High-resolution connectomic fingerprints: Mapping neural identity and behavior

Connectomes are typically mapped at low resolution based on a specific brain parcellation atlas. Here, we investigate high-resolution connectomes independent of any atlas, propose new methodologies to facilitate their mapping and demonstrate their utility in predicting behavior and identifying individuals. Using structural, functional and diffusion-weighted MRI acquired in 1000 healthy adults, we aimed to map the cortical correlates of identity and behavior at ultra-high spatial resolution. Using methods based on sparse matrix representations, we propose a computationally feasible high-resolution connectomic approach that improves neural fingerprinting and behavior prediction. Using this high-resolution approach, we find that the multimodal cortical gradients of individual uniqueness reside in the association cortices. Furthermore, our analyses identified a striking dichotomy between the facets of a person's neural identity that best predict their behavior and cognition, compared to those that best differentiate them from other individuals. Functional connectivity was one of the most accurate predictors of behavior, yet resided among the weakest differentiators of identity; whereas the converse was found for morphological properties, such as cortical curvature. This study provides new insights into the neural basis of personal identity and new tools to facilitate ultra-high-resolution connectomics.


Introduction
, function [10-12], 27 and white matter microstructure [13,14]. 28 The features of a person's neural fingerprint that 29 make an individual unique and identifiable are not 30 well understood. While the friction ridges on a human 31 finger provide approximately equal utility in identify-32 ing an individual, this is not the case for the brain,

138
Given that cortical structure is more akin to a true 139 fingerprint than functional brain properties, we hy-140 pothesized that brain structure and white matter con-  Fig. 1. Schema of study design, methodology, and neuroimaging modalities and properties. (A) Using structural, functional and diffusion-weighted magnetic resonance imaging (MRI), whole-brain structural and functional networks (top), as well as cortical maps of myelination, sulcal depth, cortical thickness and curvature (bottom), were mapped for each of 1000 individuals. (B) Each cortical map and network was represented at the resolution of an established cortical atlas (180 regions in each cerebral hemisphere) and at the high resolution of the underlying cortical mesh (fsLR-32k mesh with ∼32,000 cortical vertices in each cerebral hemisphere). Each low-and high-resolution cortical map and network yielded a candidate neural fingerprint for each individual. Measures of similarity in the form of matrices (Stest−test, Stest−retest) and maps (Ψ) were computed between individual fingerprints and used to: i) evaluate the accuracy with which each individual could be identified and differentiated from the group, otherwise known as neural fingerprinting; ii) identify regional loci that most strongly differentiated individuals; and, iii) predict interindividual variation in measures of behavior and cognition. and behavior prediction. Fig. 1  into CIFTI scalars in the PostFreeSurfer step [30]. 229 The latter step also provided surface myelin maps      The selected threshold constructed a sparse fully con- In this approach, uniqueness U was the residual re-482 sulting from the regression of intra-individual dissim- i.e. the medial wall. The actual correspondence be- The Cohen's d effect size difference in the mean of 536 these two distributions was denoted as identifiability: Where µ intra and µ inter are the mean of the two 538 distributions. Thus, the term |µ intra − µ inter | mea- Where Φ is the cumulative density function. There-          Fig. S1).

869
We conclude that high-resolution networks and cor- Asterisks positioned between pairs of violet and pink violins indicate a significant difference in the prediction performance of the low-and high-resolution connectomes and cortical maps (p < 0.05). Asterisks positioned between neighboring neuroimaging properties indicate a significant difference in the performance between these properties (high-resolution only). Performance was quantified by the Pearson's correlation between the predictions and the actual behavior measures. Significance testing was performed using the Wilcoxon signed rank test over 100 splits of paired prediction performances. Each split comprised of 300 training and 100 testing samples randomly selected from familialy unrelated individuals. The number of consecutive asterisks quantify effect sizes ( * * * : large effect size (d > 0.8), * * : medium effect size (d > 0.5), * : small effect size (d > 0.2)). The dashed red lines denote the 95% percentiles of the null distribution resembling a prediction by chance. Null distributions are based on the performances resulted from randomly sampled covariance matrices. The null hypothesis of an absence of a brain-behavior relation could be rejected for each distribution mean exceeding the dashed line. Age, sex, and motion confounds were regressed out of behaviors before fitting the prediction models. For age predictions, only sex and motion confounds were regressed. mapped here suggests that structural connectivity 1052 provides a potent predictor and should be investi-1053 gated in future brain-behavior studies. Moreover, 1054 cortical curvature and properties of brain morphol-1055 ogy should be considered in future neural fingerprint-1056 ing studies. A key finding of the present study was 1057 that high-resolution connectomes universally outper-1058 formed their atlas-based counterparts with respect 1059 to identifying individuals from a group and behav-1060 ior prediction. They also enabled more precise spa-1061 tial localization of the cortical loci that most strongly 1062 expressed neural identity. Although high-resolution 1063 connectome mapping imposes a substantial compu-1064 tational burden, we have developed new tools and 1065 methodologies for researchers to efficiently map high-1066 resolution connectomes. We found that cortical morphology and connectiv-1069 ity within the association cortices most strongly ex-1070 pressed an individual's unique neural identity. In con-1071 trast, unimodal and sensory regions shared relatively 1072 common patterns of cortical structure and connectiv-1073 ity between individuals. The distinction in unique-1074 ness between unimodal and association cortices was 1075 replicated across multiple modalities and neuroimag-1076 ing properties, including cortical thickness, curvature, 1077 sulcal depth, and connectivity. This was further con-1078 firmed by finding significant spatial correlations be-1079 tween cortical maps of uniqueness and a previously 1080 established unimodal-to-transmodal gradient in func-1081 tional connectivity [59].   neural fingerprints and behavioral predictors. All 1168 other neuroimaging properties were consistent with 1169 the identification-prediction dichotomy. For example, 1170 functional connectivity was a comparatively accurate 1171 behavioral predictor, but did not accurately identify 1172 individuals compared to other properties. On the 1173 other hand, measures of cortical structure, especially 1174 curvature, were significantly better at identifying in-1175 dividuals from a database, yet performed significantly 1176 poorer in prediction of behavior.

1177
The identification-prediction dichotomy implies 1178 that neural identity is decoupled from the cortical 1179 properties that most explain an individual's behav-1180 ior and cognition. Identity and behavioral associ-1181 ations are by no means orthogonal characteristics, 1182 but our findings point to the difference between what 1183 makes a property good at identification and what as-1184 sociates it with behavior. Indirect inference on behav-1185 ior from neural identification should thus be avoided. 1186 We believe the optimal identification-prediction per-1187 formance of structural connectivity, especially in the 1188 higher resolution, is because it contains complemen-1189 tary information from function [74-78] and structure 1190 [79, 80] that reflect unique dimensions of population 1191 heterogeneity. For all neuroimaging properties evaluated in this 1194 study, we found that high-resolution cortical maps 1195 and connectomes generally outperformed their atlas-1196 based counterparts with respect to identification and 1197 overall behavior prediction performance. This was ex-1198 ceptionally evident for connectomic properties. Neu-1199 roimaging properties are often averaged across vox-1200 els/vertices comprising a broad atlas before further 1201 analyses are undertaken [22,81]. Furthermore, con-1202 nectomes are typically mapped such that regions com-1203 prising a cortical atlas serve as network nodes, even 1204 though streamline endpoints can be resolved at much 1205 finer resolutions. Our results suggest that maintain-1206 ing cortical maps in the resolution of voxels/vertices 1207 and mapping connectomes at the voxel/vertex resolu-1208 tion can capture greater inter-individual heterogene-1209 ity in cortical structure and function, compared to 1210 imposing a regional cortical atlas. While imposing 1211 an atlas may be unavoidable in some circumstance in 1212 order to reduce the computational and storage burden 1213 of high-resolution maps, we developed openly avail-1214 able tools and pipelines to overcome some of the com-1215 putational challenges of high-resolution connectomics 1216 (Data Availability).

1217
The difference in identification and prediction per-1218 formance between voxel/vertex and atlas resolutions 1219 was greatest for structural connectomes, although 1220 this difference was also substantial for anatomical 1221 properties (Fig. 6). It is important to emphasize that 1222 the analysis of voxel/vertex resolution maps of corti-1223 cal thickness and other structural brain properties is 1224 not new and is indeed the default resolution for many 1225 neuroimaging analysis pipelines [82][83][84][85]. In contrast, 1226 despite the advantages [24], whole-brain connectomes are rarely mapped at the resolution of individual vox-1228 els/vertices due to the unwieldy dimensionality of the 1229 resulting connectivity matrices.  Several limitations require consideration. First, im-1290 plementations of behavior prediction models are in-1291 herently influenced by the vast degrees of freedom 1292 in model selection, confound regression, and evalua-1293 tion metric. We tried to alleviate such possible bi-1294 ases by reporting a consensus that combines various 1295 models to form a unanimous finding [57]. Further-1296 more, as motion is known to impact behavior predic-1297 tions of functional imaging properties [88, 89], motion 1298 was regressed out of the behavioral measures prior to 1299 prediction. Supplemental prediction analyses without 1300 control for the effects of motion yielded comparable 1301 results (Supplemental Information S.4.1). Additional 1302 supplemental analyses also demonstrate the minimal 1303 impact of the evaluation metric on the reported per-1304 formance (Supplemental Information S.4.2).

1305
Moreover, the relative rankings reported were 1306 largely replicated in a multitude of alternate design 1307 strategies as part of supplementary analyses (Supple-1308 mental Information S.4.1, S.4.2, S.4.3). Mainly, the 1309 higher prediction accuracy of the neuroimaging prop-1310 erties of connectivity (compared to morphology), es-1311 pecially in the higher resolution, was replicated in all 1312 supplemental analyses. However, the exact order of 1313 the rankings slightly varied depending on the behav-1314 ior prediction analysis design. Additionally, our mul-1315 timodal approach mainly focused on the cerebral cor-1316 tex as most properties were only available on the cor-1317 tical surface. Hence, with the exception of functional 1318 connectivity, this study did not include the subcortex. 1319 Finally, a possible confound inherent to the cortical 1320 maps of individual uniqueness is the impact of image 1321 misregistration. This confound impacts regions prone 1322 to poor registration such as the intraparietal sulcus, 1323 the cingulate sulcus, and the superior temporal sul-1324 cus [33,90]. The use of surface based registration, as 1325 used in the present study, in contrast to volumetric 1326 registration is known to reduce the effects of misreg-1327 istration on intersubject variations [34, 91, 92] and 1328 therefore minimizes the registration confounds.  We found converging evidence from three different 1331 MRI modalities suggesting that an individual's neu-1332 ral identity is most strongly expressed within the as-1333 sociation cortices. The structure, morphology and 1334 connectivity of these phylogenetically-evolved cortical 1335 areas were shown to be unique to an individual and 1336 provided potent predictors of inter-individual varia-1337 tion in behavior and cognitive performance. More-1338 over, we found a dichotomy between the specific cor-1339 tical properties that enabled accurate identification 1340 of an individual from a group and those that best ex-1341 plained inter-individual variation in behavioral mea-1342 sures. Structural connectivity was the only feature 1343 that performed accurately for both these tasks. We 1344 also found that high-resolution connectomes and cor-