Material deprivation is associated with neural resilience for late-life depression

late-life depression, material deprivation, neural resilience, neuroanatomical signatures . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice. 2

1 Late-life depression (LLD) is a major source of global morbidity and mortality, 2 influenced by multiple risk factors. Yet, a major challenge is to quantify the 3 degree of resilience or vulnerability to LLD at the individual level, which could 4 offer neurobiological insight and ultimately inform future interventions and 5 treatment. Here, applying a non-parametric regression model to the UK Biobank 6 data (N=1,988), we quantified brain-based resilience and vulnerability to LLD and 7 tested whether risk factors could explain individual differences in the estimated 8 magnitude of such neural resilience and vulnerability. Our results show that 9 social isolation was positively associated with the median magnitude of neural 10 vulnerability whereas material deprivation was negatively associated with the 11 greatest neural resilience (top 10 percentile). These results together highlight the 12 importance of social interaction and access to sufficient resources and services 13 in diminishing neural vulnerability and promoting neural resilience to LLD, 14 respectively. Our findings therefore provide insights into preventive strategies for 15 LLD, and thus are of importance for policy makers as well as the broader society. 16 17 18 19 20 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 3

(529) 21 Late life depression (i.e., depression in older adults aged 60+; LLD) has been 22 associated with increased risks of disability and mortality (Schulz et al., 2000). It 23 is intertwined with conditions and factors that are primarily ageing-related, 24 yielding distinctive and complex etiological and clinical profiles in contrast to 25 depression in younger age groups (Blazer, 2003; Alexopoulos, Schultz and 26 Lebowitz, 2005). Previous work has shown that risk factors such as physical 27 disability, medical illness, cognitive impairment, worse socioeconomic status, 28 greater exposures to traumatic events, less social support, and living an 29 unhealthy lifestyle contribute to higher chances of depression (Blazer, 2003; 30 Chang-Quan et al. , 2010; Chang et al. , 2016). However, these studies did not 31 investigate individual vari ation in LLD vulnerability and/or resilience and how 32 such variation may be linked to various risk factors. Quantifying LLD resilience 33 and vulnerability in patients is of great clinical relevance as it may ultimately 34 inform future interventions. Here we employed a novel approach to evaluate 35 neural resilience and vulnerability to LLD at the individual level and determined 36 whether known risk factors of LLD explain individual differences in such neural 37 resilience and vulnerability. 38 In previous studies, LLD has been associated with abnormalities in structural and 39 functional properties (Manning and Steffens, 2018). It has also been associated 40 with accelerated brain age, the magnitude of which further showed a correlation 41 with declined cognitive performance (Christman et al. , 2020). These findings 42 demonstrate the potential of proxy measures that capture the neural 43 underpinnings of the disorder. Here, we build on this prior work to assess 44 resilience and vulnerability to depressive symptom severity, using similar 45 aggregate neural measures. Specifically, we sought to quantify such resilience 46 and vulnerability from neuroanatomical patte rns that are related to recent 47 depressive symptoms and to examine whether LLD risk factors could explain 48 individual differences in the quantifi ed neural resilience and vulnerability. 49 Specifially, here we focused on a vulnerable group of older adults who have 50 experienced at least one depression episode before the assessment in this 51 study. This provides an opportunity to enhance the variance in depression 52 symptom severity by including individuals along the spectrum of potential LLD. 53 Inspired by brain-age models that predict the age of the brain based on 54 neuroanatomical features in healthy individuals (Franke and Gaser, 2019), here 55 we predicted the individual-level brain-based depression score (BDS) from 56 multimodal neuroimaging features (n=4,632; see more details in Materials and 57 Methods). The difference between the predicted BDS and the original symptom 58 score (delta BDS; ∆ BDS) was calculated to index “neural resilience” or “neural 59 vulnerability” to LLD and then linked to a set of LLD risk factors that covered a 60 wide range of sociopsychological, medical and lifestyle variables (see variable 61 full list in Materials and Methods ). We specifically focused on individuals with 62 highest neural resilience or vulnerability (i.e., top 10% negative and positive 63 ∆ BDS respectively) that represent older adults with the greatest clinical 64 relevance. In practice, the focus on these highly resilient and vulnerable 65 individuals also minimized the correlations between ∆ BDS and original symptom 66 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 4 score (i.e., similar to correlations between brain-age delta and chronological age 67 in brain-age literature; (Franke and Gaser, 2019; Smith et al. , 2019). This 68 enabled us to identify risk factors showing unique associations with neural 69 resilience and/or vulnerability. 70 71

72 73 Participants 74 75 Participants from the UK Biobank (UKB) dataset aged 60 or above at the time of 76 imaging acquisition with a history of probable major depression status (i.e., prior 77 experiences in single or recurrent depression episodes; (Smith et al. , 2013)) 78 were included to ensure sufficient variance in depression symptomatology 79 (Mage=66.35; 60.3% female). Approximately 70% (N=1,405) of the resulting 80 sample (N=1,988) had a status of recurrent major depression. Due to missing 81 data in risk factor variables, a sub-sample of N=1,464 participants were included 82 in partial correlation and quantile regression analyses with similar demographic 83 characteristics and prior depression histories as in the full sample (M age=66.18; 84 59.4% female; 70% recurrent MDD). All participants in this study provided 85 informed consent. UK Biobank has ethical approval from the North West Multi-86 Centre Research Ethics Committee (MREC). Data access was obtained under 87 UK Biobank application ID 47267. 88 89 Data acquisition and measurement 90 91 Depressive symptoms 92 A touchscreen questionnaire was implemented to collect information on 93 sociodemographic characteristics and mental health. Recent depressive 94 symptoms were measured using four items (RDS-4) that assess depressed 95 mood, disinterest, restlessness, and tiredness in the past two weeks (Dutt et al., 96 2022). This continuous measure is highly comparable to several standardized 97 self-report depression scales including PHQ-9, CES-D, and MASQ-30 (Dutt et 98 al., 2022), and shows an area under the curve of 0.79 for its correlation with 99 depression diagnosis (Khubchandani et al., 2016). In response to each of these 100 four questions, participants indicated their experiences from “not at all” (scoring 101 1) to “nearly every day” (scoring 4) such that the total symptom score ranged 102 from 4 to 16. 103 104 Imaging preprocessing and multimodal imaging features 105 All brain imaging data from the UKB were acquired using standard Siemens 106 Skyra 3T scanners with a standard Siemens 32-channel RF receive head coil. 107 Detailed information for UKB imaging acquisition can be found in the UK Biobank 108 Imaging Documentation, hosted on the Oxford FMRIB UK Biobank resource 109 page (https://www.fmrib.ox.ac.uk/ukbiobank/). A fully automated processing and 110 QC pipeline was developed for UKB brain imaging data, which included T1, T2, 111 FLAIR, susceptibility-weighted MRI, resting-state MRI, task-evoked MRI and 112 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 5 diffusion MRI (Alfaro-Almagro et al. , 2018). Additionally, this pipeline also 113 generated a set of imaging-derived phenotypes (IDPs) such as cortical and 114 subcortical structure volumes, microstructural measures in major tracts, and 115 functional connectivity metrics. In this study, a total of 4,632 IDPs was included 116 as multimodal imaging features in the prediction models for estimating brain-117 depression score (see section 3.1). Approximately three quarters of these 118 features were derived from the resting-state and task-evoked fMRI data, and the 119 remaining from structural and diffusion MRI data: 120 - Resting-state fMRI features: Amplitude of ICA100 nodes and ICA25 121 nodes; Edges of full correlation matrix from ICA100 and ICA25; Edges of 122 partial correlation matrix from ICA100 123 - Task-evoked fMRI features: Median and 90th percentile BOLD signals for 124 shapes and faces, as well as shape-face contrasts using a group mask 125 and an amygdala mask respectively; Median and 90th percentile Z-126 statistics for shapes and faces, as well as shape-face contrasts using a 127 group mask and an amygdala mask respectively 128 - Structural MRI features: Volumes of cortical and subcortical (including 129 sub-segments) structures; Total volume of white matter hyperintensities 130 - Diffusion MRI features: Mean FA, MD, MO, L1-L3, ICVF, OD, ISOVF 131 based on Standardized FA Skeleton; Weighted-mean FA, MD, MO, L1-L3, 132 ICVF, OD, ISOVF in White-matter tracts 133 134 Late-life Depression Risk factors 135 A list of risk factors for late-life depression (LLD) was curated based on the 136 literature that covers information on demographics, lifestyle, medical conditions, 137 adverse experiences, and psychosocial factors. Measures of these variables 138 were collected from participants either via a touchscreen questionnaire or a 139 verbal medical history interview on the same day of imaging acquisition. 140 Specifically, several variables such as self-reported health, long-standing illness 141 and Townsend deprivation index, which is a census-based index of material 142 deprivation calculated by the combination of four indicators of deprivation: non-143 home ownership, non-car ownership, unemployment and overcrowding 144 (Townsend, Phillimore and Beattie, 1988) were available directly from the UKB, 145 whereas aggregate measures for healthy lifestyle, sleep quality, vascular risk 146 factors, adverse or traumatic experiences, social isolation and loneliness were 147 derived using multiple items as follows: 148 - A healthy lifestyle score was constructed based on smoking status, 149 physical activity, diet, and alcohol consumption that are well documented 150 as depression risk factors (Sarris et al., 2020; van Lee et al., 2020; Kang 151 et al., 2021). Based on national recommendations, participants were given 152 1 for healthy, and 0 for unhealthy behaviors. Detailed coding information 153 can be found in (Lourida et al., 2019). 154 - A sleep (low) risk score was calculated using five sleep questions. Low-155 risk sleep factors were defined as a) having an early chronotype, b) 156 s leeping 7–8 hours per day, c) never having or rarely having insomnia 157 symptoms, d) not reporting snoring, and e) not reporting frequent daytime 158 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 6 sleepiness (Fan et al. , 2020). Participants received a score of 1 if their 159 behaviors were classified as low risk for that factor and a sum score was 160 calculated across five factors, where higher scores represent healthier 161 sleep patterns or low risks of sleep issues (Hepsomali and Groeger, 162 2021). 163 - An aggregate measure of vascular risk factors was calculated for each 164 participant by counting instances of having a high BMI (>25), having a 165 high waist-hip ratio (WHR>0.85 for females and WHR>0.90 for males), 166 having ever smoked, and a self-reported diagnosis of hypertension, 167 diabetes, or hypercholesterolemia (Cox et al. , 2019). The resulting sum 168 score of instances represent potential vascular risks, where higher scores 169 indicating higher risks. 170 - scores were calculated separately for childhood, adulthood, and lifetime 171 experiences with dichotomization of responses to each individual question 172 (Yapp et al. , 2021). Specifically, responses of “never true” to negative 173 experiences (e.g., hit hard) scored a 0 and responses of “rarely true” and 174 more often (i.e., “sometimes true”, “often”, “very often true”) scored a 1, 175 which was reversed for positive experiences (e.g., in a confiding 176 relationship). Binary responses (“yes”, “no”) were scored 1 and 0 177 respectively. Separate sum scores were calculated to indicate the 178 magnitude of traumatic experiences in different time periods and included 179 as separate predictors in the models. 180 - Psychosocial factors comprised two measures: loneliness and social 181 isolation. Participants were classified as lonely if they reported feeling 182 lonely often and if they could confide to someone close only occasionally 183 (e.g., less than once every few months), and socially isolated if they met at 184 least two criteria of 1) living alone, 2) visiting their family or friends less 185 than once a month, and 3) participating in none of the listed leisure/social 186 activities (Mutz, Roscoe and Lewis, 2021). These psychosocial factors 187 were included in statistical models as separate predictors. 188 189 190 Statistical analysis 191 192 Estimating the delta of brain-depression score 193 Our approach was inspired by the estimation of brain-age delta from 194 neuroimaging features that is defined as the difference between the estimated 195 brain age and chronological age in a given individual, which has been used to 196 indicate underlying problems in outwardly healthy people and related to the risk 197 of cognitive ageing or age-associated brain disease (Franke et al., 2010; Cole 198 and Franke, 2017; Baecker et al., 2021). In this study, we employed Multivariate 199 Adaptive Regression Splines models (Friedman and Roosen, 1995; Friedman, 200 2007) to estimate the brain-depression score (BDS) as informed by multimodal 201 imaging features and quantified the difference between the estimated and the 202 actual depression symptom scores (i.e., ∆ BDS), with a positive ∆ BDS indicating 203 neural vulnerability to depression (i.e., actual reported symptom score higher 204 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 7 than BDS) and negative ∆ BDS indicating neural resilience to depression (i.e., 205 actual symptom score lower than BDS). 206 Multivariate Adaptive Regression Splines (MARS) is a flexible regression 207 technique that can capture the intrinsic nonlinear and multidimensional 208 relationship of variables with an ensemble of linear functions joined together by 209 one or more spline basis functions, where the number of basis functions and the 210 parameters associated with each function (e.g., product degree and knot 211 locations) are determined by the data (Friedman and Roosen, 1995; Friedman, 212 2007). Specifically, MARS builds a model in two phases: the forward pass and 213 the backward pruning, similar to growing and pruning of tree models. In the 214 forward pass, MARS starts with a model consisting of just the intercept term (i.e., 215 the mean of the response values), followed by the assessment of every single 216 predictor to find a basis function pair that produces the maximum improvement in 217 the model error. This process iterates until either the model reaches a predefined 218 limit number of terms, or the error improvement reaches a predefined limit. The 219

of the forward pass is a MARS basis matrix with rows of observations and 220 columns of basis functions (e.g., hinge functions). To avoid overfitting by the full 221 terms in the basis matrix from the forward pass, the backward pass is to find the 222 subset of these terms that gives the best generalized cross validation (GCV) via 223 a stepwise term deletion procedure. This backward pruning process continues 224 until only one term remains (the intercept term) and the final model with the best 225 GCV is selected (Friedman and Roosen, 1995; Friedman, 2007). 226 In this study, IDPs from multimodal imaging data and depressive symptom 227 sum scores were included in MARS models as independent ( X) and dependent 228 (Y) variables. Additionally, age, sex, head size, head motion (i.e., the averaged 229 head motion across space and time points) during fMRI acquisition (i.e., for both 230 the resting-state and task-evoked fMRI), scanner site, scanner table position and 231 data acquisition dates were included in all MARS models as confounding 232 variables. Per partition of the full data in the nested cross-validations (see details 233 below), principal component analysis (PCA) was employed to decompose high-234 dimensional X and components collectively explaining more than 50% variance 235 were retained. As the outcome measure Y (i.e., symptom sum score) in our study 236 is highly skewed with a long tail, transformation of Y was performed per data 237 partition before modeling to enforce Gaussianity. This was realized by using a 238 data-drive approach that finds the optimal normalization method from a suite of 239 possible transformation options including the Box-Cox transformation, Yeo-240 Johnson transformation, the ordered quantile technique, Arcsinh transformation, 241 exponential transformation, square root transformation and the Lambert W x F 242 transformation (Peterson, 2021). We also applied dummy coding for all factorial 243 variables such as sex and scanner site before data partitioning to ensure the 244 same number of predictors across models. 245 Nested cross-validations were applied to increase robustness and 246 generalizability of our MARS estimations, with 3 iterations in the outer loop and 247 10 folds per iteration in the inner loop. As mentioned earlier, data processing 248 including PCA (on /g1850 ) and transformation (on /g1851 ) was performed within each 249 iteration of the outer loop and parameters obtained from the training and 250 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 8 validating set were applied to the held-out testing set. Per training fold, a grid 251 search was performed to identify the optimal combination of two parameters: the 252 maximum degree of interactions among terms ( ndegree) and the number of 253 terms retained in the final model ( nprune). Given the available data points and 254 the empirical evidence that 3 rd-degree interactions never benefited model fit 255 using a subset of data, expansion of interaction degree was restricted to 2 (e.g., 256 testing ndegree = 1 and = 2) and nprune up to 5 times of the predictor numbers 257 (i.e., varied between 690 and 695 due to changing numbers of principal 258 components in each fold). Per prediction model on the testing set, we further 259 partitioned data to estimate prediction uncertainty as the noise and model 260 variance of the out-of-fold predicted values over 5 iterations of 50 cross-261 validations, using variance models of MARS (Milborrow, 2015), and an averaged 262 r-squared of 0.19 in contrast to 0.09 from the training models. 263 For each participant, we quantified ∆ BDS as the difference between the 264 predicted BDS (/g1877 /g3548) and the reported symptom sum score ( /g1851 ) while accounting for 265 the prediction uncertainty including irreducible “aleatoric” or noise variance ( /g2026 /g3028 /g2870 ) 266 and model variance of prediction or “epistemic” uncertainty ( /g2026 /g3032 /g2870 ) such that 267 ∆/g1828/g1830/g1845 /g3404 /g3026/g2879/g3052 /g3548 /g3495 /g3097 /g3328/g3118 /g2878 /g3495 /g3097 /g3280/g3118 ).By definition, the resulting ∆ BDS for each participant indexed 268 less depressed or resilient brain patterns if the quantity was negative, and more 269 depressed or vulnerable brain patterns if positive. Here, we focused on 270 individuals with high neural resilience (i.e., top 10 percentile of the negative 271 ∆ BDS) and individuals with high n eural vulnerability (i.e., top 10 percentile of the 272 positive ∆ BDS). This was determined to drive a good balance between sufficient 273 statistical power (i.e., n= 144 with 11 regressors) and high clinical relevance, and 274 to reduce dependency of BDS as a function of original symptom score (i.e., 275 minimize high correlation between ∆/g1828/g1830/g1845 and original symptom score). 276 277 Testing risk factor effects 278 All analyses were conducted for indi viduals with high neural resilience and 279 vulnerability separately. We first conducted partial rank correlation analyses to 280 test the associations between each risk factor and the ∆ BDS while controlling for 281 the reported symptom sum score (i.e., resulting in effects not influenced by the 282 sum score). This is a common practice to deconfound the associative effects 283 driven by chronological age rather than the brain-age delta in the brain age 284 literature (Smith et al. , 2019). False Discover Rate (FDR) corrections were 285 applied to adjust for multiple testing. We further conducted quantile regression 286 analyses to identify unique contributions of individual risk factors, while 287 accounting for the effects of reported symptom sum score. 288 289 290

(289) 291 For each participant, we estimated the BDS using the Multivariate Adaptive 292 Regression Splines (MARS) model ( Figure 1 ). Our model yielded successful 293 prediction of the BDS from the held-out sample via nested cross validations 294 (RMSE=0.90, R 2=0.09), with a comparable effect size to recent brain-wide 295 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 9 association studies that had sufficiently large sample sizes (Dick et al. , 2021; 296 Marek et al. , 2022). Importantly, our MARS model captures information about 297 brain structural and functional features associated with LLD symptoms and thus 298 indicates the disorder manifestation at the neurobiological level (i.e., a proxy of 299 how “depressed” an individual’s brain is). 300 301 302 303 Figure 1. Prediction Models for Brain-depression Score. (A) Multimodal imaging features 304 including cortical thickness, functional connectivity, gray matter volume, task activity and white 305 matter microstructure measures including fractional anisotropy (FA), mean diffusivity (MD) and 306 intracellular volume fraction (ICVF), as well as late-life depressive (LLD) symptom score and 307 nuisance covariates were included into multivariate adaptive regression splines models. (B) In 308 each model, multimodal imaging features ( X) were used to predict symptom sum score ( Y) while 309 controlling for sex, age, in-scanner motion, head size, scanner site, scanner table position and data 310 acquisition dates ( Cov), via nested cross-validations. In the outer loop, full data were partitioned 311 into 3 splits and in the inner loop, each split of training and validating set (TRAIN+VAL) was fed 312 into a 10-fold cross-validation to validate the model. 313 314 315 In this study sample, about 58% (n=1,144) participants appeared to possess a 316 resilient brain as their BDS was lower than their reported depressive symptom 317 scores (i.e., negative ∆ BDS), and 42% showed the opposite pattern with a 318 vulnerable brain (i.e., positive ∆ BDS). Results from partial correlation analyses 319 show that the Townsend deprivation index was negatively correlated with the 320 magnitude of neural resilience after FDR corrections (r=-0.26, FDR corrected 321 p=0.018). Findings from quantile regression analyses further show that this effect 322 is mostly robust in participants with the highest level of neural resilience (e.g., top 323 10 percentile with the largest negative ∆ BDS) when all risk factors were 324 considered in one model (t=3.72, p<0.001; Figure 2). Interestingly, this 325 association with the material deprivation was not observed for individuals with the 326 highest neural vulnerability (i.e., top 10 percentile with the largest positive ∆ BDS). 327 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 10 Neither did we observe this associative effect for the self-reported depressive 328 symptom scores from individuals with th e most resilient or vulnerable neural 329 patterns. These findings together suggested a specific effect of material 330 deprivation on the LLD-related neural resilience. Additionally, we observed a 331 positive association between social is olation and the median quantile magnitude 332 of neural vulnerability with adjustment for all other risk factors (t=2.01, p<0.05). 333 This associative effect was absent in participants showing resilient patterns at the 334 neuroanatomical level. 335 336 337 Figure 2. Impact of Risk factors on Brain-depression Index. (A) The dot-and-whisker plot 338 compares the estimated coefficient and its standard error for each risk factor from separate 339 multiple linear regression models for high neural resilience and vulnerability. (B) Material 340 deprivation was significantly associated with the delta of brain-depression score ( ∆ BDS) only for 341 individuals with high neural resilience. 342 343 344

(312) 345 In this study, we quantified individual-specific neural resilience and vulnerability 346 to LLD using multimodal imaging features and linked known risk factors of LLD to 347 such neural resilience and vulnerability. Our findings provide empirical evidence 348 that risk factors may exert varying impact on the neurobiological manifestation of 349 LLD. Specifically, we found a negative association between material deprivation 350 and the ∆ BDS in individuals with hi gh neural resilience. This result indicated a 351 beneficial effect of sufficient material resources on neural resilience to LLD. 352 Interestingly, on the other hand, more material deprivation was not associated 353 with neural vulnerability, suggesting that insufficient material resources alone 354 may not adequately contribute to the differential patterns of brain structure and 355 function that put individuals at increased risks of depression. Additionally, we 356 observed a positive association between neural vulnerability to LLD and social 357 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 11 isolation, which was absent in participants showing neuroanatomically resilient 358 patterns. These findings are consistent with the observations that older adults 359 with sufficient resources to age successfully are relatively healthy, active, 360 independent and maintain high levels of mental well-being (Terraneo, 2021), and 361 that increases in socioeconomic status decreases the odds for depression 362 (Freeman et al., 2016; Zhou et al., 2021). Furthermore, our results also resonate 363 with previous findings that social isolation and/or loneliness in older adults is 364 associated with increased risk of all-cause mortality (Holt-Lunstad et al., 2015), 365 as well as clinically significant depression and anxiety (Schwarzbach et al., 2014; 366 Taylor et al., 2018; Domènech-Abella et al., 2019; Donovan and Blazer, 2020). 367 It is however, important to note that the BDS was estimated from the symptom 368 sum score in the current study and thus might fail to capture subtleties in 369 neuroanatomical features associated with specific clinical subtypes. It is highly 370 likely that the disorder manifests in a heterogenous manner and the presentation 371 of vulnerability at the neuroanatomical level can vary across individuals. Future 372 studies may consider using individual symptom-level or dimensional measures of 373 LLD to estimate BDS and test risk factor effects on those potential subtypes. 374 In conclusion, our results demonstrate a link between material deprivation and 375 neural resilience, as well as between social isolation and neural vulnerability to 376 LLD. These results are of importance for policy makers as well as the broader 377 society, as they provide evidence that sufficient material resources can improve 378 neural resilience to depression for older adults and that compensational solutions 379 to improve human interactions are in urgent need to offset the potential 380 vulnerability to LLD when in-person social contacts are restricted. 381 382 383 384 Acknowledgments 385 386 This research was performed under UK Biobank application number 47267. This 387 research was supported by the NIH (1 R34 NS118618-01) and the McDonnell 388 Center for Systems Neuroscience. 389 390 391 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 12

392 Alexopoulos, G. S., Schultz, S. K. and Lebowitz, B. D. (2005) ‘Late-Life Depression: A 393 Model for Medical Classification’, Biological Psychiatry. Elsevier, 58(4), p. 283. doi: 394 10.1016/J.BIOPSYCH.2005.04.055. 395 Alfaro-Almagro, F. et al. (2018) ‘Image processing and Quality Control for the first 396 10,000 brain imaging datasets from UK Biobank’, NeuroImage. doi: 397 10.1016/j.neuroimage.2017.10.034. 398 Baecker, L. et al. (2021) ‘Machine learning for brain age prediction: Introduction to 399

and clinical applications’, EBioMedicine. doi: 10.1016/j.ebiom.2021.103600. 400 Blazer, D. G. (2003) ‘Depression in Late Life: Review and Commentary’, The Journals of 401 Gerontology: Series A. Oxford Academic, 58(3), pp. M249–M265. doi: 402 10.1093/GERONA/58.3.M249. 403 Chang-Quan, H. et al. (2010) ‘Education and risk for late life depression: A meta-404 analysis of published literature’, International Journal of Psychiatry in Medicine. doi: 405 10.2190/PM.40.1.i. 406 Chang, S. C. et al. (2016) ‘Risk factors for late-life depression: A prospective cohort 407 study among older women’, Preventive Medicine. doi: 10.1016/j.ypmed.2016.08.014. 408 Christman, S. et al. (2020) ‘Accelerated brain aging predicts impaired cognitive 409 performance and greater disability in geriatric but not midlife adult depression’, 410 Translational Psychiatry. doi: 10.1038/s41398-020-01004-z. 411 Cole, J. H. and Franke, K. (2017) ‘Predicting Age Using Neuroimaging: Innovative Brain 412 Ageing Biomarkers’, Trends in Neurosciences. doi: 10.1016/j.tins.2017.10.001. 413 Cox, S. R. et al. (2019) ‘Associations between vascular risk factors and brain MRI 414 indices in UK Biobank’, European Heart Journal. doi: 10.1093/eurheartj/ehz100. 415 Dick, A. S. et al. (2021) ‘Meaningful associations in the adolescent brain cognitive 416 development study’, NeuroImage. doi: 10.1016/j.neuroimage.2021.118262. 417 Domènech-Abella, J. et al. (2019) ‘Anxiety, depression, loneliness and social network in 418 the elderly: Longitudinal associations from The Irish Longitudinal Study on Ageing 419 (TILDA)’, Journal of Affective Disorders. doi: 10.1016/j.jad.2018.12.043. 420 Donovan, N. J. and Blazer, D. (2020) ‘Social Isolation and Loneliness in Older Adults: 421 Review and Commentary of a National Academies Report’, American Journal of 422 Geriatric Psychiatry. doi: 10.1016/j.jagp.2020.08.005. 423 Dutt, R. K. et al. (2022) ‘Mental health in the UK Biobank: A roadmap to self-report 424 measures and neuroimaging correlates’, Human Brain Mapping. doi: 425 10.1002/hbm.25690. 426 Fan, M. et al. (2020) ‘Sleep patterns, genetic susceptibility, and incident cardiovascular 427 disease: A prospective study of 385 292 UK biobank participants’, European Heart 428 Journal. doi: 10.1093/eurheartj/ehz849. 429 Franke, K. et al. (2010) ‘Estimating the age of healthy subjects from T1-weighted MRI 430 scans using kernel methods: Exploring the influence of various parameters’, 431 NeuroImage. doi: 10.1016/j.neuroimage.2010.01.005. 432 Franke, K. and Gaser, C. (2019) ‘Ten years of brainage as a neuroimaging biomarker of 433 brain aging: What insights have we gained?’, Frontiers in Neurology. doi: 434 10.3389/fneur.2019.00789. 435 Freeman, A. et al. (2016) ‘The role of socio-economic status in depression: Results from 436 the COURAGE (aging survey in Europe)’, BMC Public Health. doi: 10.1186/s12889-016-437 3638-0. 438 Friedman, J. H. (2007) ‘Multivariate Adaptive Regression Splines’, The Annals of 439 Statistics. doi: 10.1214/aos/1176347963. 440 Friedman, J. H. and Roosen, C. B. (1995) ‘An introduction to multivariate adaptive 441 regression splines’, Statistical Methods in Medical Research. doi: 442 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 13 10.1177/096228029500400303. 443 Hepsomali, P. and Groeger, J. A. (2021) ‘Diet, sleep, and mental health: Insights from 444 the uk biobank study’, Nutrients. doi: 10.3390/nu13082573. 445 Holt-Lunstad, J. et al. (2015) ‘Loneliness and Social Isolation as Risk Factors for 446 Mortality: A Meta-Analytic Review’, Perspectives on Psychological Science. doi: 447 10.1177/1745691614568352. 448 Kang, M. et al. (2021) ‘The relationship of lifestyle risk factors and depression in Korean 449 adults: A moderating effect of overall nutritional adequacy’, Nutrients. doi: 450 10.3390/nu13082626. 451 Khubchandani, J. et al. (2016) ‘The Psychometric Properties of PHQ-4 Depression and 452 Anxiety Screening Scale Among College Students’, Archives of Psychiatric Nursing. doi: 453 10.1016/j.apnu.2016.01.014. 454 van Lee, L. et al. (2020) ‘Multiple modifiable lifestyle factors and the risk of perinatal 455 depression during pregnancy: Findings from the GUSTO cohort’, Comprehensive 456 Psychiatry. doi: 10.1016/j.comppsych.2020.152210. 457 Lourida, I. et al. (2019) ‘Association of Lifestyle and Genetic Risk With Incidence of 458 Dementia’, JAMA, 322(5), p. 430. doi: 10.1001/jama.2019.9879. 459 Manning, K. J. and Steffens, D. C. (2018) ‘State of the Science of Neural Systems in 460 Late-Life Depression: Impact on Clinical Presentation and Treatment Outcome’, Journal 461 of the American Geriatrics Society. NIH Public Access, 66(Suppl 1), p. S17. doi: 462 10.1111/JGS.15353. 463 Marek, S. et al. (2022) ‘Reproducible brain-wide association studies require thousands 464 of individuals’, Nature. Springer US, 603(7902), pp. 654–660. doi: 10.1038/s41586-022-465 04492-9. 466 Milborrow, S. (2015) ‘Notes on the earth package [WWW Document]’, URL http://www. 467 milbo. org/doc/earth-notes. pdf, pp. 1–69. 468 Mutz, J., Roscoe, C. J. and Lewis, C. M. (2021) ‘Exploring health in the UK Biobank: 469 associations with sociodemographic characteristics, psychosocial factors, lifestyle and 470 environmental exposures’, BMC Medicine. doi: 10.1186/s12916-021-02097-z. 471 Peterson, R. A. (2021) ‘Finding Optimal Normalizing Transformations via bestNormalize’, 472 R Journal. doi: 10.32614/rj-2021-041. 473 Sarris, J. et al. (2020) ‘Multiple lifestyle factors and depressed mood: a cross-sectional 474 and longitudinal analysis of the UK Biobank (N = 84,860)’, BMC Medicine. doi: 475 10.1186/s12916-020-01813-5. 476 Schulz, R. et al. (2000) ‘Association between depression and mortality in older adults: 477 The Cardiovascular Health study’, Archives of Internal Medicine. doi: 478 10.1001/archinte.160.12.1761. 479 Schwarzbach, M. et al. (2014) ‘Social relations and depression in late life - A systematic 480 review’, International Journal of Geriatric Psychiatry. doi: 10.1002/gps.3971. 481 Smith, D. J. et al. (2013) ‘Prevalence and characteristics of probable major depression 482 and bipolar disorder within UK Biobank: Cross-sectional study of 172,751 participants’, 483 PLoS ONE, 8(11), pp. 1–7. doi: 10.1371/journal.pone.0075362. 484 Smith, S. M. et al. (2019) ‘Estimation of brain age delta from brain imaging’, 485 NeuroImage. doi: 10.1016/j.neuroimage.2019.06.017. 486 Taylor, H. O. et al. (2018) ‘Social Isolation, Depression, and Psychological Distress 487 Among Older Adults’, Journal of Aging and Health. doi: 10.1177/0898264316673511. 488 Terraneo, M. (2021) ‘The Effect of Material and Social Deprivation on Well-Being of 489 Elderly in Europe’, International Journal of Health Services. doi: 490 10.1177/0020731420981856. 491 Townsend, P., Phillimore, P. and Beattie, A. (1988) Health and deprivation: Inequality 492 and the North. Edited by P. Townsend, P. Phillimore, and A. Beattie. London: Croom 493 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint 14 Helm. 494 Yapp, E. et al. (2021) ‘Sex differences in experiences of multiple traumas and mental 495 health problems in the UK Biobank cohort’, Social Psychiatry and Psychiatric 496 Epidemiology. doi: 10.1007/s00127-021-02092-y. 497 Zhou, S. et al. (2021) ‘Socioeconomic status and depressive symptoms in older people 498 with the mediation role of social support: A population-based longitudinal study’, 499 International Journal of Methods in Psychiatric Research. doi: 10.1002/mpr.1894. 500 501 502 503 . CC-BY-NC-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 25, 2022. ; https://doi.org/10.1101/2022.07.25.22277997doi: medRxiv preprint