2.1 Sample

A total of 318 healthy adolescents and young adults between the ages of 14 and 24, recruited in London and Cambridgeshire, underwent MRI scanning as part of the *Neuroscience in Psychiatry Network project. Subjects were selectively recruited from a larger questionnaire cohort (N > 2,400) to ensure sex, age and ethnicity distribution representative of the community. Exclusion criteria were any self-reported neurological, developmental or psychiatric disorders. Subjects were scanned at baseline and a subset returned for a follow-up scan on average 1.3 (SD = 0.3) years later. After quality control, data from 293 subjects were included in further analyses. The study was approved by the Cambridge Central Research Ethics Committee (12/EE/0250) and all participants (and their legal guardian where participants were under 16 years old) gave written informed consent.

2.2 P-factor

We utilised p-factor or “general symptomatology” scores at baseline and follow-up, as previously validated and published by St Clair et al. (2017), where full modelling details and fit indices can be found. In brief, exploratory and confirmatory factor analyses yielded an optimal model with a bifactor structure, including a general p-factor and five specific factors. The general factor explained 92% of the reliable variance (Rodriguez, Reise, & Haviland, 2016) in total scores and as such, the specific factors explained little beyond p. Due to this and our a priori interest in general psychopathology, the specific factors were not the focus of further analysis here. Analyses were conducted at item level on items derived from self-report measures of anxiety (Revised Children's Manifest Anxiety Scale), depression (Moods and Feelings Questionnaire), psychotic-spectrum symptoms (Schizotypal Personality Questionnaire), obsessionality (Revised Leyton Obsessional Inventory), antisocial behaviour (Antisocial Behaviour Questionnaire, self-esteem (Rosenberg Self-Esteem Questionnaire) and well-being (Warwick-Edinburgh Mental Well-Being Scale) (Goodyer et al., 2011; Kiddle et al., 2017). Here, we assess p-factor scores in the MRI subsample; however, importantly, bifactor analysis was conducted on the full questionnaire cohort (N = 2,228) (St Clair et al., 2017), resulting in more reliable and representative factor scores. As a result, the MRI sample demonstrated satisfactory stability of scores over time, with a correlation of R = .65 (p < .001) between baseline and follow-up p-factor scores.

2.3 MRI acquisition and processing

Subjects were scanned on identical Siemens Magnetom TIM Trio whole-body 3T MRI scanners in Cambridge and London as per the multi-parameter mapping (MPM) protocol, which has previously undergone multi-centre validation and is described in detail elsewhere (Weiskopf et al., 2013; Weiskopf et al., 2015). Acquisition parameters were identical across sites. Whole-brain multi-echo FLASH MT weighted contrast were acquired at 1 mm isotropic resolution (TR: 23.7, α = 6°, 176 sagittal slices, FOV = 256 × 240 mm², matrix = 256 × 240 × 176). Semi-quantitative MT saturation maps were derived using biophysical models as described in more detail in Helms, Dathe, Kallenberg, and Dechent (2008) and Tabelow et al. (2019) using the hMRI toolbox (www.hmri.info) for SPM (Wellcome Centre for Human Neuroimaging, London, UK, http://www.fil.ion.ucl.ac.uk/spm). These MT maps represent the percentage of signal loss as a result of an off-resonance MT pre-pulse, which preferentially saturates the macro-molecular bound water pool, known to be a sensitive measure for myelin content (Schmierer et al., 2004; Turati et al., 2015). In contrast to the commonly used MT ratio, this semi-quantitative MT saturation metric explicitly removes bias induced by spatially varying T1 relaxation times and flip angle inhomogeneities.

MT maps were longitudinally pre-processed using a custom pipeline, as described in detail in Ziegler, Hauser, et al. (2019). Briefly, each subject's baseline and follow-up scans were first registered to each other using symmetric diffeomorphic registration (Ashburner & Ridgway, 2013), resulting in a midpoint image for each subject. Midpoint images were subsequently segmented using the Computational Anatomy Toolbox (CAT, http://www.neuro.uni-jena.de/cat/). The maps were then normalised to MNI space using SPM's geodesic shooting and tissue-weighted smoothing with a Gaussian kernel of 3 mm full width at half maximum was applied. Tissue-weighted smoothing (Draganski et al., 2011) was chosen so as to preserve MT values within grey/white matter tissue classes and thereby account for small discrepancies in spatial normalisation, ensuring that obtained MT values within the tracts would not be contaminated by non-white matter tissue values.

Rigorous quality assessment included manual inspection for motion artefacts by an expert (G.Z.) and the use of statistical covariance-based inhomogeneity measures (as implemented in the CAT toolbox), which detects subjects with extreme overall deviation of quantitative values. In addition, as a proxy of motion during the scan, the SD parameter of R2* exponential decay residuals (SDR2*) in white matter areas was computed. This has been shown to provide a reliable measure of motion across scans in the context of MPMs (Castella et al., 2018; Ziegler, Hauser, et al., 2019). Scans with SDR2* values above 2.5 SDs from the sample mean were removed from all further analyses.

The final data set included in our analyses consisted of 283 baseline scans and 203 follow-up scans (from a total of 293 subjects).

2.4 Region of interest histogram extraction

We defined 10 tract ROIs based on the JHU White Matter Tractography Atlas (Hua et al., 2008) corresponding to the ATR, dorsal cingulum bundle, hippocampal cingulum bundle, corticospinal tract (CST), forceps major, forceps minor, inferior fronto-occipital fasciculus (IFOF), ILF, SLF and UF. Probability maps for each of these tracts were thresholded at 20% to minimise partial voluming. Separate unilateral masks for each hemisphere were created for all ROIs except for forceps major and forceps minor, which remained bilateral. For each MT map, we summed the number of voxels within each ROI for 100 uniform bins within a range of MT values from 0.8 to 2.2. We then normalised each histogram with respect to its area and extracted the histogram mean (first moment), variance (second moment) and skewness (third moment). Therefore, each ROI yielded three metrics for every subject at each available time point within each hemisphere (with the exception of forceps major and minor, which were not split by hemisphere).

2.5 Statistical analysis

3.1 Identifying tracts showing developmental increase of myelin-sensitive MT

The null model (M0) was not significantly improved by the addition of age in the CST (χ² = 1.31, p = .520), forceps major (χ² = 2.64, p = .267) and forceps minor (χ² = 5.20, p = .074), suggesting there were no (longitudinal or cross-sectional) age effects on mean MT in these tracts.

The additive model (M1) including main effects of longitudinal and cross-sectional age was the winning model for the ATR (χ² = 36.34, p < .0001), dorsal cingulum (χ² = 32.63, p < .0001), hippocampal cingulum (χ² = 37.83, p < .0001) and SLF (χ² = 23.25, p < .0001). All four ROIs showed a significant positive effect of longitudinal age (all ps < .0001), indicating a within-subject increase of MT in these tracts (see Figure 1). Only the hippocampal cingulum showed an additional significant effect of cross-sectional age, indicating higher MT in older subjects in this cohort, p = .002. A significant effect of laterality indicated higher MT values in the right compared to the left hemisphere for the ATR, dorsal cingulum, hippocampal cingulum, IFOF and ILF, and higher MT in the left compared to right hemisphere for the SLF and UF, all p < .001.

There was a significant negative interaction between longitudinal age and cross-sectional age (M2) in the IFOF (χ² = 11.40, p < .0001), ILF (χ² = 9.39, p = .002) and UF (χ² = 8.66, p = .003). As seen in Figure 2, longitudinal MT increase was strongest in youngest adolescents in these tracts, whereas virtually no increase was seen in early adulthood. IFOF and ILF showed higher MT values in the right compared to the left hemisphere, whereas there were higher mean MT values in the left compared to the right UF, all ps < .001. Neither the rate of change of MT nor cross-sectional age effects on MT differed significantly between the two hemispheres or between the two sexes in any of the reported ROIs.

3.2 Association between development of MT and p-factor scores

We conducted additional model comparisons for tracts showing significant MT increase (ATR, dorsal and hippocampal cingulum, SLF, IFOF, ILF and UF) as well as tracts not showing an overall MT increase (CST, forceps major and forceps minor). Including the p-factor in the model improved the model significantly for dorsal cingulum (χ² = 20.56, p < .0001) and UF (χ² = 13.13, p = .004). In both of these tracts, there was a significant negative interaction between cross-sectional p-factor and longitudinal age, indicating that longitudinal MT change over study visits in the dorsal cingulum and UF differed as a function of p. As seen in Figure 3, subjects showing high p-factor scores showed the least MT increase, whereas subjects with low p-factor scores showed the strongest MT increase in these tracts. The models for both tracts were not improved by the addition of a three-way interaction between cross-sectional p-factor, longitudinal age and sex, indicating that the depicted effect did not differ between male and female participants.

None of the models revealed significant main effects of cross-sectional or longitudinal p-factor, indicating no direct correlation between mean levels of p and mean level of myelination, and no correlation between longitudinal change in p and longitudinal change in MT.

3.3 Histogram variance and skewness

Histogram variance in the CST and forceps minor, as well as skewness in the CST and ATR, were best explained by a model including longitudinal and cross-sectional age effects (all χ² > 16.08, p < .005), as seen in Supplementary Table 1. Specifically, histogram variance in the CST and forceps minor decreased as a function of cross-sectional age, indicating a narrower distribution of MT values in older subjects in these regions. Skewness in the ATR and CST was also negatively impacted by cross-sectional age, suggesting that older subjects exhibit more negatively skewed distributions compared to younger subjects. In the CST, this tendency was also reflected in a negative effect of longitudinal age on histogram skewness, p < .001. There were no significant age by laterality interactions in any of these tracts. Furthermore, the addition of the p-factor did not improve model fit for any of the ROIs. Further details of ROI-specific model comparisons for histogram variance and skewness are found in Supplementary Table 1.

4.1 Developmental findings

MT, likely reflecting myelin (Turati et al., 2015), increased in the majority of white matter tracts throughout adolescence, a finding that aligns with much existing literature on white matter development, albeit a literature that mostly relies on diffusion imaging metrics (Bava et al., 2010; Krogsrud et al., 2016; Lebel, Walker, Leemans, Phillips, & Beaulieu, 2008; Schmithorst & Yuan, 2010; Simmonds et al., 2014; Wang et al., 2012). These metrics, such as fractional anisotropy (FA), tend to be sensitive to white matter microstructural change, but are less specific to myelin changes per se (Beaulieu, 2002; Wheeler & Voineskos, 2014). Recent reviews of developmental change using diffusion metrics indicate a consensus in a view that there are considerable regional differences in both rate and timing of white matter maturation (Lebel et al., 2017; Lebel & Deoni, 2018; Schmithorst & Yuan, 2010; Tamnes, Roalf, Goddings, & Lebel, 2018). Our study lends support to a literature, which suggests a protracted maturation of frontotemporal connections, including within the cingulum and SLF that extends well into adulthood (Lebel et al., 2008; Tamnes et al., 2009). Maturation of dorsal cingulum, in particular, has been shown to occur much later compared to other tracts (Lebel et al., 2017; Simmonds et al., 2014). However, we did not replicate a similar pattern of late maturation of the UF, which is often observed in terms of FA (Lebel et al., 2017). Instead, in our data, the longitudinal change in UF decelerated post-adolescence. This observation aligns more closely with findings of earlier UF maturation in terms of mean diffusivity (MD) compared to FA (Lebel et al., 2008; Wang et al., 2012), indicating that MD may be more reflective of myelin changes.

We found evidence for a deceleration in myelin change in the IFOF and ILF, manifest as MT increase levelling off with age. This is consistent with reports that both these tracts reach 90% of their adult FA values before adulthood, and certainly earlier than the cingulum bundle, which continues to undergo change up to the late 20s (Lebel et al., 2008; Simmonds et al., 2014). Nonlinear trajectories of white matter development have been reported in a number of large neuroimaging studies, with regionally specific peaks reached at variable time points in late adolescence and early adulthood (Lebel & Deoni, 2018). We note that nonlinear development could not be explicitly modelled longitudinally in our analysis due to the presence of only two scan time points (Fjell et al., 2010), though we were able to test for nonlinear development indirectly by assessing whether longitudinal change differed over the sampled age range.

We found no evidence for age-dependent myelin growth in the forceps major and minor, compatible with previous diffusion imaging findings that microstructural development in the anterior (genu) and posterior (splenium) corpus callosum is mostly complete by early adolescence (Lebel et al., 2008; Pohl et al., 2016; Rollins et al., 2010). Callosal microstructural changes in adolescence and adulthood, if present, are more frequently reported in the callosum body (Lebel & Beaulieu, 2011; Rollins et al., 2010), which was not included in our analyses. The forceps major, containing the splenium of the corpus callosum, shows a particularly fast FA increase before the age of 13 (Mah, Geeraert, & Lebel, 2017), consistent with a posterior-to-anterior gradient of myelin maturation in the brain.

Interestingly, we did not observe any longitudinal change in mean MT within the CST. This is inconsistent with previous studies suggesting that CST is one of the latest maturing tracts in terms of diffusion parameters (Lebel et al., 2017). The absence of MT increase in our sample could be due to the specific studied age range, as slopes may temporarily flatten off consistent with the idea of a putative interim period in some tracts during which there is no myelin growth during adolescence (Simmonds et al., 2014). However, intriguingly, we observed an effect of cross-sectional age on both variance and skewness of MT in the CST, findings reflecting a narrowing of the MT distribution with age as well as a shift of the mass of the distribution towards higher values. The latter was also evident longitudinally. MT saturation imaging and diffusion parameters likely index different aspects of white matter microstructure, with MT showing higher specificity for myelin. It has been suggested that discrepancies in the developmental findings from different white matter neuroimaging techniques might be explained by a greater sensitivity of diffusion metrics to, for example, axonal packing (Lebel & Deoni, 2018). As such, a more protracted change in axonal packing (rather than or in addition to myelin per se) might be more readily reflected in a change in variance (rather than the mean) of MT values due to increased tissue homogeneity. Furthermore, if myelin maturation does indeed continue for longer on average than is the case for other tracts, but shows greater variability in the exact timing of the peak, this may be more readily reflected in changes in skewness than in mean values. This highlights the importance of not only cross-modal imaging and replication but also the use of within-modality metrics that go beyond simple averages (which also appear to be differentially sensitive to longitudinal and cross-sectional effects).

4.2 P-factor findings

Crucially, we found that increased expression of p-factor scores was associated with reduced rates of MT increase, which we interpret as slower myelination, in dorsal cingulum and UF. Both tracts constitute core-connecting pathways of distinct limbic networks and have been implicated in several neuropsychiatric disorders (Buckholtz & Meyer-Lindenberg, 2012; Passamonti et al., 2012). The temporo-amygdala-orbitofrontal network, connected through UF, is implicated in the integration of emotion and cognition (Catani, Dell'Acqua, De Schotten, & Reviews, 2013), and is particularly important for processes such as flexible reward learning (Schoenbaum, Roesch, Stalnaker, & Takahashi, 2009; Stalnaker, Franz, Singh, & Schoenbaum, 2007) and emotional memory (LaBar & Cabeza, 2006; Rauch et al., 1996). The DMN, the medial portion of which (posterior cingulate/precuneus to anterior cingulate/medial prefrontal cortex) is connected through the dorsal cingulum, shows deactivation during goal-directed tasks and is implicated in introspective and self-referential thought (Raichle et al., 2001).

Functional and structural imaging studies point towards an involvement of the DMN and temporo-amygdala-orbitofrontal networks in general psychopathology during development. In one study, delayed maturation of functional connectivity within the DMN was associated with childhood psychopathology (Sato et al., 2016), while in young adults functional hyperconnectivity between visual association cortex and DMN has been associated with higher p-factor scores (Elliott et al., 2018). A recent cross-sectional diffusion imaging study in over 700 healthy adolescents showed that general psychopathology, as well as cognitive performance, was associated with frontotemporal white matter disconnectivity at the intersection of the UF and IFOF (Alnæs et al., 2018). Consistent with our findings, the latter suggests that the structural integrity of frontotemporal circuits may be a transdiagnostic brain phenotype for psychiatric vulnerability. Our results extend these earlier findings in two crucial ways: first, we show that myelination may underpin this association; second, by using a longitudinal design we demonstrate that—with respect to MT—it is the rate of change that is the most relevant feature for general psychopathology during development. Indeed, absolute MT values showed no relationship to p-factor scores in our study, underscoring intra- rather than inter-individual maturational status as a critical factor. Importantly, our findings provide a potential neural mechanism underlying the ability of the p-factor to predict future functioning, as found in adolescents (Patalay et al., 2015) and adults (Lahey et al., 2012) over periods as long as 3 years. In this framework, the persistence of problems likely reflected in the p-factor appears to be underpinned by slowed maturation over time of white matter tracts central to cognitive and emotional processing.

With respect to cortical maturation, our findings dovetail with observations in this same data set that highlight adolescence as a key period for consolidation of highly connected hubs in association cortices, driven in large part by myelination and linked to the expression of genes enriched for risk of schizophrenia (Whitaker et al., 2016). In addition, aberrant myelin development in intra- and juxta-cortical regions has been specifically linked to an expression of compulsive and impulsive traits (Ziegler, Hauser, et al., 2019), highlighting the link between non-normative longitudinal brain maturation trajectories and psychiatric vulnerability in adolescence.

4.3 Limitations and outlook

An open question remains as to whether altered myelin maturation is an underlying cause of increased psychopathology, an adverse outcome thereof, or a concomitant effect of other disrupted brain processes more central to psychopathology. Myelin is known to be subject to substantial experience-dependent changes (Cicchetti, 2002; Grossman et al., 2003), a notion firmly embedded in the childhood adversity literature (Daniels et al., 2013; Ziegler et al., 2019). For example, chronic hyperactivity of the hypothalamic–pituitary–adrenal axis as a result of life stressors is thought to result in a slowing down of myelin development (Sapolsky, 1992). Conversely, there is tentative evidence that genetic dysregulation of myelin development causally underlies specific psychiatric symptoms, particularly in schizophrenia (Hakak et al., 2001; Nave & Ehrenreich, 2014), a finding strengthened by animal studies that experimentally manipulate genes specifically in oligodendrocytes (Roy et al., 2007). Therefore, the most likely account of the association between neurodevelopment and psychopathology may be one of bidirectional and mutually exacerbating influences between the two (Nigg, 2016), but further neuroepidemiological studies are needed to explicitly address this hypothesis.

A further potential limitation concerns the specificity of MT saturation imaging to myelin content. MT saturation is one of several MRI metrics commonly used to index myelin, including estimations of the myelin water fraction (MWF) (e.g., derived from multicomponent T1 and T2 relaxometry, Deoni, Rutt, Arun, Pierpaoli, & Jones, 2008), T1w/T2w ratio, or the longitudinal relaxation rate R1. While there is evidence for substantial correlations with histological measures of myelin for MT (Schmierer et al., 2004), MWF (Laule et al., 2006) and R1 (Stueber et al., 2014), imperfect correlations between these metrics suggest that each may be differentially sensitive to distinct aspects of tissue microstructure (Geeraert et al., 2017; Hagiwara et al., 2018; O'Muircheartaigh et al., 2019). Quantitative, or semi-quantitative, MT related metrics are suggested as more robust measures of myelin compared to MWF (Dula, Gochberg, Valentine, Valentine, & Does, 2010) and T1w/T2w ratio (Hagiwara et al., 2018), as well as more specific to myelin compared to R1 and R2* (Callaghan et al., 2014), which show additional sensitivity to iron content. However, MT measures may also be sensitive to the effects of neuroinflammation or oedema (Gareau, Rutt, Karlik, & Mitchell, 2000; Stanisz, Webb, Munro, Pun, & Midha, 2004) and as such remain an imperfect measure of myelin. Future research might usefully address the specificity of our longitudinal findings to myelination by comparing MT saturation effects with other white matter microstructure metrics, an approach adopted in recent work in studies of late childhood and early adolescence, albeit reports on cross-sectional and not longitudinal data (Geeraert, Lebel, & Lebel, 2019; Moura et al., 2016).

Since we excluded subjects with self-reported mental health diagnoses, our sample reflects a relatively healthy population at the time of study. This selection approach is in line with the general notion surrounding the p-factor and comes with a number of advantages. First, the p-factor as measured in healthy community cohorts is thought to reflect “distress” that is phenomenologically continuous with clinical expression of psychopathology (Caspi et al., 2014; St Clair et al., 2017) and subject to identical neurodevelopmental, as well as environmental and genetic, influences. Second, several studies in non-clinical community samples have shown that the p-factor is predictive of later adverse mental health and life outcomes (Laceulle, Chung, Vollebergh, & Ormel, 2019; Lahey et al., 2012; Pettersson, Lahey, Larsson, & Lichtenstein, 2018), suggesting that assessing p in healthy individuals, particularly during apparently healthy development, is a useful tool for estimating risk for distal clinical outcomes (Snyder, Young, & Hankin, 2017). Nevertheless, we recognise that certain neurobiological discontinuities may also exist between healthy, subclinical and clinical populations and that future research is necessary to test explicitly the correspondence of neural mechanisms underlying general psychopathology construct in healthy and clinical samples.