{"paper_id":"8034e0d4-d196-4f29-ac48-c2d76e8bb850","body_text":"Cortical myelination networks reflect neuronal gene \nexpression and track adolescent age in marmosets \n \nE D Hutchings1, *, S J Sawiak2, R A I Bethlehem3, A C Roberts2, E T Bullmore1, 4 \n \n1Department of Psychiatry, University of Cambridge, Cambridge, UK \n2Department of Physiology, Development and Neuroscience, University of Cambridge, Downing Street, \nCambridge, UK \n3Department of Psychology, University of Cambridge, Cambridge, UK \n4School of Academic Psychiatry, Institute of Psychiatry, Psychology & Neuroscience, King’s College \nLondon, London, UK \n \n* Correspondence to E D Hutchings. Email: eh636@cam.ac.uk  \n \nAbstract \n \nStructural similarity provides a powerful framework for measuring coordinated macro- \nand microstructural variation across the cortex of a single brain. Similarity networks \nderived from myelin-sensitive MRI sequences undergo marked reorganisation during \nadolescence, linked to psychosocial outcomes in humans and rodents. However, the \ncellular mechanisms of myelination similarity and its development in non -human \nprimate cortex remain unexplored. Here, we used myelin -sensitive T1w/T2w ratio \nimages from a cross-sectional sample of 446 common marmosets (aged 0.62 to 12.75 \nyears) to estimate cortical similarity networks in individual animals. Cortical areas with \nsimilar myeloarchitecture showed highly similar patterns of gene co -expression in \nglutamatergic neurons and PV+ and VIP+ interneurons, reflecting the activity \ndependence of myelination. A reliable age-based signal exists within network features, \nwith coordinated developmental trajectories observed across the cortical hierarchy \nfrom primary to transmodal asso ciation cortices - a pattern that mirrors findings in \nhuman cortex. Taken together, marmosets demonstrate phylogenetically conserved \npatterns of myelination network development, potentially underpinned by key neuronal \ncell types that shape the functional specialisation of cortical areas. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n2 \nIntroduction \n \nNetwork science has significantly advanced our understanding of brain structure \n(Bullmore and Sporns 2009; Bassett and Sporns 2017). Magnetic resonance imaging \n(MRI) has been a key driver of this progress because it has enabled the collection of \nwhole brain images at the millimetre scale in vivo (Bullmore and Sporns 2009 ). \nAnalysis of MRI images allows quantification of brain structure both at the macro -\nscale, through metrics such as cortical thickness, surface area, and mean curvature; \nand at the micro-scale, below the minimal resolution of single voxels, using sequences \nsuch as diffusion imaging and magnetisation transfer (Lerch et al. 2017) . Recent \ntechnical advances in structural similarity have permitted the integration of these MRI-\nderived structural features into a single network representation, with edges between \nnodes defined by the statistical relationship between the MRI features mea sured in \npairs of regional nodes (J. Wang and He 2024; Sebenius , Dorfschmidt et al. 2025) . \nSimilarity networks therefore represent spatially patterned structural variation across \nthe cortex (the architectome), rather than directly measuring connectivity (the \nconnectome), as is the case in diffusion tensor imaging (DTI) tractography. However, \nareas that are structurally similar are more likely to be axonally connected due to the \nnetwork formation principle of homophilic attachment, or “like attracts like” (Sebenius, \nDorfschmidt et al. 2025) . Accordingly, the architectome represented by MRI -derived \nstructural similarity networks has been correlated empirically with the connectome \nrepresented by axonal tract tracing data in the macaque (Sebenius et al. 2023; Seidlitz \net al. 2018) and rat (Smith et al. 2024) cortex. \n \nA recently developed method for estimating structural similarity networks from MRI \ndata is Morphometric Inverse Divergence (MIND), which uses Kullback -Leibler \ndivergence to quantify how similar a pair of brain regions is in terms of their voxel - or \nvertex-wise distributions of one or more structural features (Sebenius et al. 2023) . \nMIND and adjacent methods are technically robust and reliable (H. Wang et al. 2016; \nSebenius et al. 2023) , heritable (Sebenius et al. 2023) , associated with similarity in \ngene expression (Sebenius et al. 2023; Seidlitz et al. 2018), developmentally sensitive \n(Sebenius et al. 2023; J. Wang and He 2024; Smith et al. 2024) , and altered in \npsychiatric disorders (J. Wang and He 2024; García-San-Martín et al. 2025). \n \nAnimal models enhance our understanding of the biological underpinnings of structural \nsimilarity networks and how they are reorganised during development and disease. \nThe common marmoset ( Callithrix jacchus), a New World primate, represents a key \nrung on the translational ladder between rodent and human neuroscience, combining \nthe tractability of rodent research with the translational relevance of primate research \n(Okano et al. 2012). Like rodents, marmosets have an accelerated lifespan compared \nto humans, reaching adulthood at 18-20 months (Schultz-Darken, Braun, and Emborg \n2016; Sawiak et al. 2018), which facilitates longitudinal MRI studies that collect repeat \nmeasures over the lifespan. Like humans, marmosets show prolonged postnatal \ndevelopment in the context of familial nurturing (Burkart and Finkenwirth 2015; Miller \net al. 2016), and protracted development of association cortex (Sawiak et al. 2018) , \nparticularly those areas involved in social cognition (Cerrito et al. 2024).  \n \nRepresentations of the marmoset structural connectome generated from “gold \nstandard” tract tracing data (Majka et al. 2020; Watakabe et al. 2023)  have revealed \ntopological similarities to the human brain, with a rich club of highly interconnected \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n3 \nhubs and exponentially distributed connection distances or wiring costs (Liu, Zheng, \nand Misic 2020; Theodoni et al. 2021) . These topological parallels between human \nand marmoset brain networks, combined with the suitability of marmosets for MRI \n(Schaeffer et al. 2020)  and the drive towards increasing non -invasive imaging of \nprimates as part of the 3Rs in neuroscience (Prescott and Poirier 2021) , have \nmotivated efforts to generate MRI structural networks in the common marmoset, using \nDTI tractography (Hata et al. 2023)  and structural covariance (Quah et al. 2022) . \nHowever, there are problems with both of these approaches. DTI tractography suffers \nfrom a high false positive rate, i.e., identifying tracts between cortical areas that are \nnot validated by comparable tract tracing data (Maier-Hein et al. 2017) ; and a bias \nagainst long -range connections, i.e., failing to identify tracts between physically \nseparated cortical areas that are demonstrated by tract tracing data (Donahue et al. \n2016). Whereas, structural covariance network analysis is constrained to estimate a \nnetwork representing between-subject covariance, typically from a dataset of N~100 \nsingle feature MRI scans, and therefore cannot provide the single -brain network \nanalysis that is required for MRI studies of brain network development. We anticipated \nthat similarity network analysis of MRI data in the common marmoset could circumvent \nthese issues and provide a robust methodological platform for measuring the \nneurodevelopmental trajectories of primate brain network maturation. \n \nTo investigate the biological significance of structural similarity network analysis in the \ncommon marmoset, we focused on a myelin -sensitive MRI parameter, the T1w/T2w \nratio (Glasser and Van Essen 2011) . Myelin is important to normative brain function \n(Nave and Werner 2014), shows characteristic patterns of variation across the cortex \n(Burt et al. 2018; García -Cabezas, Hacker, and Zikopoulos 2020) , and a clear \nprogramme of developmental changes in human (Flechsig 1901 , Yakovlev and \nLecours 1967). However, developmental changes in myeloarchitecture of the \nmarmoset cortex have not previously been investigated. To do this, we estimated \nMIND networks from an open cross -sectional dataset containing T1 -weighted (T1w) \nand T2 -weighted (T2w) MRI images from 446 animals spanning pre -puberty to \nsenescence (0.62 to 12.75 years; Hata et al. 2023). Specifically, we estimated the ratio \nbetween T1w and T2w images to yield parameter maps with enhanced myelin -\nsensitive contrast (Glasser and Van Essen 2011)  and estimated similarity networks \nfrom parcellated voxel-wise T1w/T2w ratio images using MIND. \n \nWe used these MRI-based network estimates to test two key hypotheses grounded in \ngene expression and development of the marmoset brain. First, assuming that \nT1w/T2w MIND network metrics are representative at macro scale of cortical (co -\n)variation in microscopic myeloarchitectonics, they should be significantly co -located \nwith molecular and cellular markers of myeloarchitecture. To test this, we \nbenchmarked T1w/T2w MIND networks against two transcriptomic datasets: (i) \nspatially-resolved in -situ hybridisatio n images of myelin basic protein ( MBP) \nmessenger RNA (mRNA) in an infant and adult marmoset (Shimogori et al. 2018; Kita \net al. 2021) ; and (ii) whole genome single nucleus RNA sequencing data from 12 \ncortical regions in 6 adult marmosets (Krienen et al. 2023) . Second, assuming that \nT1w/T2w MIND networks can be validated as cortical myelination networks, we \npredicted: (i) that T1w/T2w MIND networks would show age -related changes during \nadolescence which are consistent across individuals and predictive of chrono logical \nage; and (ii) that the cortical patterning of network changes is conserved across \nspecies, such that cortical areas occupying a similar position on the hierarchy from \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n4 \nlow-level primary sensory areas to higher -order transmodal association areas \n(Mesulam 1998; Sydnor et al. 2021)  would show similar trajectories of myelination \nnetwork development, as has been observed in human (Paquola et al. 2019). \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n5 \nResults \n \nEstimation and anatomical validation of T1w/T2w similarity networks \n \nWe estimated cortical T1w/T2w ratio (henceforth T1w/T2w) similarity networks in 446 \nindividual marmosets aged between 0.62 and 12.75 years (Hata et al. 2023)  using \nMorphometric Inverse Divergence (MIND; Sebenius et al. 2023 ; Figure 1A , \nSupplementary Figure 1 ). MIND networks use Kullback -Leibler (KL) divergence \n(Kullback and Leibler 1951) to quantify how similar or convergent every pair of cortical \nareas is in terms of their voxel -wise T1w/T2w distributions. Several methods have \nbeen used to estimate KL divergence in the context of MRI structural similarity \nnetworks (Supplementary Methods), though these methods have not been formally \ncompared. We therefore compared two estimators KL divergence, based on either \nglobal or local comparisons of distribution density, and variable parameters of each of \nthese estimators (Supplementary Table 2, Supplementary Figure 2-3), on a set of \nbiologically informed benchmarks and an age prediction task. While both algorithms \nproduced similar networks, the local estimator out -performed the global estimator \nacross all benchmarks and age prediction ( Supplementary Figure 4 ). Our \ndownstream analyses therefore used the local KL estimator, as previously \nimplemented in MIND (Sebenius et al. 2023) , to construct similarity networks from \nvoxel-level data. \n \nThe average adult (N=340, mean age = 4.3 years)  T1w/T2w similarity network \nestimated using MIND displayed high average similarity within anatomically and \nfunctionally-defined cortical zones (Paxinos et al. 2012) and strong divergence \nbetween primary sensory and prefrontal association cortical zones ( Figure 1B ). \nFurthermore, the average adult network recapitulated biological relationships with \nEuclidean distance, axonal connectivity, and cytoarchitecture at the edge level shown \nto be present in structural similarity networks in other non-human species (Seidlitz et \nal. 2018; Sebenius et al. 2023; Smith et al. 2024) . T1w/T2w similarity was anti -\ncorrelated with the Euclidean distance between regions (Spearman r = -0.27, Figure \n1C). Using retrograde tract tracing data as a “gold standard” estimate of the average \nadult marmoset connectome (Majka et al. 2020) , we observed a significant positive \ncorrelation between T1w/T2w similarity and axonal connectivity (Spearman r = 0.34, \npvariogram < 0.001, Figure 1D), reflecting homophily in cortical connectivity (Sebenius, \nDorfschmidt et al. 2025) . Using a cytoarchitectonic class assignment map from \n(Atapour et al. 2024)  (Supplementary Figure 5B), the proportion of edges between \nregions of the same cortical type increased as the network was thresholded to \nprogressively lower densities, remaining above the average of 1000 spatially \nconstrained null models (Figure 1E). This suggests that high values of T1w/T2w MIND \nrepresent the high cytoarchitectonic and myeloarchitectonic similarity between cortical \nareas of the same histologically defined class.  \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n6 \nFigure 1: T1w/T2w similarity reflects axonal connectivity and cytoarchitectonic \nsimilarity. A Similarity between pairs of cortical areal T1w/T2w voxel distributions was \nestimated using Morphometric Inverse Divergence (MIND; Sebenius et al. 2023). The \naverage adult T1w/T2w similarity network is shown on the right of the panel. B Mean \nsimilarity between cortical zones. Top and bottom 10% of mean zone -zone similarity \nis shown. C T1w/T2w similarity decays with distance. D T1w/T2w similarity is \nsignificantly correlated with axonal connectivity. E The proportion of edges between \nregions of the same cytoarchitectonic class increases as the network is thresholded \nto lower densities. The empirical network is shown in blue. The mean of 1000 spatial \nautocorrelation-preserving null networks is shown in grey, with shading indicating the \n95% confidence interval. \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n7 \nT1w/T2w similarity networks align with molecular and cellular \ntranscriptomic markers of myeloarchitecture \n \nOwing to the strong correlation of T1w/T2w ratio MRI with myelin (Glasser and Van \nEssen 2011; Glasser et al. 2014) , we hypothesised that T1w/T2w similarity would \nmirror similarity in a molecular transcriptomic marker of myeloarchitecture. To test this \nhypothesis, we downloaded spatially -resolved images of myelin basic protein ( MBP) \nmessenger RNA (mRNA) expression from the Marmoset Gene Atlas (Shimogori et al. \n2018; Kita et al. 2021), measured using in-situ hybridisation, and warped these images \nto MRI space ( Supplementary Figure 6 ). The resulting MBP mRNA volume was \ncompared with an average adult T1w/T2w volume, formed by averaging N=340 \nspatially co-registered T1w/T2w scans with a mean age of 4.3 years. Both modalities \nshow a drop-off in myelin-related signal in superficial layers of cortex, though t his is \nmore obvious in the MBP mRNA image ( Figure 2A). Regionally parcellated cortical \nmaps of mean MBP expression and mean T1w/T2w ( Figure 2B) were positively \ncorrelated (Spearman r = 0.68, pvariogram < 0.001, Figure 2C). Standard deviation and \nskewness, quantifying the shape of each regional distribution, were less strongly, \nthough still significantly, correlated between MBP mRNA and T1w/T2w MRI images \n(Supplementary Figure 7).  \n \nGiven the high correspondence between T1w/T2w and MBP expression, we further \nhypothesised that MBP MIND networks would show a highly similar organisation to \nT1w/T2w MIND networks. Since each cortical area in the MBP mRNA image contains \na distribution of voxel values of MBP expression, we could estimate the MBP MIND \nnetwork using identical methods and a commensurate cortical parcellation as was \nused to estimate T1w/T2w MIND networks ( Figure 2D ). Across all 26,335 edges, \nT1w/T2w and MBP similarity were significantly correlated (Spearman r = 0.37), and \ncorrelations were lowest for regions highest in the cortical hierarchy (Spearman r = -\n0.46, pvariogram < 0.001, Figure 2E). \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n8 \nFigure 2: T1w/T2w similarity aligns with inter -areal similarity of myelin basic \nprotein (MBP) gene expression. A Qualitative comparison of cortical myelination in \nMBP mRNA and T1w/T2w images in an example region, the left primary motor cortex. \nFor T1w/T2w, the colour scale is capped at the 90th percentile of intracortical values \nto improve visualisation. B Mean MBP expression and T1w/T2w per cortical area. C \nRegional mean MBP expression and T1w/T2w were significantly correlated across all \ncortical areas. D Adult MIND networks of inter-areal similarity of MBP expression and \nT1w/T2w, grouped into cortical zones. E Regional correlations between the MIND \nnetwork edge weights linking each cortical area to the rest of the brain in MBP and \nT1w/T2w MIND networks. Spearman’s rho for the correlation between MBP MIND and \nT1w/T2w MIND edge weights is shown for each area on a  cortical map (top panel). \nThe strength of positive correlation was greater for specialised sensory and motor \nareas positioned lower in the cortical hierarchy (bottom panel). \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n9 \nStudies using bulk transcriptomic data on human cortex have found that: (1) mean \nT1w/T2w correlates strongly with the first principal component of gene expression \n(Burt et al. 2018); and (2) macroscale structural similarity is correlated with gene co -\nexpression (Sebenius et al. 2023) . We therefore hypothesised that T1w/T2w MIND \nsimilarity would be correlated with gene co-expression. To assess this hypothesis, we \nused single nucleus RNA sequencing data (Krienen et al. 2023) from 12 cortical areas \n(Figure 3A , Supplementary Table 3 ). The use of single cell transcriptomic data \nallowed us to assess the relationship between T1w/T2w MIND and gene co -\nexpression within distinct cell types while avoiding cell type compositional differences \nas a potential confounder (Repsilber et al. 2010) . We estimated whole genome co -\nexpression matrices for 6 cell types (Figure 3A) and correlated these with a matched \naverage adult T1w/T2w MIND network (N=340, mean age = 4.3 years). The cell type-\nspecific correlation between whole genome co -expression and MIND similarity was \nstrongest for glutamatergic neurons (Spearman r = 0.43; uncorrected p = 0.038); \nalthough no cell types remained significant after correction for multiple (6) \ncomparisons (Figure 3B). When we repeated this analysis substituting MBP MIND \nedge weights for T1w/T2w MIND edge weights, we found much stronger relationships \nwith cell -type specific genome co -expression. Inter -areal co -expression by \nglutamatergic neurons was most strongly correlated with MBP MIND  (Spearman r = \n0.67, pFDR < 0.001). Whole genome co -expression was significantly correlated with \nMBP MIND in all neuronal and macroglial cell types after correction for multiple \ncomparisons (Figure 3C). \n \nCortical cell types contain considerable heterogeneity (Krienen et al. 2020, 2023). We \ntherefore hypothesised that there may be cell subtypes driving the relationship \nbetween gene co-expression and MIND similarity. We elected to test this relationship \nwithin GAD+ interneurons as previous work has indicated that markers of distinct \ninterneuron subtypes are differentially correlated with T1w/T2w and other measures \nof cortical hierarchy in diverse species, with parvalbumin (PV+) interneurons being \nparticularly strongly co-located with cortical hierarchy (Kim et al. 2017; Burt et al. 2018; \nFulcher et al. 2019; Chen et al. 2023) . We therefore estimated the correlations \nbetween interneuronal subtype -specific whole genome co -expression and T1w/T2w \nMIND edge weights (as in Figure 3B-C) for a set of 10 interneuron subtypes defined \na priori by hierarchical clustering (Krienen et al. 2023, Supplementary Figure 8). We \nobserved that both T1w/T2w MIND and MBP MIND were most strongly correlated with \nwhole genome co -expression by PV+ and vasoactive intestinal peptide (VIP+) \ninterneuronal subtypes ( Figure 3D ). Correlations between T1w/T2w MIN D and \ngenome co-expression by type 2 PV+ (Spearman r = 0.42, uncorrected pvariogram = \n0.034) and VIP+ (Spearman r = 0.37, uncorrected pvariogram = 0.047) interneurons were \nnominally significant. Correlations between MBP MIND similarity and whole genome \nco-expression by PV+ and VIP+ interneurons remained significant after correction for \nmultiple comparisons. \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n10 \nFigure 3: Myeloarchitectonic similarity is associated with co -expression of \ngenes in glutamatergic neurons and PV+ and VIP+ interneurons . A Cell type -\nspecific gene co-expression matrices estimated using single nucleus RNA sequencing \ndata from 12 cortical areas (Krienen et al. 2023). In the following plots, bars represent \nthe correlation (Spearman’s rho) between myeloarchitectonic (T1w/T2w or MBP \nexpression) similarity networks and whole genome co-expression networks. Asterisks \nindicate the level of significance after comparison with  spatial autocorrelation \npreserving nulls and Bonferroni correction for multiple comparisons: * p < 0.05, ** p < \n0.01, *** p < 0.001. B T1w/T2w MIND is most strongly associated with whole genome \nco-expression by glutamatergic neurons. C MBP MIND was signif icantly correlated \nwith whole genome co -expression by all neuronal and macroglial cell types, most \nstrongly with glutamatergic neurons. D Both T1w/T2w MIND and MBP MIND were \nmost strongly correlated with gene co-expression in PV+ and VIP+ interneurons, with \nMBP correlations remaining significant after correction for multiple comparisons. \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n11 \nAge-related changes in cortical T1w/T2w \n \nWe modelled age-related changes in a subset of N=210 marmosets with ages ranging \nfrom 7 to 36 months (pre-puberty to mature adulthood; Figure 4A). We used multiple \nlinear regressions to estimate the linear slope of mean T1w/T2w change over this age \nrange ( Figure 4B ). Corroborating previous findings in humans (Baum et al. 2022; \nBoroshok et al. 2023) , we found that primary sensory regions lower in the cortical \nhierarchy showed greater age -related increases in T1w/T2w, while regions higher in \nthe hierarchy showed smaller age-related increases in T1w/T2w (Spearman r = -0.61, \npvariogram = 0.05, Figure 4C), though this trend did not reach significance. Furthermore, \nβage coefficients were significantly positively correlated with pre -pubertal (< 1 year) \nmean T1w/T2w (Spearman r = 0.69, p variogram < 0.001), indicating that regions which \nare more highly myelin ated before puberty become relatively more myelinated over \nthe course of adolescence: a “rich get richer” pattern of myelination ( Supplementary \nFigure 9A ). β age coefficients were robust to the inclusion of covariates acting as \nadditional T1w/T2w image normalisations ( Supplementary Methods , \nSupplementary Figure 9B). Furthermore, these results were not unique to MRI; rate \nof change in mean T1w/T2w was highly correlated with the difference in mean MBP \nexpression between a 6-month-old (pre-pubertal) male and 4-year-old (mature adult) \nmale animal (Spearman r = 0.60, pvariogram < 0.001, Figure 4D). \n \nFigure 4: Marmosets show conserved adolescent increases in myelination. A \nAnimals used for developmental analysis ranged from 7 months (pre -puberty) to 3 \nyears (mature adulthood). B Rate of change in mean T1w/T2w for each region during \nthis period, assessed using multiple linear regressions. C Rate of change in mean \nT1w/T2w is negatively associated with a region’s position in the cortical hierarchy. D \nRate of change in mean T1w/T2w is significantly correlated with change in mean MBP \nexpression. \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n12 \nDevelopmental changes in myelination similarity \n \nWe hypothesised that T1w/T2w MIND would also show age -related changes \nrepresentative of the inter-areally coordinated maturation of cortical myeloarchitecture. \nTo qualitatively assess whether T1w/T2w MIND networks showed age -related \nchanges, we assigned individual networks for animals aged <3 years to one of 7 age-\ndefined bins (each spanning ≥4 months) and estimated the mean network for each \nage bin and the correlations between all pairs of averaged networks ( Figure 5A ). \nCorrelations between networks progre ssively decreased as the separation in age \nbetween them increased, suggesting that there were indeed graded network changes \noccurring during development. After 20 months, roughly corresponding to the \ntermination of adolescence (Sawiak et al. 2018), correlations between networks were \nmore robust to variation in the age difference between them indicating a slowing or \ntermination of developmental change in T1w/T2w MIND. This maturational plateau \nwas consistently found to begin between 18 -21 months in sensitivity analyses using \nbin sizes of 3 and 6 months (Supplementary Figure 10). \n \nTo assess the extent to which T1w/T2w MIND networks capture developmental \nchanges which are chronologically consistent across individuals, we used machine \nlearning models to predict subject age. Models were trained to predict age from an \nindividual scan using three distinct sets of features: 1) regional mean T1w/T2w signal \nitself, 2) MIND networks summarised by the weighted degree of each node and 3) all \nedge weights in the MIND network. We measured performance using the partial \ncorrelation between predicted and true age, corrected for sex and eTIV (Sebenius et \nal. 2023), and the mean absolute error (MAE) between predicted and true ages. Over \n50 splits of the data, T1w/T2w MIND network features out -performed regional mean \nT1w/T2w in predicting age ( Figure 5B), indicated by higher partial correlations and \nlower MAE. Example model predictions ( Figure 5C ) illustrate a plateau in age \npredictions around 2 years, evident in the degree -based prediction. MIND edge \nweights and nodal degree gave more accurate predictions than regional mean \nT1w/T2w, supporting the added value of MIND  network metrics in characterising \nintracortical maturation of myelin. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n13 \n  \n \nFigure 5: T1w/T2w MIND network edges and nodal degrees predict age more \naccurately than regional mean T1w/T2w. A Networks averaged into ≥4-month age \nbins become decreasingly correlated as the age between them increases. Correlations \nremain high after 20 months, indicating that there is relatively little change in T1w/T2w \nsimilarity beyond this point. B Performance of age prediction models trained on mean \nT1w/T2w and network features (degree and edge weights) over 50 splits of the data. \nC Typical model performance for each predictor. Performance of the model with the \nmedian partial R across 50 folds is shown here. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n14 \nTo identify specific age-related changes in nodal degree and edge weights of T1w/T2w \nMIND networks that were driving the superior predictive performance of MIND network \nfeatures (Figure 5) we used multiple linear regression. At the degree level, 73/230 \nregions showed significant linear changes after correction for multiple comparisons \n(Supplementary Figure 11A ). Anterior temporal, mainly auditory, areas showed \nsignificant degree decreases, reflecting an increasingly differentiated \nmyeloarchitectonic structure relative to the rest of the cortex (Supplementary Figure \n11B). Conversely, some somatosensory, primary motor, and visual areas showed \nsignificant degree increases, reflecting myeloarchitectonic convergence with the rest \nof the cortex (Supplementary Figure 11B). This pattern of changes was robust to the \naddition of estimated total intracranial volume as a covariate (Supplementary Figure \n11C). \n  \nWe performed the same regression analyses at each edge. The β age coefficient of \neach edge, representing the linear rate of change in similarity between each pair of \ncortical nodes, is shown in Figure 6A , grouped into cortical zones. This revealed \nbidirectional changes in similarity of certain cortical areas which were not apparent in \nthe degree-level analyses. For example, orbitofrontal cortex (OFC) showed decreases \nin similarity to posterior parietal (PPC), auditory (AUD), and visual (VIS) cortices, and \nincreases in similarity to dorsolateral prefrontal (DLP) and anterior cingulate (ACC) \ncortices. Mean age -related changes between pairs of cortical zones is visualised in \nSupplementary Figure 12A. \n \nWe next used agglomerative hierarchical clustering as a data driven approach to \nexplore how cortical areas were coordinated in terms of their adolescent maturation. \nThis approach iteratively groups cortical areas into clusters based on how statistically \nsimilar their pattern of edge-level changes are. The two-cluster solution separated the \ncortex into one cluster containing mainly primary sensory and motor areas (“sensory”) \nand one cluster containing mainly frontal and paralimbic association areas \n(“association”: Figure 6B, Supplementary Figure 12B ). In keeping with these \nidentities, clusters were anatomically differentiable, with sensory cluster having \nsignificantly higher mean degree of lamination (t = 2.46, pvariogram = 0.048) and \nsignificantly lower mean hierarchical position (t = -2.98, pvariogram = 0.027; Figure 6C). \nClusters also showed distinct mean T1w/T2w trajectories, the sensory cluster \nbecoming significantly more myelinated during adolescence and increasingly \ndivergent from the association cluster (Figure 6D). When repeating this analysis using \nthree or four clusters, the association cluster remained largely intact, suggesting that \nits constituent regions are strongly maturationally coupled ( Supplementary Figure \n12C). \n \nWe replicated edge -level findings using MBP expression data. Here, developmental \nchange was approximated using the difference between a mature adult (4 -year-old) \nand pre -pubertal (6 -month-old) male MBP MIND network ( Supplementary Figure \n13A). This revealed large decreases in similarity between primary sensory and \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n15 \nprefrontal cortical regions (Supplementary Figure 13B). Clustering of the MBP MIND \ndevelopmental matrix aligned with the T1w/T2w MIND developmental matrix, yet also \nreflected modality-specific developmental trends. This was evidenced by a moderate \ncophenetic distance correlation (Spearman r = 0.20; Supplementary Methods) and \ndifferences in overlap of cortical zones with clusters (Supplementary Figure 13C-D). \nHowever, MBP MIND developmental clusters demonstrated the same pattern of \nanatomical differences bet ween clusters ( Supplementary figure 15E ), supporting \noverall correspondence between the two modalities in estimating developmental \nchanges in cortical myelination networks. \n \n \nFigure 6: Sensory and association cortex show distinct age -related changes in \nmyelination network properties. A Annual change in T1w/T2w similarity between all \npairs of cortical areas grouped by cortical zones. Matrix shows the left hemisphere. B \nHierarchical clustering reveals two cortical clusters with similar patterns of maturation. \nMatrix shows both hemispheres. C The sensory cluster had significantly higher mean \ndegree of lamination and significantly lower mean hierarchical position. Reported p -\nvalues were derived from two -tailed, two-sample t-tests evaluated against spatially \nautocorrelated null models. D Maturational clusters were differentiated by their linear \nmean T1w/T2w trajectories. Asterisks represent the statistical significance of the β age \ncoefficient in multiple linear regressions predicting mean T1w/T2w of each cluster \nusing age and sex as covariates. \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n16 \nDiscussion \nWe estimated structural similarity networks from myelin -sensitive MRI images in the \ncommon marmoset and largely confirmed our prior hypotheses: (1) that MRI -derived \ncortical similarity networks are closely co -located with commensurate maps derived \nfrom mic roscopic measurements of myelin basic protein gene expression and \ntranscription at the cellular level; and (2) that MRI -derived cortical myeloarchitectonic \nsimilarity shows developmental changes which are stable across individuals and are \ndifferent for regions occupying different levels of the sensory-association hierarchy, as \nin humans. We discuss these hypotheses in light of the existing literature. \nT1w/T2w ratio calculation enhances contrast related to myelin in humans (Glasser and \nVan Essen 2011) . We show, via co -location with spatially resolved brain images of \nmyelin basic protein mRNA, that this applies to T1w/T2w images obtained from \ncommon marmosets at high magnetic field strengths. T1w/T2w and MBP expression \nsimilarity network edges were highly correlated, most so for edges to primary sensory \nand motor areas lowest in the cortical hierarchy. The reduced cross -modal \nconcordance of similarity in higher -order cortical areas may arise due to their \nprolonged developmental and enhanced adult plastic ity (Glasser et al. 2014; Sydnor \net al. 2021), which would render their myeloarchitectonic structures more sensitive to \nindividual-specific environmental influences and thus increase inter -individual \nvariability in similarity of these areas. \nWe found that cortical areas with high myeloarchitectonic similarity also had similar \nwhole genome expression profiles in neurons. We interpret this finding in two ways. \nFirst, axonal connectivity is guided by the degree of matching between specific \npatterns of adhesion molecules on neuronal cell surfaces (Sperry 1963; de Wit and \nGhosh 2016) and axonally connected neurons typically exhibit greater similarity in their \nwhole genome expression profiles (Fornito, Arnatkevičiūtė, and Fulcher 2019). Thus, \nif we interpret similarity as a weak proxy of axonal connectivity emerging due to \nhomophily (Sebenius, Dorfschmidt et al. 2025) – supported by a moderate (r = 0.34) \ncorrelation with tract tracing – similarity and neuronal gene co-expression should also \nbe correlated. Second, despite oligodendrocytes being the key cell type responsible \nfor myelination in the brain, myelination is a lso dependent on neuronal factors, \nincluding the expression of permissive surface molecules and neuronal activity (Emery \n2010). This interdependence may establish a coupling between local \nmyeloarchitecture and neuronal gene expression patterns.  \nThe coupling between myeloarchitectonic similarity and neuronal gene co-expression \nwas strongest for glutamatergic and PV+ and VIP+ interneurons. Previous work has \nshown that transcriptomic markers of glutamatergic neurons and interneurons (most \nstrongly PV+ but also VIP+ and SST+) correlate strongly with mean T1w/T2w (Burt et \nal. 2018; Fulcher et al. 2019)  and other indices of cortical hierarchy (Kim et al. 2017; \nChen et al. 2023). These neuronal subtypes form tightly anatomically and functionally \ninterconnected cortical microcircuits (Douglas, Martin, and Whitteridge 1989; Pfeffer \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n17 \net al. 2013)  that strongly reflect local laminar structure (Beul and Hilgetag 2014) . \nCortical areas with more similar myeloarchitecture are therefore likely to contain \nphenotypically similar cortical microcircuits and similar gene expression profiles in the \ncell types forming these circuits. \nOur second main hypothesis centred around adolescence, the age range with peak \nonset of psychiatric disorders (Solmi et al. 2022) and during which increases in cortical \nmyelin contribute to the closure of critical windows of enhanced sensitivity – and \nvulnerability – to environmental stimuli (Larsen et al. 2023) . Previous studies have \nshown that linear increases in mean T1w/T2w vary along the cortical hierarchy, with \nprimary sensory areas showing steepest increases and association areas showing \nshallowest increases over the course of adolescence (Baum et al. 2022; Boroshok et \nal. 2023). For the first time, to our knowledge, we show that this pattern is conserved \nin the common marmoset, highlighting its translational utility as a primate \nneurodevelopmental model. \nWe further demonstrate that T1w/T2w ratio similarity networks show adolescent \nchanges which are stable across individuals and can therefore predict age. Previous \nwork has shown that MIND networks generated from multiple microscale and \nmacroscale MRI featur es can predict adolescent age in humans (Sebenius et al. \n2023). Our work generalises this finding to a simpler context, using a single clinically \nubiquitous MRI feature in a primate model. We found that network features predicted \nage more accurately than mean T1w/T2w alone, suggesting that relative \nmyeloarchitectonic changes between cortical areas are more developmentally \ninformative than absolute changes in mean myelin. Though the enhanced \nperformance of edge -based prediction is unsurprising due to the large number of \nedges, it is striking that degree -based predictions, with one value per cortical area, \nout-perform mean T1w/T2w. A prior study using morphometric similarity networks \n(Seidlitz et al. 2018)  estimated by pairwise correlations of 7 mean macrostructural \nfeatures found that network features were poorer at predicting human adolescent age \nthan mean cortical thickness or volume alone (Griffiths-King, Wood, and Novak 2023). \nThough our study did not use the same features, our results suggest that when using \nvoxel-level information to estimate similarity, networks contain enhanced \ndevelopmental sensitivity relative to regional mean values. \nSimilarity network changes were organised along an axis from primary sensory to \ntransmodal association areas (Mesulam 1998; Sydnor et al. 2021) . Cortical areas \noccupying similar positions along this axis showed similar patterns of edge level \nmaturation, with particularly strong maturational coupling observed within a \nrostromedial transmodal association cluster that overlapped a putative human default \nmode network homologue identified by previous analysis of tract tracing data (Buckner \nand Margulies 2019). In human cortex, myelination networks estimated as the pairwise \ncorrelations between regional MTsat depth profiles showed conserved adolescent \nreorganisation, with developmental changes organised along a latent axis mirroring \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n18 \nthe sensory -association gradient (Paquola et al. 2019). MTsat profiles of primary \nsensory (idiotypic) cortex became increasingly divergent from paralimbic cortex, and \nintermediate areas became either more idiotypic -like or paralimbic -like, indicating a \nreinforcement of the sensory-association axis. Consistent with this, we observed that \nareas occupying distant locations in sensory -association space tended to diverge in \ntheir myeloarchitecture, with large decreases in similarity between primary sensory \nareas and prefrontal, temporal, and paralimbic association areas. Parallel evidence \nfrom rat cortex indicates that prefrontal areas become myeloarchitectonically \ndifferentiated from primary sensory areas between late adolescence and mid \nadulthood in MIND similarity networks estimated from myelin-sensitive Magnetisation \nTransfer Ratio MRI images (Smith et al. 2024) . Taken together, these findings may \nindicate that the sensory -association axis represents a conserved gradient of \nmyelination network reorganisation during adolescence, with areas co-localised along \nthis gradient a) showing similar patterns of change to a ll other cortical areas and b) \nbecoming themselves convergent in terms of their myeloarchitecture.  \nOur study contains limitations. Although sensitive to myelination, T1w/T2w ratio is not \na direct measure of myelin. Hence, though we observed strong concordance between \nmean T1w/T2w and mean MBP expression across the cortical surface, there were \ndifferences in the signal distribution shapes (standard deviation and skewness) \nbetween MRI and MBP histology. These differences may arise because: a) though \nT1w/T2w ratio calculation enhances contrast related to myelin, images are sensitive \nto non-myelin factors such as iron and water content (Glasser and Van Essen 2011; \nUddin et al. 2019)  and b) noise and partial voluming effects introduced during MRI \nacquisition and preprocessing may influence regional T1w/T2w distribution shapes, \nparticularly at the white matter and pial surfaces. These factors may contribute to the \ndifferences we observ ed in myeloarchitectonic network structure when estimated \nusing MRI or MBP expression, which in turn may explain why we observed only \nnominally significant correlations between T1w/T2w similarity and gene co -\nexpression. Here, we further anticipate that exp anding the number of cortical areas \nbeyond 12 (and the number of edges beyond 66) would allow for a finer -grained \ncomparison between structural similarity and gene co-expression.   \nSecond, we used linear modelling approaches as an interpretable way to model age -\nrelated changes, though we expect that cortical myelination similarity would show non-\nlinear trajectories. In humans, developmental increases in cortical mean T1w/T2w \nfollow non-linear trajectories, with primary sensory areas reaching peak myelination \nfirst, and association areas last (Grydeland et al. 2019; Baum et al. 2022) . A \ncomparable chronology has been observed for grey matter volume decreases in \nmarmoset cortex (Sawiak et al. 2018) , thought to reflect progressive increases in \ncortical myelination (Natu et al. 2019) . Furthermore, cortical myelin undergoes \ndynamic changes across the lifespan. Indeed, myelination begins in utero for sensory \nand motor areas (Flechsig 1901, Yakovlev and Lecours 1967), while ageing involves \ncortical de-myelination. The regional sequencing of ageing -related myelin decreases \nis thought to be the chronological reverse of the sequence of adolescent myelin \nincreases, reflecting a “first in last out” hypothesis of senescence (Grydeland et al. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n19 \n2019). The field would benefit from an exploration of non-linear lifespan trajectories of \nmyelination network features, for example using normative modelling approaches \n(Bethlehem et al. 2022), and from an investigation of how these trajectories relate to \nbrain function, cognition, and psychiatric outcomes. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n20 \nMethods \n \nT1w/T2w ratio data and preprocessing \n \nWe used T1-weighted (T1w) and T2-weighted (T2w) images acquired at 9.4T from the \nBrain/MINDS project, currently the world’s largest open -access marmoset MRI data \nresource (Hata et al. 2023) . The dataset contains cross -sectional MRI images from \nN=446 marmosets, spanning 0.62 to 12.75 years (N female = 238). Volumetric images \nwere parcellated according to the 2019 Brain/MINDS Atlas (BMA 2019; Woodward et \nal. 2018), which is based on the histological atlas defined by Paxinos et al. (2012). \nCortical areas used in our analyses are listed in Supplementary Table 1 and the full \npreprocessing pipeline is visualised in Supplementary Figure 1. Briefly, T2w images \nwere skull stripped using BrainSuite18a (Shattuck and Leahy 2002)  to improve the \nquality of subsequent image registration, for which we used advanced normalisation \ntools (ANTs: Avants et al. 2011) . An ex -vivo T2w template (Woodward et al. 2018) \nwas warped to each T2w image, after which the warp was applied to the BMA 2019 \natlas to generate a label image for each subject. Voxel-wise division of the T1w image \nby the T2w image generated a T1w/T2w rati o image which increased myelin -related \ncontrast and reduced intensity inhomogeneity (bias; Glasser and Van Essen 2011) . \nT1w and T2w images were not co -registered prior to calculation of T1w/T2w images \nbecause marmosets were anaesthetised and head-fixed before scanning, minimising \nhead motion between T1w and T2w image acquisitions. \n \nSimilarity network estimation using Morphometric Inverse Divergence \n \nFor each parcellated T1w/T2w image, a distribution of voxel values was extracted from \neach cortical area. Similarity between each pair of voxel value distributions was \nestimated using Morphometric Inverse Divergence (MIND; Sebenius et al. 2023)  \nyielding a univariate (single feature) similarity network. MIND is the inverse of the \nKullback-Leibler (KL) divergence, a statistical measure of divergence or dissimilarity, \nbetween two distributions (Kullback and Leibler 1951) . For two sets of voxels (or \nvertices from a surface reconstruction) in cortical areas a and b following distributions \nPa and Pb, the KL divergence between them is calculated as: \n \n \n(1) \n \n Where pa(x) and pb(x) are the probability densities at particular values of x. KL is then \nnormalised and inverted, converting it to a measure of similarity: \n  \n \n \n(2) \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n21 \nEstimation of MIND for each possible pair of areas in the cortical parcellation results \nin a similarity matrix, where each element (edge) represents the similarity in voxel \ndistributions between a pair of cortical areas, ranging from >0 to 1. Low values indicate \na highly dissimilar, divergent or differentiated pair of cortical areas and high values \nindicate a highly similar, convergent, or near-identical pair of cortical areas. \n \nTwo elementary features of each MIND network were the focus of further analysis: \nnodes and edges. Edge weights represent the pair -wise similarity between cortical \nnodes, and the weighted degree of each node is simply the average of the weights of \nall edges linking that node to the rest of the cortical network.    \n \nEstimation of KL divergence \n \nNon-parametric estimation of the KL divergence between two (voxel) distributions \nrequires a systematic comparison of probability densities, which can be done either \nglobally or locally. Global methods involve comparing coarse partitions of the \ndistributions, for example corresponding histogram bins, while local methods involve \ncomparing densities in the local neighbourhoods surrounding individual distribution \nvalues, i.e. voxel values (Supplementary Methods). Though both global and local KL \nestimation methods have previously been used to estimate structural similarity \nnetworks from MRI data (Kong et al. 2014; Sebenius et al. 2023), they have not been \ndirectly compared. We addressed this gap in the literature by systematically comparing \na global and local KL estimator. We used a global estimator of KL which binned the \ndata into successive histograms (Wang, Kulkarni, and Verdu 2005)  using in-house \ncode. Secondly, we used the local estimator implemented in the current MIND \nalgorithm (Sebenius et al. 2023)  based on a k -Nearest Neighbours (k -NN) \napproximation (Supplementary Methods). \n \nTo facilitate comparisons between similarity networks generated using different KL \nestimators, and variable parameters of each of those estimators, we adopted a \npreviously used set of anatomical benchmarks (Sebenius et al. 2023) . Anatomical \nbenchmarking assumes that the “best -performing” or “most optimal” MRI similarity \nnetworks are ones that correspond most strongly with the underlying anatomy. In our \ncase, benchmarks were based on cytoarchitecture, interhemispheric symmetry, an d \naxonal connectivity ( Supplementary Methods ). We additionally compared MRI \nsimilarity networks resulting from these KL estimation algorithms using an age \nprediction task. Here, the networks producing most accurate age predictions, \nassessed via their corre lation and mean squared error relative to actual ages, were \ndeemed “most optimal”. \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n22 \nMarmoset anatomical data \n \nCytoarchitecture \n  \nWe used the cytoarchitectonic class assignment map of Atapour et al. (2024), in which \ncortical regions were assigned to one of six classes based on expert classification of \nhistological sections (Supplementary Figure 5B). In order of increasing lamination, \nthe classes are: Agranular (Ag), Dysgranular (Dys), Eulaminate I (Eu I), Eulaminate II \n(Eu II), Eulaminate (Eu III), and Koniocortex (Kon). \n \nAxonal connectivity \n  \nTract tracing data were obtained from Majka et al. (2020 ). Data from 143 retrograde \ntracer injections in 52 adult marmosets (N female = 21, aged 1.4-4.6 years) were used \nto generate a single {55 x 116} tract tracing matrix. Retrograde tracer injections were \nplaced in 55 target regions, labelling axonal tracts o riginating from cell bodies in 116 \nsource regions. Quantitative analysis of histological sections was used to estimate the \nfraction of labelled neurons (FLN) originating in each source brain region for each of \nthe 55 target regions receiving a tracer injec tion. The logarithm of the estimated \nfraction of labelled neurons in each region, Log10(FLN), averaged across animals, was \nused as the key metric of axonal connectivity. Log 10(FLN) ranges from 0 (100% of \ntracts originate in source) to -∞ (0% of tracts originate in source). Because this \napproach fails to capture axonal pathways terminating in the 61 regions not injected \nwith tracer, we used the edge-complete matrix of {55 x 55} regions that had received \nat least one injection. For comparab ility, MIND similar ity matrices were restricted to \nthe same 55 regions in all analyses involving tract tracing data. \n \nPosition in the cortical hierarchy \n \nA cortical area can be assigned to a position in the cortical processing hierarchy based \non its laminar pattern of incoming and outgoing axonal connections (Felleman and \nVan Essen 1991). This is based on the observation that feedforward connections, \ncarrying information up the cortical hierarchy from sensory to association regions, \narise mainly from supragranular layers and terminate in layer IV (Rockland and \nPandya 1979) . The proportion of tracts a cortical area receives which originate in \nsupragranular layers has therefore been used to estimate cortical maps of hierarchical \nposition in the marmoset (Theodoni et al. 2021) , with regions receiving a high \nproportion of feedforward information being located high in the hierarchy and regions \nreceiving a low proportion of feedforward information being located low. We used the \nhierarchical position map from Theodoni et al. (2021 ), which assigns to 110 of 115 \ncortical areas a value ranging from 0 (bottom of the hierarchy: primary visual cortex) \nto 1 (top of the hierarchy: agranular insula). This map is visualised in Supplementary \nFigure 5C. \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n23 \nMarmoset transcriptional data and pre-processing   \n \nIn-situ hybridisation data \n \nWe downloaded spatially resolved images of myelin basic protein ( MBP) messenger \nRNA (mRNA) from the Marmoset Gene Atlas (Shimogori et al. 2018; Kita et al. 2021). \nImages were collected using in -situ hybridisation (ISH), a histochemical technique \nallowing spatial localisation of specific nucleic acids within a tissue sample (Wilcox \n1993). We downloaded in -situ hybridisation images of MBP mRNA from a 6 -month-\nold (pre -pubertal) male and a 4 -year-old (mature adult) male marmoset . For both \nanimals, MBP mRNA was measured in 90 coronal brain slices, spaced 196 μm apart \n(Shimogori et al. 2018; Kita et al. 2021) . Though the posterior pole of occipital cortex \nand the anterior pole of frontal cortex were not completely covered, each brain region \nin the Paxinos et al. (2012) parcellation had at least one representative tissue sample \nanalysed at both ages. \n \nTo facilitate a quantitative comparison of T1w/T2w and MBP mRNA images, in -situ \nhybridisation slices were stacked to form a volume, which was then linearly and \nnonlinearly deformed to the BMA 2019 T2w template. To do this, we implemented a \npreprocessing pipeline, fully described in Supplementary Figure 6 , that used a \ncombination of in-house code and code made available from Tong et al. (2022). This \npipeline allowed us to estimate the distribution of MBP mRNA expression in each \ncortical area, with comparable spatial resolution to the T1w/T2w images (isotropic \nvoxel resolution of MBP mRNA images = 0.10mm; T1w/T2w images = 0.27mm). The \nresulting cortical maps were subsequently used to estimate MIND networks \nrepresenting inter-areal similarity of MBP expression on the same spatial scale and \nusing the same areal parcellation as the MIND networks derived from T1w/T2w \nimages. \n \nSingle-cell, whole genome mRNA data \n \nWe used open-access single nucleus RNA sequencing data from Krienen et al. (2023). \nThis dataset contains whole genome transcripts from multiple cortical and subcortical \nregions comprising six cell types: glutamatergic excitatory neurons (glut), glutamic acid \ndecarboxylase-expressing inhibitory interneurons (GAD), oligodendrocytes (oligo), \nastrocytes (astro), microglia and macrophages (immune), and endothelial cells (endo). \nFurther details on data collection, quality control, and cell type classification are  \nprovided by Krienen et al. (2023). \n \nCortical dissections in Krienen et al. (2023) were guided by the Paxinos et al. (2012) \natlas. Some dissections were restricted to individual Paxinos regions, while others \nwere inclusive of several neighbouring regions, constituting a larger amalgamated \narea (Krienen, personal communication). This re sulted in a partial parcellation of \ncortex into 12 relatively coarse-grained areas (Figure 3A, Supplementary Table 3). \nWe used this scheme to coarse -grain the more finely parcellated T1w/T2w and MBP \nmRNA images so that they were anatomically aligned with t he cortical regions for \nwhich single -cell sequencing data were available. Further details on the coarse \nparcellation can be found in Supplementary Table 3 , and the sex and age of the \nsampled animals are described in the figure legend. \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n24 \nWe used these data to generate cell type -specific correlated gene expression (CGE) \nmatrices. Cell -by-gene matrices for each cell type were downloaded from the \nCELLxGENE data repository ( RRID:SCR_021059). Library size normalisation, i.e. \nnormalising gene expression across cell types to correct for differences in sequencing \ndepth, had already been done by Krienen et al. (2023). We retained the top 10% most \nvariable genes for CGE network calculation using Scanpy’s highly_variable_genes \nfunction (Wolf, Angerer, and Theis 2018) . Our reasoning was that genes with low \nvariability in expression across the cortex are unlikely to drive systemic variation in \ncortical myeloarchitecture. Indeed, previous work has shown that transcriptomic \nvariation in a subset of 200 genes that are most highly variable across cortex is closely \ncoupled to variation in mean T1w/T2w of human visual cortex (Gomez, Zhen, and \nWeiner 2019). Within each cell type, we averaged log normalised expression across \ncells with the same regional annotation to form a region -by-gene expression matrix. \nPearson correlations between each pair of regional gene expression vectors were \nused to generate the final CGE matrix for each cell type ( Figure 3A ). The same \nprocess was applied to interneuron subtypes to generate interneuron subtype-specific \ncorrelated gene expression matrices. \n \nStatistical analysis \n \nMultiple linear regression \n \nTo estimate linear changes in nodal properties (mean T1w/T2w and degree), and edge \nweights during development, we used multiple linear regressions with age as the \nindependent variable. Multiple comparisons testing was conducted using Benjamini \nHochberg correction. We included sex as a covariate in all models as studies have \ndemonstrated sex differences in regional cortical mean T1w/T2w in humans \n(Küchenhoff et al. 2024). Furthermore, some studies indicate that humans display sex-\nspecific adolescent trajectories of cortical mean T1w/T2w (Norbom et al. 2020) , \ntherefore we tested for the significance of an age by sex interaction term in our models. \nHowever, this interaction term was ultimately dropped from all models as it did not \nreach significance in any model after multiple comparisons correction. Furthermo re, \nsensitivity analyses revealed that addition of two normalisation covariates to our \nT1w/T2w models, and the addition of estimated total intracranial volume (eTIV), to \ndegree models did not greatly change βage values (Supplementary Figure 9B, 11C). \nFor further discussion on covariate considerations, see Supplementary Methods. \n \nSpatial null models \n \nIn spatially embedded systems, such as the brain, regions that are physically closer \nto each other tend to be more similar across a wide range of features, including axonal \nconnectivity profiles, cytoarchitectonic structure, and gene expression (Fornito, \nArnatkevičiūtė, and Fulcher 2019; Markello and Misic 2021). This spatial dependence, \nor autocorrelation, between neighbouring brain areas violates the assumption of \nindependence in standard statistical inference procedures and can lead to inflated type \nI error rates if not properly addressed (Markello and Misic 2021) . We therefore used \nspatial autocorrelation-preserving null models, generated using BrainSMASH (Burt et \nal. 2020), for statistical inference on parcellated cortical maps. BrainSMASH \nimplements an algorithm that estimates for a particular brain map (for instance, mean \nT1w/T2w) the variance in regional values as a function of the Euclidean distance \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n25 \nbetween regions (the variogram) and generates surrogate brain maps with matched \nvariograms (Viladomat et al. 2014)  that can be used to construct an appropriate null \ndistribution of test statistics. BrainSMASH permuted maps were computed from left \nhemispheric data only. We used a similar procedure to assess the significance of \ncorrelations with similarity networks at the edge level (Supplementary Methods). All \nstatistical tests used 1000 permutations. \n \nAge prediction \n \nTo assess the extent to which MIND networks capture developmental changes in \nmyelination, we trained machine learning models to predict age from T1w/T2w scans \nin N=210 animals aged 7 months (pre-pubertal) to 36 months (mature adult). Repeated \n5-fold cross validation with 50 random permutations of the data was used to compare \nage prediction performance for models trained using either i) mean T1w/T2w per \ncortical area, ii) MIND weighted degree per cortical area, or iii) MIND edges between \neach pair of cortical areas. Each fold was balanced for sex and subjects in the top or \nbottom 1% of mean T1w/T2w or MIND weighted degree (averaged over all cortical \nareas) were excluded to reduce the impact of outliers. For each outer test fold, models \nwere evaluated using th e mean absolute error (MAE) and the partial correlation \nbetween actual and predicted age values, corrected for sex and estimated total \nintracranial volume. For details on quality control, model specification, and model \nselection, please refer to Supplementary Methods. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n26 \nData availability \nAll data analysed in this study are publicly available. Retrograde tract tracing data are \navailable from Majka et al. (2020). Region-wise hierarchical position assignments are \navailable from Theodoni et al. (2021). Region -wise cytoarchitectonic class \nassignments are available from Atapour et al. (2024). MRI data (Hata et al. 2023) are \navailable from the Brain/MINDs 2.0 data portal \n(https://dataportal.brainminds.jp/marmoset-mri-na216). In -situ hybridisation data \n(Shimogori et al. 2018 and Kita et al. 2021) a re available from the Brain/MINDs \nMarmoset Gene Atlas data portal (https://gene-atlas.brainminds.jp/). Processed single \nnucleus RNA -sequencing data (Krienen et al. 2023) are available from the \nCELLxGENE data repository (RRID:SCR_021059; \nhttps://cellxgene.cziscience.com/collections/0fd39ad7-5d2d-41c2-bda0-\nc55bde614bdb). Marker gene sets used to assign biological identities to interneuron \nclusters are available from Bakken et al. (2021). \nCode availability \nPreprocessing of MRI data used Brainsuite18a (Shattuck and Leahy 2002) and ANTs \n(Avants et al. 2011 ). Preprocessing of in -situ hybridisation data used a modified \nversion of previously published code (Tong et al. 2022; \nhttps://github.com/TrangeTung/marmoset_gradient) and ANTs. Computation of MIND \nnetworks used previously published code (Sebenius et al. 2 023; \nhttps://github.com/isebenius/MIND). Downstream analyses were conducted in python. \nProcessing of single cell data used Scanpy (Wolf et al. 2018). All custom code used \nin this study will be made publicly available on GitHub upon publication of the final \narticle. \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n27 \nReferences \n \n1. Atapour N, Rosa MGP, Bai S, Bednarek S, Kulesza A, Saworska G, et al. \nDistribution of calbindin -positive neurons across areas and layers of the \nmarmoset cerebral cortex. PLoS Comput Biol. 2024 Sep;20(9):e1012428.  \n2. Avants BB, Tustison NJ, Song G, Cook PA, Klein A, Gee JC. A reproducible \nevaluation of ANTs similarity metric performance in brain image registration. \nNeuroimage. 2011 Feb;54(3):2033–44.  \n3. Bassett DS, Sporns O. Network neuroscience. Nat Neurosci. 2017 \nFeb;20(3):353–64.  \n4. Baum GL, Flournoy JC, Glasser MF, Harms MP, Mair P, Sanders AFP, et al. \nGraded variation in T1w/T2w ratio during adolescence: Measurement, caveats, \nand implications for development of cortical Myelin. J Neurosci. 2022 \nJul;42(29):5681–94.  \n5. Bethlehem RAI, Seidlitz J, White SR, Vogel JW, Anderson KM, Adamson C, et \nal. Brain charts for the human lifespan. Nature. 2022 Apr;604(7906):525–33.  \n6. Beul SF, Hilgetag CC. Towards a “canonical” agranular cortical microcircuit. \nFront Neuroanat. 2014;8:165.  \n7. Boroshok AL, McDermott CL, Fotiadis P, Park AT, Tooley UA, Gataviņš MM, et \nal. Individual differences in T1w/T2w ratio development during childhood. Dev \nCogn Neurosci. 2023 Aug;62(101270):101270.  \n8. Buckner RL, Margulies DS. Macroscale cortical organization and a default -like \napex transmodal network in the marmoset monkey. Nat Commun. 2019 \nApr;10(1):1976.  \n9. Bullmore E, Sporns O. Complex brain networks: graph theoretical analysis of \nstructural and functional systems. Nat Rev Neurosci. 2009 Mar;10(3):186–98.  \n10. Burkart JM, Finkenwirth C. Marmosets as model species in neuroscience and \nevolutionary anthropology. Neurosci Res. 2015 Apr;93:8–19.  \n11. Burt JB, Demirtaş M, Eckner WJ, Navejar NM, Ji JL, Martin WJ, et al. Hierarchy \nof transcriptomic specialization across human cortex captured by structural \nneuroimaging topography. Nat Neurosci. 2018 Sep;21(9):1251–9.  \n12. Burt JB, Helmer M, Shinn M, Anticevic A, Murray JD. Generative modeling of \nbrain maps with spatial autocorrelation. NeuroImage. 2020 Oct;220:117038.  \n13. Cerrito P, Gascon E, Roberts AC, Sawiak SJ, Burkart JM. Neurodevelopmental \ntiming and socio -cognitive development in a prosocial cooperatively breeding \nprimate (Callithrix jacchus). Sci Adv. 2024 Nov;10(44):eado3486.  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n28 \n14. Chen A, Sun Y, Lei Y, Li C, Liao S, Meng J, et al. Single -cell spatial \ntranscriptome reveals cell -type organization in the macaque cortex. Cell. 2023 \nAug;186(17):3726-3743.e24.  \n15. de Wit J, Ghosh A. Specification of synaptic connectivity by cell surface \ninteractions. Nat Rev Neurosci. 2016 Jan;17(1):22–35.  \n16. Donahue CJ, Sotiropoulos SN, Jbabdi S, Hernandez -Fernandez M, Behrens \nTE, Dyrby TB, et al. Using diffusion tractography to predict cortical connection \nstrength and distance: A quantitative comparison with tracers in the monkey. J \nNeurosci. 2016 Jun;36(25):6758–70.  \n17. Douglas RJ, Martin KAC, Whitteridge D. A canonical microcircuit for neocortex. \nNeural Comput. 1989 Dec;1(4):480–8.  \n18. Emery B. Regulation of oligodendrocyte differentiation and myelination. \nScience. 2010 Nov;330(6005):779–82.  \n19. Felleman DJ, Van Essen DC. Distributed hierarchical processing in the primate \ncerebral cortex. Cereb Cortex. 1991;1(1):1–47.  \n20. Flechsig P. Developmental (myelogenetic) localisation of the cerebral cortex in \nthe human subject. Lancet. 1901 Oct;158(4077):1027–30.  \n21. Fornito A, Arnatkevičiūtė A, Fulcher BD. Bridging the gap between connectome \nand transcriptome. Trends Cogn Sci. 2019 Jan;23(1):34–50.  \n22. Fulcher BD, Murray JD, Zerbi V, Wang XJ. Multimodal gradients across mouse \ncortex. Proc Natl Acad Sci U S A. 2019 Mar;116(10):4689–95.  \n23. García-Cabezas MÁ, Hacker JL, Zikopoulos B. A protocol for cortical type \nanalysis of the human neocortex applied on histological samples, the Atlas of \nVon Economo and Koskinas, and magnetic resonance imaging. Front \nNeuroanat. 2020 Dec;14:576015.  \n24. García-San-Martín N, Bethlehem RAI, Mihalik A, Seidlitz J, Sebenius I, Alemán-\nMorillo C, et al. Molecular and micro -architectural mapping of gray matter \nalterations in psychosis. Mol Psychiatry. 2025 Apr;30(4):1287–96.  \n25. Glasser MF, Goyal MS, Preuss TM, Raichle ME, Van Essen DC. Trends and \nproperties of human cerebral cortex: correlations with cortical myelin content. \nNeuroimage. 2014 Jun;93 Pt 2:165–75.  \n26. Glasser MF, Van Essen DC. Mapping human cortical areas in vivo based on \nmyelin content as revealed by T1 - and T2 -weighted MRI. J Neurosci. 2011 \nAug;31(32):11597–616.  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n29 \n27. Gomez J, Zhen Z, Weiner KS. Human visual cortex is organized along two \ngenetically opposed hierarchical gradients with unique developmental and \nevolutionary origins. PLoS Biol. 2019 Jul;17(7):e3000362.  \n28. Griffiths-King D, Wood AG, Novak J. Predicting ‘Brainage’ in late childhood to \nadolescence (6 -17yrs) using structural MRI, morphometric similarity, and \nmachine learning. Sci Rep. 2023 Sep;13(1):15591.  \n29. Grydeland H, Vértes PE, Váša F, Romero -Garcia R, Whitaker K, Alexander -\nBloch AF, et al. Waves of maturation and senescence in micro -structural MRI \nmarkers of human cortical myelination over the lifespan. Cereb Cortex. 2019 \nMar;29(3):1369–81.  \n30. Hata J, Nakae K, Tsukada H, Woodward A, Haga Y, Iida M, et al. Multi -modal \nbrain magnetic resonance imaging database covering marmosets with a wide \nage range. Sci Data. 2023 Apr;10(1):221.  \n31. Kim Y, Yang GR, Pradhan K, Venkataraju KU, Bota M, García Del Molino LC, \net al. Brain -wide maps reveal stereotyped cell -type-based cortical architecture \nand subcortical sexual dimorphism. Cell. 2017 Oct;171(2):456-469.e22.  \n32. Kita Y, Nishibe H, Wang Y, Hashikawa T, Kikuchi SS, U M, et al. Cellular -\nresolution gene expression profiling in the neonatal marmoset brain reveals \ndynamic species - and region -specific differences. Proc Natl Acad Sci U S A. \n2021 May;118(18).  \n33. Kong XZ, Wang X, Huang L, Pu Y, Yang Z, Dang X, et al. Measuring individual \nmorphological relationship of cortical regions. J Neurosci Methods. 2014 \nNov;237:103–7.  \n34. Krienen FM, Goldman M, Zhang Q, C H Del Rosario R, Florio M, Machold R, \net al. Innovations present in the primate interneuron repertoire. Nature. 2020 \nOct;586(7828):262–9.  \n35. Krienen FM, Levandowski KM, Zaniewski H, Del Rosario RCH, Schroeder ME, \nGoldman M, et al. A marmoset brain cell census reveals regional specialization \nof cellular identities. Sci Adv. 2023 Oct;9(41):eadk3986.  \n36. Küchenhoff S, Bayrak Ş, Zsido RG, Saberi A, Bernhardt BC, Weis S, et al. \nRelating sex -bias in human cortical and hippocampal microstructure to sex \nhormones. Nat Commun. 2024 Aug;15(1):7279.  \n37. Kullback S, Leibler RA. On Information and Sufficiency. Annals of Mathematical \nStatistics. 1951 Mar;22(1):79–86.  \n38. Larsen B, Sydnor VJ, Keller AS, Yeo BTT, Satterthwaite TD. A critical period \nplasticity framework for the sensorimotor -association axis of cortical \nneurodevelopment. Trends Neurosci. 2023 Oct;46(10):847–62.  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n30 \n39. Lecours, I.R. Y PA. The Myelogenetic Cycles of Regional Maturation of the \nBrain. In: Regional Development of the Brain in Early Life. Oxford: Blackwell; \n1967. p. 3–70.  \n40. Lerch JP, van der Kouwe AJW, Raznahan A, Paus T, Johansen-Berg H, Miller \nKL, et al. Studying neuroanatomy using MRI. Nat Neurosci. 2017 Feb;20(3):314–\n26.  \n41. Liu ZQ, Zheng YQ, Misic B. Network topology of the marmoset connectome. \nNetw Neurosci. 2020 Dec;4(4):1181–96.  \n42. Maier-Hein KH, Neher PF, Houde JC, Côté MA, Garyfallidis E, Zhong J, et al. \nThe challenge of mapping the human connectome based on diffusion \ntractography. Nat Commun. 2017 Nov;8(1):1349.  \n43. Majka P, Bai S, Bakola S, Bednarek S, Chan JM, Jermakow N, et al. Open \naccess resource for cellular-resolution analyses of corticocortical connectivity in \nthe marmoset monkey. Nat Commun. 2020 Feb;11(1):1133.  \n44. Markello RD, Misic B. Comparing spatial null models for brain maps. \nNeuroimage. 2021 Aug;236(118052):118052.  \n45. Mesulam MM. From sensation to cognition. Brain. 1998 Jun;121 ( Pt 6):1013 –\n52.  \n46. Miller CT, Freiwald WA, Leopold DA, Mitchell JF, Silva AC, Wang X. \nMarmosets: A neuroscientific model of human social behavior. Neuron. 2016 \nApr;90(2):219–33.  \n47. Natu VS, Gomez J, Barnett M, Jeska B, Kirilina E, Jaeger C, et al. Apparent \nthinning of human visual cortex during childhood is associated with myelination. \nProc Natl Acad Sci U S A. 2019 Oct;116(41):20750–9.  \n48. Nave KA, Werner HB. Myelination of the nervous system: mechanisms and \nfunctions. Annu Rev Cell Dev Biol. 2014;30(1):503–33.  \n49. Norbom LB, Rokicki J, Alnaes D, Kaufmann T, Doan NT, Andreassen OA, et \nal. Maturation of cortical microstructure and cognitive development in childhood \nand adolescence: A T1w/T2w ratio MRI study. Hum Brain Mapp. 2020 \nNov;41(16):4676–90.  \n50. Okano H, Hikishima K, Iriki A, Sasaki E. The common marmoset as a novel \nanimal model system for biomedical and neuroscience research applications. \nSemin Fetal Neonatal Med. 2012 Dec;17(6):336–40.  \n51. Paquola C, Bethlehem RA, Seidlitz J, Wagstyl K, Romero -Garcia R, Whitaker \nKJ, et al. Shifts in myeloarchitecture characterise adolescent development of \ncortical gradients. Elife. 2019 Nov;8.  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n31 \n52. Paxinos G. The marmoset brain in stereotaxic coordinates. 1st ed. London: \nAcademic Press; 2012.  \n53. Pfeffer CK, Xue M, He M, Huang ZJ, Scanziani M. Inhibition of inhibition in \nvisual cortex: the logic of connections between molecularly distinct interneurons. \nNat Neurosci. 2013 Aug;16(8):1068–76.  \n54. Prescott MJ, Poirier C. The role of MRI in applying the 3Rs to non -human \nprimate neuroscience. Neuroimage. 2021 Jan;225(117521):117521.  \n55. Quah SKL, McIver L, Bullmore ET, Roberts AC, Sawiak SJ. Higher-order brain \nregions show shifts in structural covariance in adolescent marmosets. Cereb \nCortex. 2022 Sep;32(18):4128–40.  \n56. Repsilber D, Kern S, Telaar A, Walzl G, Black GF, Selbig J, et al. Biomarker \ndiscovery in heterogeneous tissue samples -taking the in -silico deconfounding \napproach. BMC Bioinformatics. 2010 Jan;11(1):27.  \n57. Rockland KS, Pandya DN. Laminar origins and terminations of cortical \nconnections of the occipital lobe in the rhesus monkey. Brain Res. 1979 \nDec;179(1):3–20.  \n58. Sawiak SJ, Shiba Y, Oikonomidis L, Windle CP, Santangelo AM, Grydeland H, \net al. Trajectories and milestones of cortical and subcortical development of the \nmarmoset brain from infancy to adulthood. Cereb Cortex. 2018 \nDec;28(12):4440–53.  \n59. Schaeffer DJ, Liu C, Silva AC, Everling S. Magnetic resonance imaging of \nmarmoset monkeys. ILAR J. 2020 Dec;61(2–3):274–85.  \n60. Schultz-Darken N, Braun KM, Emborg ME. Neurobehavioral development of \ncommon marmoset monkeys: Marmoset Neurobehavioral Development. Dev \nPsychobiol. 2016 Mar;58(2):141–58.  \n61. Sebenius I, Dorfschmidt L, Seidlitz J, Alexander-Bloch A, Morgan SE, Bullmore \nE. Structural MRI of brain similarity networks. Nat Rev Neurosci. 2025 \nJan;26(1):42–59.  \n62. Sebenius I, Seidlitz J, Warrier V, Bethlehem RAI, Alexander -Bloch A, Mallard \nTT, et al. Robust estimation of cortical similarity networks from brain MRI. Nat \nNeurosci. 2023 Aug;26(8):1461–71.  \n63. Seidlitz J, Váša F, Shinn M, Romero -Garcia R, Whitaker KJ, Vértes PE, et al. \nMorphometric similarity networks detect microscale cortical organization and \npredict inter-individual cognitive variation. Neuron. 2018 Jan;97(1):231-247.e7.  \n64. Shattuck DW, Leahy RM. BrainSuite: an automated cortical surface \nidentification tool. Med Image Anal. 2002 Jun;6(2):129–42.  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n32 \n65. Shimogori T, Abe A, Go Y, Hashikawa T, Kishi N, Kikuchi SS, et al. Digital gene \natlas of neonate common marmoset brain. Neurosci Res. 2018 Mar;128:1–13.  \n66. Smith RL, Sawiak SJ, Dorfschmidt L, Dutcher EG, Jones JA, Hahn JD, et al. \nDevelopment and early life stress sensitivity of the rat cortical microstructural \nsimilarity network. bioRxivorg. 2024 Dec;2024.12.20.629759.  \n67. Solmi M, Radua J, Olivola M, Croce E, Soardo L, Salazar de Pablo G, et al. \nAge at onset of mental disorders worldwide: large -scale meta-analysis of 192 \nepidemiological studies. Mol Psychiatry. 2022 Jan;27(1):281–95.  \n68. Sperry RW. Chemoaffinity in the orderly growth of nerve fiber patterns and \nconnections. Proc Natl Acad Sci U S A. 1963 Oct;50(4):703–10.  \n69. Sydnor VJ, Larsen B, Bassett DS, Alexander-Bloch A, Fair DA, Liston C, et al. \nNeurodevelopment of the association cortices: Patterns, mechanisms, and \nimplications for psychopathology. Neuron. 2021 Sep;109(18):2820–46.  \n70. Theodoni P, Majka P, Reser DH, Wójcik DK, Rosa MGP, Wang XJ. Structural \nattributes and principles of the neocortical connectome in the marmoset monkey. \nCereb Cortex. 2021 Nov;32(1):15–28.  \n71. Tong C, Liu C, Zhang K, Bo B, Xia Y, Yang H, et al. Multimodal analysis \ndemonstrating the shaping of functional gradients in the marmoset brain. Nat \nCommun. 2022 Nov;13(1):6584.  \n72. Uddin MN, Figley TD, Solar KG, Shatil AS, Figley CR. Comparisons between \nmulti-component myelin water fraction, T1w/T2w ratio, and diffusion tensor \nimaging measures in healthy human brain structures. Sci Rep. 2019 \nFeb;9(1):2500.  \n73. Viladomat J, Mazumder R, McInturff A, McCauley DJ, Hastie T. Assessing the \nsignificance of global and local correlations under spatial autocorrelation: a \nnonparametric approach: Test for the Correlation When the Variables Are \nSmoothed. Biometrics. 2014 Jun;70(2):409–18.  \n74. Wang H, Jin X, Zhang Y, Wang J. Single-subject morphological brain networks: \nconnectivity mapping, topological characterization and test-retest reliability. Brain \nBehav. 2016 Apr;6(4):e00448.  \n75. Wang J, He Y. Toward individualized connectomes of brain morphology. \nTrends Neurosci. 2024 Feb;47(2):106–19.  \n76. Wang Q, Kulkarni SR, Verdu S. Divergence estimation of continuous \ndistributions based on data -dependent partitions. IEEE Trans Inf Theory. 2005 \nSep;51(9):3064–74.  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint \n\nMapping myelination networks in the common marmoset cortex            Hutchings et al. 2025 \n33 \n77. Watakabe A, Skibbe H, Nakae K, Abe H, Ichinohe N, Rachmadi MF, et al. Local \nand long-distance organization of prefrontal cortex circuits in the marmoset brain. \nNeuron. 2023 Jul;111(14):2258-2273.e10.  \n78. Wilcox JN. Fundamental principles of in situ hybridization. J Histochem \nCytochem. 1993 Dec;41(12):1725–33.  \n79. Wolf FA, Angerer P, Theis FJ. SCANPY: large -scale single -cell gene \nexpression data analysis. Genome Biol. 2018 Feb;19(1):15.  \n80. Woodward A, Hashikawa T, Maeda M, Kaneko T, Hikishima K, Iriki A, et al. The \nBrain/MINDS 3D digital marmoset brain atlas. Sci Data. 2018 Feb;5(1):180009.  \n \nAcknowledgements \nE.D.H was generously supported by a Rosetrees  Trust grant MB2023 \\100002 and \nfunding from the University of Cambridge School of Clinical Medicine. S.J.S. and \nA.C.R were supported by a Medical Research Council Programme grant \nMR/V033492/1. E.T.B. was supported by a National Institute for Health and Care \nResearch (NIHR) Senior Investigator award.  \nAuthor information \nContributions \nE.D.H. and E.T.B. conceived of the primary methodology. E.D.H. performed all \nanalyses and drafted the manuscript. R.A.I.B., A.C.R., and E.T.B. provided \nsupervision. S.J.S. advised on analyses of age -related changes. E.D.H., S.J.S., \nR.A.I.B., A.C.R., and E.T.B. reviewed and edited the manuscript. \nEthics declaration \nCompeting interests \nThe authors declare the following competing interests: E.T.B. has recently consulted \nBoehringer Ingelheim, SR One, Novartis, GlaxoSmithKline, Sosei Heptares, and \nMonument Therapeutics. R.A.I.B. and E.T.B. hold equity in and are cofounders of \nCentile Bioscience Inc. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 28, 2025. ; https://doi.org/10.1101/2025.09.26.678906doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}