Network communication models improve the behavioral and functional predictive utility of the human structural connectome

The connectome provides the structural substrate facilitating communication between brain regions. We aimed to establish whether accounting for polysynaptic communication in structural connectomes would improve prediction of interindividual variation in behavior as well as increase structure-function coupling strength. Connectomes were mapped for 889 healthy adults participating in the Human Connectome Project. To account for polysynaptic signaling, connectomes were transformed into communication matrices for each of 15 different network communication models. Communication matrices were (a) used to perform predictions of five data-driven behavioral dimensions and (b) correlated to resting-state functional connectivity (FC). While FC was the most accurate predictor of behavior, communication models, in particular communicability and navigation, improved the performance of structural connectomes. Communication also strengthened structure-function coupling, with the navigation and shortest paths models leading to 35–65% increases in association strength with FC. We combined behavioral and functional results into a single ranking that provides insight into which communication models may more faithfully recapitulate underlying neural signaling patterns. Comparing results across multiple connectome mapping pipelines suggested that modeling polysynaptic communication is particularly beneficial in sparse high-resolution connectomes. We conclude that network communication models can augment the functional and behavioral predictive utility of the human structural connectome.

racy and the t-statistics of pairwise comparisons of models to their null counterparts). In addition, when comparing accuracies across all possible pairs of empirical and null predictors, we found that most empirical models significantly outperformed most null models. This is evidenced by the predominantly warm-and cool-colored off-diagonal blocks in the effect size matrix shown in Figure S4c. Together, these findings provide further evidence that network communication models are capable of predicting interindividual variation in human behavior, and that observed differences in empirical accuracies reflect meaningful distinctions in the predictive utility of different models. Disruptions to SC topology, even when preserving degree distribution, led to significant reductions in the predictive utility of communication models. This suggests that the relationship between network communication models and behavior is contingent on high-order organizational features of the human structural connectome. The behavioral predictive utility analyses performed in this paper were based on distributions comprising 100 estimates of prediction accuracy, computed by performing 10 repetitions of 10-fold cross validation. In this section, we evaluate whether our results are stable when considering single repetitions of the 10-fold cross validation. In other words, we aimed to compare communication models based on distributions comprising 10 estimates of prediction accuracy, one from each cross validation fold. Our goal was to determine whether statistical differences between predictors (or the lack thereof) reported when considering 100 accuracy estimates would be observed when only taking into account the accuracies from each individual cross validation repetition.
To this end, we separately recomputed the results shown in Fig. 3 for each of the 10 repetitions of the cross validation process. We derived 10 rankings of predictors (as in Fig. 3a) and 10 17 × 17 effect size matrices of pairwise statistical comparisons (as in Fig. 3b). We also computed a ranking of predictors and effect size matrix for the accuracy distributions obtained when averaging results across the cross validation repetitions. This resulted in a set of 11 predictor rankings and effect size Seguin, C., Tian, Y. & Zalesky, A. (2020). Supporting information for "Network communication models improve the behavioral and functional predictive utility of the human structural connectome." Network Neuroscience, 4(4), 980-1006. https://doi.org/10.1162/netn_a_00161 matrices.
To facilitate comparisons, effect size matrices were binarized. Entries equal to 1 and 0 denoted, respectively, evidence for significant and non-significant differences between accuracy distributions (Bonferroni-corrected threshold for multiple comparisons α = 0.05/(17 × 16/2) = 3.67 × 10 −4 ). This categorization of statistical test outcomes allowed us to compare our original results to the ones from individual cross validation repetitions by means of Fisher's exact tests. Rejection of the null hypothesis provides evidence of nonrandom associations between significance outcomes computed on 100 accuracy estimates to those based on 10 accuracy estimates.
Results are shown in Table S1. We found that the average prediction accuracies shown in Fig. 3a were strongly correlated to those obtained for individual repetitions of the cross validation process. In addition, for each individual repetition, the outcome of pairwise statistical comparisons between predictors was significantly associated to the original outcomes shown in Fig. 3b. We have characterized the predictive utility of connectivity measures and communication models by considering a pooled prediction accuracy between the cognition and tobacco use dimensions. As we have seen, these behavioral dimensions led to the most accurate and consistent predictions in our sample. Considering the average prediction accuracy across the two traits facilitated the comparison of communication models by providing us with a single score on which to evaluate prediction accuracy. However, this approach may potentially obscure nuanced phenotype-specific relationships between brain and behavior. Indeed, we observed no correlation between the prediction accuracies obtained for cognition and tobacco use (Spearman rank correlation p = 0.49, 0.69 for lasso and NBS, respectively). Therefore, in this section, we sought to separately examine behavioral predictions for the cognition and tobacco use dimensions.
We observed that the outstanding performance of FC was mostly due to the cognition dimension ( Fig. S5 and S8; this can also be observed in Fig. 2c,d). This is in line with several findings on the relationship between cognitive processes and the architecture of functional networks [2]. Interestingly, while FC still yielded top-ranking predictions of tobacco use, several communication models showed comparable predictive utility in this dimension ( Fig. S6 and S9; again evident in Fig. 2c,d). For instance, navigation was the best predictor of tobacco use, with binary and distance navigation occupying the first positions under lasso and NBS, respectively. Although none of the communication models statistically outperform FC, this points towards a tighter margin of difference between the utility of structural and functional mea-sures in predicting behavioral phenotypes not directly related to cognition.
Another interesting observation was that weighted search information was the best communication model in predicting cognition, but showed near bottom-ranking predictive utility of tobacco use, which resulted in a moderate performance when combining predictions from both behavioral dimensions. Along similar lines, we observed that rankings of connection weight definitions were different across behavioral dimensions and prediction methods, painting an unclear picture of what weighting schemes best contribute to predict human behavior from structural connectomes.
Therefore, collectively, our results did not point towards a single communication model as the best predictor of human behavior. Despite the overall good performance of communicability and navigation, our observations indicate that different communication models may be better suited to predict different behavioral dimensions, possibly suggesting the presence of behavior-specific signaling mechanisms in the human brain. Importantly, across the multiple analyses performed, our results consistently suggested that network communication models, in particular communicability and navigation, improve the behavioral predictive utility of the human connectome.

Note 4: Additional analyses of structure-function coupling
We further examined structure-function relationships by stratifying FC predictions across anatomically connected and unconnected regions pairs (Fig. S11a). In accordance to previous work [3,4], associations to FC were stronger for connected regions. Despite these changes in association strength, rankings of communication models in terms of FC predictions were consistent across the scenarios explored (Spearman rank correlation r = 0.90, 0.68, 0.85 and p = 8 × 10 −7 , 3 × 10 −3 , 0 between connected and all, all and unconnected, and connected and unconnected region pairs, respectively). Interestingly, certain communication models outperformed SC for connected regions, suggesting that indirected polysynaptic signaling may be relevant for communication even in the presence of direct anatomical links.
Estimates of structure-function coupling depend on accurate reconstruction of structural connectomes. However, white matter tractography is prone to a number of known biases, of which the underestimation of interhemispheric connections is an important concern [5]. To attenuate this issue, past studies have focused on intrahemispheric characterization of structure-function coupling [6,7]. Focusing on the right hemisphere, we found that, for all communication models, FC associations were stronger compared to whole-brain estimates (Fig. S11b). Importantly, the functional predictive utility ranking of communication models remained consistent (Spearman rank correlation r = 0.93 and p = 0 between whole-brain 3 and right hemisphere rankings). This suggests that our analyses may provide a meaningful ranking of signaling strategies despite shortcomings in connectome mapping techniques. FIG. S1. Comparison between behavioral prediction accuracies computed using Pearson correlation and mean square error evaluated using Spearman rank correlation (Lasso regression, N = 360 thresholded connectomes). The least squares line is shown in black.