State-Unspecific Modes of Whole-Brain Functional Connectivity Predict Intelligence and Life Outcomes

Recent functional magnetic resonance imaging (fMRI) studies have increasingly revealed potential neural substrates of individual differences in diverse types of brain function and dysfunction. Although most previous studies have been inherently limited to state-specific characterizations of related brain networks and their functions, several recent studies have examined the potential state-unspecific nature of functional brain networks, such as their global similarities across different experimental conditions (i.e., states) including both task and rest. However, no previous studies have carried out direct, systematic characterizations of state-unspecific brain networks, or their functional implications. Here, we quantitatively identified several modes of state-unspecific individual variation in whole-brain functional connectivity patterns, called “Common Neural Modes (CNMs)”, from a large fMRI dataset including eight task/rest states, obtained from the Human Connectome Project. Furthermore, we tested how CNMs account for variability in individual behavioral measures. The results revealed that three CNMs were robustly extracted under various different preprocessing conditions. Each of these CNMs was significantly correlated with different aspects of behavioral measures of both fluid and crystalized intelligence. The three CNMs were also able to predict several life outcomes, such as income and life satisfaction, achieving the highest performance when combined with behavioral intelligence measures as inputs. Our findings highlight the importance of state-unspecific brain networks to characterize fundamental individual variation.

of state-unspecific brain networks, or their functional implications. Here, we quantitatively 23 identified several modes of state-unspecific individual variation in whole-brain functional 24 connectivity patterns, called "Common Neural Modes (CNMs)", from a large fMRI dataset 25 including eight task/rest states, obtained from the Human Connectome Project. Furthermore, 26 we tested how CNMs account for variability in individual behavioral measures. The results 27 revealed that three CNMs were robustly extracted under various different preprocessing 28 conditions. Each of these CNMs was significantly correlated with different aspects of 29 behavioral measures of both fluid and crystalized intelligence. The three CNMs were also 30 able to predict several life outcomes, such as income and life satisfaction, achieving the 31 highest performance when combined with behavioral intelligence measures as inputs. Our 32 as represented by the default mode network (DMN) (Raichle, 2015). A wide variety of 48 individual differences in our cognition and behavior have been associated with the 49 characteristics of FC patterns and networks in the brain, including cognitive abilities (Finn et  These previous studies have investigated the relationship between individual differences and 55 brain networks while a person is experiencing a specific state. In particular, recent research 56 has intensively focused on the resting state, as it potentially reflects many types of individual 57 differences and can be measured easily (Dubois and Adolphs, 2016). The present study is 58 directly inspired by Smith et al. (2015), who revealed, in a data-driven manner, that a small 59 number of linear factors underlying individuals' whole-brain resting-state FC patterns 60 ("neural modes") can explain diverse ranges of individual differences simultaneously (Smith 61 et al., 2015). However, despite their apparent connections with behavior, the brain networks 62 and neural modes examined in these previous studies, as well as their relations to individual 63 differences, are inherently state-specific; thus, it is unclear whether these findings generalize 64 across states, indicating basic traits of individuals. Geerligs et al. (2015) demonstrated that 65 the relationship between individual differences and FC patterns may substantially change 66 across different states, including both rest and task (Geerligs et al., 2015).  Furthermore, Tavor et al. (2016) revealed that an individual's brain activity during a task state 75 can be predicted from their resting-state FC patterns. These findings clearly suggest a 76 potential state-unspecific aspect of brain networks. Unfortunately, however, no previous 77 studies have explicitly identified these state-unspecific brain networks, or quantitatively 78 investigated their relationship to individual differences in behavior. 79

80
In the present study, we conducted, for the first time, a quantitative characterization of state-81 unspecific brain networks and investigated its connection with inter-individual variability in 82 behavior. Our approach combined the large-scale database of the Human Connectome Project 83 with a sophisticated machine learning technique. Specifically, we applied multiset canonical 84 correlation analysis (M-CCA) to the FC matrices obtained from eight states, including the 85 resting state (Kettenring, 1971;Vía et al., 2007). The obtained components uniquely 86 characterize individuals' FC patterns that are common across different states, which we refer 87 to as "Common Neural Modes (CNMs)". We demonstrated that several CNMs could be 88 robustly extracted from whole-brain FC patterns. These CNMs were then found to be 89

MRI parameters 105
The fMRI data were acquired using a protocol with advanced multiband sequences. Whole-106 brain echo-planar scans were acquired with a 32-channel head coil on a modified 3T Siemens 107 Skyra with repetition time = 720 ms, echo time = 33.1 ms, flip angle = 52°, bandwidth 2,290 108 Hz/Px, in-plane field of view = 208 ×180 mm, 72 slices, 2.0 mm isotropic voxels, with a 109 multiband acceleration factor of 8 (Uǧurbil et al., 2013). Data were collected over 2 days. On 110 each day, 28 min of rest (eyes open with fixation) fMRI data across two runs were collected 111 (two runs, 56 min in total, per day), followed by 30 min of task-fMRI data collection (60 min 112 in total, per day). Each of the seven task-fMRI was completed over two consecutive fMRI 113 runs. Three task-fMRI (working memory, reward learning, and motor responses) data were 114 collected on the first day. The other four task-fMRI (emotion perception, language processing, 115 relational reasoning, and social cognition) data were collected on the second day. More 116 details about the fMRI collection method were described in previous studies (Barch et al.,  The seven task-fMRI paradigms were selected to activate different neural circuitry that 121 supports broad cognitive functions, and included emotion perception, reward learning, 122 language processing, motor responses, relational reasoning, social cognition, and working 123 memory (Barch et al., 2013;Cole et al., 2016). Briefly, the emotion task involved matching 124 fearful or angry faces to a target face. The reward learning task involved a gambling task 125 involving monetary rewards and losses. The language task involved auditory stimuli 126 consisting of narrative stories and math problems, along with questions to be answered 127 regarding the prior auditory stimuli. The motor task involved movement of the hands, tongue 128 and feet. The relational reasoning task involved higher-order cognitive reasoning regarding 129 relations among features of presented shape stimuli. The social cognition (theory of mind) 130 task used short video clips of moving shapes that interacted in some way or moved randomly, 131 with subjects making decisions about whether the shapes had social interactions. The and CNMs are defined as the average of Xkwk for k = 1,…,M (Fig. 1c). 197 To reduce redundancy among FCs, the dimensionalities of the data matrices were reduced in 198 advance using principal components analysis (PCA). The numbers of principal components 199 were varied between 10, 50 and 100 for calculating CNMs, and the numbers of CNMs were 200 also varied between 10, 50 and 100, respectively. The significance of the pairwise canonical 201 correlations was investigated using a permutation test for individual CNMs. We first shuffled 202 subject labels of all Xk, then conducted M-CCA. We ran these analyses 1,000 times and 203 obtained 1,000 instances of estimated wk. We then took the average of the absolute correlation 204 coefficients between all pairs among Xkwk for each random dataset. Finally, we calculated 205 the statistical significance by comparing the true averaged value of the correlation coefficient 206 with those obtained from shuffled datasets. 207 208

Relationship between CNMs and cognitive measures 209
To analyze how CNMs were associated with individual differences in behavior, we calculated

Effects of the number of states used to identify CNMs 240
We investigated the effects of the number of states used to identify CNMs on prediction 241 accuracy. Specifically, we conducted the same prediction analyses as above, but here we used 242 a smaller number of states for constructing the CNMs. We varied the number of states for 243 constructing the CNMs from 2 to 8. We calculated all possible combinations for each case. 244 For example, we calculated 28 CNMs (=7 8 2 :), then constructed prediction models for all 245 CNMs, when we estimated the prediction accuracy of two states. 246 247

Interpretation of CNMs 248
To facilitate the characterization of the biological substrates of the CNMs, we summarized 249 the FC patterns that were correlated with first, second and third CNMs. We focused on these 250 three CNMs because they had been robustly extracted by M-CCA. First, we averaged every 251 FC value over all eight states. We then calculated Pearson's correlation coefficients between 252 three CNMs and each averaged FC. The 268 ROIs were then grouped into eight 253 representative macroscale networks (e.g., DMN) defined functionally in a previous study 254 intelligence (Cattell, 1963;Gottfredson, 1997). Finally, CNM3 was correlated with both fluid 281 intelligence and language-related scores. Note that we confirmed that the correlations 282

Prediction of life outcomes using CNMs 296
We next investigated whether CNMs could predict life outcomes, complementing 297 conventional behavioral tests (i.e., measures of fluid intelligence). 298 Figs. 3a, 3b and 3c show that predicting with CNMs alone achieved significant predictive 299 value (P < 10 -4 for income and number of years of education; P < 2.00 × 10 -4 for life 300 satisfaction; 10,000 times permutation test). The correlation coefficient (r) was slightly 301 higher than that with fluid intelligence alone for income and life satisfaction, but worse for 302 years of education. Combining both the CNMs and fluid intelligence yielded the highest 303 performance in every case (P < 10 -4 for all income and years of education; P < 2.00 × 10 -4 304 for life satisfaction; 10,000 times permutation test).

Effects of the number of states used for the CNMs 313
We further investigated the effects of the number of states used for identifying the CNMs on 314 the prediction accuracy. Fig. 4a, 4b and 4c show the prediction accuracies using the CNMs 315 with different numbers of states. These figures indicate that the more states we used, the 316 greater accuracy we were able to achieve for predicting life outcomes. We constructed linear 317 regression models, and found that the effects of the number of states were significant for all 318 models (P = 8.15 × 10 -13 for income; P = 5.71 × 10 -13 for life satisfaction; P = 0.007 for years 319 of education). intelligence. We also found that the more states we used to identify CNMs, the higher 365 accuracy we were able to achieve when predicting life outcomes. The FCs constituting those 366 CNMs were widely distributed throughout the brain rather than being locally constrained. 367 368 Three CNMs were robustly extracted by M-CCA, which correlated significantly with 369 representative intelligence measures (Fig. 2). Intelligence measures are related to a wide 370 range of cognitive functions and predict broad social outcomes such as educational 371 achievement, job performance, health, and longevity (Cattell, 1963;Colom et al., 2010;372 Gottfredson, 1997). Therefore, the relationships between the CNMs and these measures are 373 intuitive to understand. It is also noteworthy that each CNM correlated with a different 374 dimension of intelligence. That is, CNM1 and CNM3 correlated with fluid intelligence, while 375 the CNM2 correlated with crystalized intelligence. This suggests that these CNMs may have 376 different biological substrates (Fig. 5). Importantly, the CNMs were derived in a fully data-377 driven, cross-validated manner. The relationship between CNMs and intelligence measures 378 was thus non-trivial. Although our study was inspired by the "positive-negative" neural Importantly, using a greater number of states to identify CNMs enabled us to achieve greater 397 prediction accuracy (Fig. 4). This indicates that CNMs were more reliably extracted when 398 considering a greater number of behavioral states. Indeed, the correlation between 399 representative intelligence measures and first principal components derived from each single 400 state were lower than those of the CNMs. Our findings suggest that contrasting many 401 different states, rather than considering any single (typically resting) state, can more reliably 402 identify the neural modes that are able to predict diverse types of individual differences. 403 404 Although all three CNMs were related to the subcortical-networks and motor networks, we 405 observed different trends among them in terms of the related canonical networks (Fig. 5). 406 CNM1, CNM2 and CNM3 were related to the medial frontal network, frontoparietal network, 407 and both networks, respectively. This finding is of interest because CNM1 and CNM2 408 captured different aspects of intelligence (fluid and crystalized intelligence, respectively) 409 while CNM3 was related to both. We also observed that brain regions contributing to all 410 CNMs were widely distributed rather than locally restricted. This is consistent with a 411