A task-general connectivity model reveals variation in convergence of cortical inputs to functional regions of the cerebellum

  1. Maedbh King  Is a corresponding author
  2. Ladan Shahshahani
  3. Richard B Ivry
  4. Jörn Diedrichsen
  1. Department of Psychology, University of California, Berkeley, United States
  2. Western Institute for Neuroscience, Western University, Canada
  3. Helen Wills Neuroscience Institute, University of California, Berkeley, United States
  4. Department of Statistical and Actuarial Sciences, Western University, Canada
  5. Department of Computer Science, Western University, London, Canada
7 figures and 1 additional file

Figures

Connectivity model training and evaluation.

(A) Models were trained on Task Set A (session 1 and 2) of the multi-domain task battery (MDTB; King et al., 2019). Model hyperparameters were tuned using fourfold cross validation. Three types of …

Figure 2 with 4 supplements
Performance of cortico-cerebellar connectivity models.

(A) Predictive accuracy (Pearson correlation) of the Ridge, Lasso, and WTA regression models for the test data of Task Set B (B) Predictive accuracy normalized to the noise ceiling based on …

Figure 2—source data 1

Source data file contains model evaluation predictive accuracies for each of the three methods (Ridge, Lasso, Winner-Take-All), and for each parcellation (80-1848).

https://cdn.elifesciences.org/articles/81511/elife-81511-fig2-data1-v1.csv
Figure 2—figure supplement 1
Model recovery simulations demonstrate the ability to identify different forms of cortico-cerebellar connectivity.

Predictive accuracy for Ridge, Lasso, and WTA models trained and tested on simulated data generated using (A) one-to-one connectivity with each cerebellar voxel connected only to one randomly …

Figure 2—figure supplement 2
Noise ceiling for Ridge model.

Expected predictive accuracy assuming that the fitted Ridge model (1848 parcels) reflects the true cortico-cerebellar connectivity. The noise ceiling takes into account the reliability of the …

Figure 2—figure supplement 3
Predictive accuracy for Ridge and WTA models using functional cortical parcellations.

[(Yeo et al., 2011), (Schaefer et al., 2018), (Arslan et al., 2015), (Fan et al., 2016), (Gordon et al., 2016), (Shen et al., 2013), denoted by first author]. Predictive performance is normalized to …

Figure 2—figure supplement 4
Hyper-parameter tuning for connectivity models.

Predictive accuracy for Ridge (A) and Lasso (B) models using different regularization coefficients (log-lambda values) across five levels of granularity, and cross-validated over 4 folds of the …

Figure 2—figure supplement 4—source data 1

Source data file contains predictive accuracies for each hyperparameter (-2, 0, 2, 4, 6, 8, 10) of the Lasso model, and for each parcellation (80-1848).

https://cdn.elifesciences.org/articles/81511/elife-81511-fig2-figsupp4-data1-v1.csv
Figure 2—figure supplement 4—source data 2

Source data file contains predictive accuracies for each hyperparameter (-5, -4, -3, -2, -1) of the Ridge model, and for each parcellation (80-1848).

https://cdn.elifesciences.org/articles/81511/elife-81511-fig2-figsupp4-data2-v1.csv
Figure 3 with 3 supplements
Cortical connectivity weight maps for the Ridge regression model with 1848 cortical parcels for each of 10 functional cerebellar regions.

Each region is denoted by the most important functional term (King et al., 2019). Results are averaged across participants. Regression weights are in arbitrary units. See Figure 3—animation 1 for a …

Figure 3—figure supplement 1
Cortical connectivity weight maps for the Lasso model with 1848 cortical regions.

As in Figure 4, the results are averaged across individuals for each of the 10 functional regions defined on the MDTB data set, with each region denoted by the most important functional term. …

Figure 3—animation 1
Animated cortico-cerebellar connectivity maps for the Ridge regression model.
Figure 3—animation 2
Animated cortico-cerebellar connectivity maps for the Lasso regression model.
Figure 4 with 2 supplements
Cortico-cerebellar convergence measures using the Lasso model.

(A) Map of the cerebellum showing percentage of cortical parcels with non-zero weights for the Lasso model (n=80 parcels). (B) Percentage of parcels with non-zero weights for functional subregions …

Figure 4—figure supplement 1
Cortico-cerebellar convergence measures using the Ridge model.

(A) Map of the cerebellum showing percentage of cortical parcels with coefficients for the Ridge model (n=80 parcels) above threshold (see Methods). (B) Percentage of parcels with weights above …

Figure 4—figure supplement 2
Percentage of cortical surface across levels of granularity for Lasso regression model.

Calculation is performed as in Figure 3B. The results are averaged across the two hand regions of the cerebellum (MDTB functional parcellation, regions 1 and 2), and regions related to narrative and …

Generalization to new dataset.

Models of cortico-cerebellar connectivity are tested in a new experiment. Each model is tested across different levels of cortical granularity. Predictive accuracy is the Pearson correlation between …

Author response image 1
Author response image 2

Additional files

Download links