated brain region were used to represent the axonal connectivity strength.
2 Percent of total projection was defined by the GFP-positive pixel count within each 3 parcellated brain region (or brain lobe) normalized to the total of all GFP-positive pixels.

4
Additionally, the percent innervation density was calculated as the proportion of density of 5 GFP pixel counts covering the maximal density of GFP pixel counts of the brain. To create 6 plots that display the data along the anterior-posterior axis (e.g. % density innervation), the 7 location of ear bar zero was used as the origin. The percent innervation density of each 8 cortical region innervated by vlPFC was rendered onto a brain surface.

10
Ex-vivo MRI scanning and data preprocessing 11 We collected dMRI data using an 11.7 T horizontal MRI system (Bruker Biospec 117/16

22
The scanning parameters were set: TR/TE = 82/22.19 ms, FOV = 64×54 mm, acquisition 23 matrix = 128×108, slice thickness = 0.5 mm, and averages = 3. In addition, whole brain 24 T1-weighted and T2-weighted structural images were obtained using 3D FLASH and 2D  area. Note that the axonal density map was also filtered by setting an intensity threshold of 28 10 3.2 to minimize false positives due to segmentation artefacts 4 . After co-registered the 29 probabilistic maps and the axonal density maps onto the same template, both Dice 30 coefficients and pixel-wise Pearson coefficients were calculated to quantitatively assess the 1 spatial overlap 78, 79 .

2
As described recently 63 , the inferior fronto-occipital fasciculus was reconstructed 3 using streamline-based probabilistic tractography. We ran this probabilistic tractography 4 tool in MRTrix3 (www.mrtrix.org) via bootstrapping 80 . Streamlines were seeded over the 5 whole brain area that encapsulated the tract of interest. Two inclusion masks were used to 6 define two regions that each tract must pass through, and only streamlines that pass through 7 both regions are retained. One exclusion mask was used to restrict tracking to the 8 contralateral hemisphere of the brain. The inclusion and exclusion masks were drawn 9 manually as described previously 63  were then detected with using ImageJ and FSL software in both 2D and 3D space (Fig. 7B).

22
The Szymkiewicz-Simpson overlap coefficient was adopted to quantify the spatial 23 relationship between the IFOF tract and vlPFC projectome, which was defined as the size 24 of the union of them over the size of the smaller set: The Szymkiewicz-Simpson overlap coefficient ranges from 0 (no overlap) to 1 (if the 27 IFOF tract is found in its entirety in vlPFC projectome). can copy any or all of it to their own computer. We would like to thank Jinqiang Peng and Jie Xu for their assistance to data acquisition,  Erlangen, Germany) under general anesthesia. A detailed description of in-vivo MRI 4 scanning procedure has been described in our previous studies 1, 2, 3, 4, 5, 6 and briefly 5 summarized here. Anesthesia was induced by intramuscular injection of ketamine (10 mg 6 per kg). Deep anesthesia was maintained by isoflurane (1.5-3%) and vital physiological 7 signals were continuously monitored during MRI scanning. Anatomical scans were 8 acquired with an MPRAGE sequence using the following parameters: TR = 2300 ms, TE 9 = 2.8 ms, TI = 1100 ms, spatial resolution 0.5 mm isotropic. The target regions were 10 localized in each animal by warping the 3D digital atlas of Saleem and Logothetis 7 to the 11 individual T1 image using a symmetric normalization (SyN) algorithm. The location of the 12 vlPFC was then calculated with regard to the stereotaxic space. 13 All procedures for virus injection were performed in strict aseptic conditions. The 14 head of the animal was fixed in a stereotaxic apparatus, within the same coordinate space 15 as the MRI images. The target area was then labelled and an incision was made to expose 16 the skull. A burr hole with a 2 mm radius was drilled above the target according to the 17 calculated coordinates, and the dura was carefully incised to expose the cortical surface. 18 The viral vector was delivered into the cortex using a 33-gauge Hamilton syringe controlled 19 by an UltraMicroPump and a micro4 controller (WPI). The injection speed started with 20 200 nl/min and was increased to 400 nl/min; total injection volume was 10-20 μl. After 21 injection, the needle was retained for at least 15 minutes and drawn back at a rate of ~1 22 mm/min. The burr hole was then filled with bone wax and the skin was sutured. 23 Cephalosporin was given for three consecutive days after surgery (25 mg/kg/day, i.m., once 24 a day). 25 26

Cryo-sectioning 27
For virus testing, serial sections were cut on a freezing microtome. The fixed brain was 28 first cut into a block, then equilibrated sequentially in 15% and 30% sucrose in PBS until 29 it sank to the bottom of the container. A cryostat microtome (Leica CM1950) was used to 30 serially slice the brain into 50 μm sections. Brain slices were preserved in a cryoprotectant 31 solution (containing 30% ethanediol, 30% sucrose in PBS solution, pH = 7.2) for further 1 immunofluorescence staining and imaging. 2 3 Serial two-photon tomography 4 To image the monkey brain, we customized the STP tomography system which was 5 integrated a two-photon microscope (Bruker) with a vibratome (WPI) (Fig. S8), computer 6 controlled and fully automated. The XY stage covered a 50*60 mm 2 area, and the 3D 7 scanning of Z-volume stacks was achieved with using a stepper motor (Thorlabs) that 8 traveled over 70 mm. The fixed brain that was embedded with 4% agarose was scanned in 9 a 3T MRI to obtain ex-vivo T1 images. Using these T1 images as reference, the active 10 imaged region of each section was determined during STP tomography for improved 11 imaging efficiency. The embedded brain was then held via a magnetic adaptor to a stepper 12 motor and immersed in a cutting bath filled with PBS containing 0.1% sodium azide. The 13 vibratome blade was aligned in parallel with the leading edge of the specimen block. Brain 14 images were captured from the anterior PFC to posterior V1 in the coronal plane. 15 Fluorescence signals for the green channel (excitation wavelength light in 920 nm) and red 16 channel (excitation wavelength light in 1045 nm) were acquired at 30 μm below the cutting 17 surface through a Nikon 16x Water objective (N.A. = 0.8). 18 During serial scanning, the STP system was fully automated: each optical section was 19 imaged as a mosaic of fields of view on the block surface as the xy stage moved the brain 20 under the objective; once an entire section was imaged, the xy stage moved the brain to the 21 vibratome and cut off a 200 μm section from the top of the sample. The remaining specimen 22 was then moved back under the objective for imaging the next neighboring plane. Optical 23 and mechanical sectioning were repeated until the complete brain data was collected. 24 Hence fluorescent images of the whole monkey brain were continuously acquired (Fig. S9). 25 26

Histological staining 27
To perform immunofluorescence procedure, brain slices were incubated in blocking 28 solution containing 5% BSA and 0.3% Triton X-100 in PBS at room temperature for 2 hr 29 and then overnight with primary antibodies in PBS containing 3% BSA and 0.3% Triton 30 X-100 at 4 ℃. Slices were rinsed in PBS followed by Alexa Fluor-conjugated secondary 31 antibodies at room temperature for 3 hrs, and DAPI (Cell signaling Cat# 4083s) for 30 1 mins at room temperature. The following primary antibodies were used: CaMKIIa (1:200

Ex-vivo diffusion MRI 19
After a fixation period of ~ 30 days, the whole brain specimen was immersed in a 1:100 20 dilution of a 1 mmol/mL gadolinium MR contrast agent (Magnevist®, Bayer Pharma AG, 21 Germany) mixed with phosphate buffered saline (PBS) solution for 14 days. Before MRI 22 scanning, the specimen was washed and drained of water from the surface, then positioned 23 into a customized container which was 3D printed for perfect accommodation of the brain 24 sample. Thus the brain was held steadily during MRI scanning. And the container was filled 25 with FOMBLIN® perfluoropolyether (Solvay Solexis Inc. Thorofare, NJ, USA) for 26 susceptibility matching and improved magnetic field homogeneity. The specimen was 27 degassed with a vacuum pump for 24h under 0.1 atmosphere pressure to remove all air 28 bubbles in the sample at 20 ℃ (magnet room temperature). The ex-vivo macaque brain 29 was scanned on a 11.7 T animal MRI system (Bruker Biospec 117/16 USR, Ettlingen, 30 Germany), equipped with a 72 mm volume resonator and an actively shielded, high 31 performance BGA-S series gradient system (gradient strength:740 mT/m, slew rate: 1 6660T/m/s). dMRI images were acquired using a 3D diffusion-weighted spin echo pulse 2 sequence with single-line read-out, TR/TE = 82/22.19 ms, FOV = 64×54 mm, matrix = 3 128×108, slice thickness = 0.5 mm and averages = 3, which included 60 diffusion 4 directions with b = 4000 s/mm 2 (Δ/δ = 15/2.8 ms, maximum b value = 4234.97, gradient 5 amplitude = 97.19 mT/m) and five non-diffusion encoding (b = 0 s/mm 2 ) directions. For 6 the ex-vivo diffusion MRI data acquisition, the b-value was recommended to set at 4000 7 s/mm 2 8, 9 . T2 weighted images were acquired using a 2D Turbo RARE sequence with 8 TR/TE = 8353.42/28.8 ms, flip angle = 87°, matrix = 450×450, FOV = 54×45 mm, slice 9 thickness = 0.5mm, and averages = 6. T1 weighted images were acquired using 3D FLASH 10 sequence with TR/TE = 40/5.5 ms, flip angle = 15°, matrix = 290×225, FOV =58×45 mm, 11 slice thickness = 0.2 mm, and averages = 4. The total scan time was approximately 36 12 hours. 13 Visual inspection of MRI data was first performed to ensure that there were no obvious 14 image artefacts and geometric distortions. Then we calculated the signal-to-noise ratio 15 (SNR) for typical diffusion images. As diffusion images were acquired by spin warp 16 imaging (image reconstruction by a 3D Fourier transform) with a volume quadrature coil, 17 the SNRs were calculated using the "two-region" approach 10, 11 . Specifically, for each 18 gradient encoding direction, the deep white matter (WM) were extracted in subject-native 19 diffusion space to represent the signal 12 ; a region positioned in the no signal area at the 20 corner of the image was used to represent the noise. As a rule of thumb, the SNR of b = 0 21 s/mm 2 images should be minimally larger than 20 for obtaining relatively unbiased 22 measures of parameters such as FA 13 . Typical SNRs of diffusion images with b = 0 and b 23 = 4000 in the present study were 48.34 ± 8.50 and 23.13 ± 2.05, respectively. It allowed a 24 reliable seed-based 3D reconstruction for diffusion tractography, as illustrated in Fig. 5. 25

Construction of inferior fronto-occipital fasciculus 26
The streamline-based probabilistic tractography strategy was used to generate the IFOF 27 tracts in 3D 14 . The fiber orientation distribution function (FOD) was estimated with 28 MRtrix3 (www.mrtrix.org) 15 using the tournier algorithm for single-tissue Constrained 29 spherical deconvolution 16 . For fiber tracking, we then used tckgen with the Tensor_Prob 30 tracking algorithm in MRtrix3 17 . Within each image voxel, a residual bootstrap was 31 performed to obtain a unique realisation of the dMRI data in that voxel for each streamline. 1 These data are then sampled via trilinear interpolation at each streamline step, the diffusion 2 tensor model is fitted, and the streamline follows the orientation of the principal 3 eigenvector of that tensor. The following additional tckgen settings and inputs were used: 4 step size of 0.25 mm, max. angle between successive steps = 45°, max. length = 150 mm, 5 min. length value set the min. length 10 mm, cutoff FA value = 0.1, b-vectors and b-values 6 from the diffusion-weighted gradient scheme in the FSL format, b-value scaling 7 mode = true, maximum number of fibers = 10,000, and unidirectional tracking.  The stepping distance along z-axis is 200 μm. The images acquired through the red channel are used as 5 background. The injection site (vlPFC) and major projection targets can be readily observed in this 6 'montage' view.