Cerebellar growth is associated with domain-specific cerebral maturation and socio-linguistic behavior

This repository contains analysis scripts and cerebellar normative models presented in this manuscript:

https://www.nature.com/articles/s41467-026-72940-5

Instructions

If you'd like to reproduce the analyses in the manuscript, you can run the following scripts:

A. Cerebellar segmentation

Scripts in scripts/1_segmentation generate anatomical (ACAPULCO: https://www.sciencedirect.com/science/article/pii/S1053811920303062) and functional (https://github.com/DiedrichsenLab/cerebellar_atlases/tree/master) cerebellar parcellations in individual subjects' space.

• anat/1_run_acapulco.sh: Runs ACAPULCO (via Singularity).
• func/1_fsl_func_warp.sh: Warps MNI-aligned MDTB, functional fusion, and resting-state cerebellar atlases to subject space (via FSL).
• <parcellation_type>/2_<parcellation_type>_native_vols.py: Counts volumes per parcel. Functional parcellations include outlier detection (anatomical ACAPULCO masks are manually corrected when necessary via ITK-SNAP: https://www.itksnap.org/pmwiki/pmwiki.php).

Expected outputs: Dataframe and images of subject-space anatomical and functional parcellations.

Expected runtime: ~15 minutes per atlas per subject (manual corrections: ~3 hours per subject).

B. Normative modeling

Scripts in scripts/2_normative_modeling generate normative models based on the Hierarchical Bayesian Regression algorithm in the PCN toolkit (https://github.com/amarquand/PCNtoolkit). Sex and site were modeled as batch (random) effects.

• 1_qc_euler_yeo17.py: Performs QC based on the Euler index of cerebral cortex surface reconstructions (for consistency, we included the same subjects in cerebral and cerebellar normative models).
• 2_data_prep.py: Prepares and splits the data for normative modeling.
• 3_norm_modeling_hbr.py: Fits parcel-wise linear and 3rd-order b-spline normative models, including model diagnostics and quantile plots.
• 4_loocv.py: Performs model comparison of linear and b-spline models via leave-one-out cross-validation, as well as additional model diagnostics.

Expected outputs: Parcel-wise normative models, model diagnostics, and quantile plots.

Expected runtime: ~3 minutes per parcel per model type (linear or b-spline). This might vary based on computational infrastructure, model algorithm, and batch effects.

C. Associations with cerebral cortex

Scripts in scripts/3_cerebral_associations perform associations of growth trajectories between the cerebellum and the cerebral cortex (Yeo-17 atlas) via regularized regression.

• cerebral_assoc.py: Compares regularization methods, uses the winning method for cerebro-cerebellar associations, performs permutation significance testing, and plots FDR-corrected associations.

Expected outputs: Heatmap and cerebral cortex plots of significant cerebro-cerebellar associations.

Expected runtime: ~2 hours per cerebellar atlas.

D. Associations with behavioral outcomes

Scripts in scripts/4_behavioral_outcomes perform associations of cerebellar growth trajectories and individual differences in behavioral markers.

• 1_behav_pls.py: Performs multivariate association of cerebellar parcels and behavioral outcomes via Partial Least Squares (PLS) analysis. Selects number of components (latent variables), and performs cross-validation, permutation significance testing, and bootstrapping of the PLS model.
• 2_behav_regressions.py: Unpacks the PLS results via univariate association of cerebellar parcels and behavioral outcomes (mass-univariate linear regression).
• 3_behav_cereb_cortex.py: Compares the relative contribution of cerebellar and cerebral (or combined) growth to behavioral outcomes via regularized regression. Compares regularization methods across parcel set combinations, and then compares model performance of the winning method for each parcel set combination across behaviors.

Expected outputs: Plots of cerebellar-behavioral associations.

Expected runtime: ~30 minutes per cerebellar atlas.

Data

This study is based on openly accessible data from the Lifespan Baby Connectome Project (BCP; https://nda.nih.gov/edit_collection.html?id=2848) and the Lifespan 2.0 Human Connectome Project in Development (HCP-D; https://nda.nih.gov/general-query.html?q=query=featured-datasets:HCP%20Aging%20and%20Development). Cerebellar atlases are available here: https://github.com/DiedrichsenLab/cerebellar_atlases/tree/master.

Linear and b-spline normative models for cerebellar anatomical and functional parcels are available in the normative_models folder. Labels for cerebral and cerebellar anatomical and functional parcellations are available in the atlas_labels folder.

Requirements and installation

All Python scripts were executed using Python 3.11.8 (https://www.python.org/downloads/) on macOS (please see associated scripts for required Python libraries). The ACAPULCO container was downloaded from https://gitlab.com/shuohan/acapulco and executed via Singularity (https://docs.sylabs.io/guides/3.0/user-guide/installation.html). Functional cerebellar segmentations were performed via FSL (https://fsl.fmrib.ox.ac.uk/fsl/docs/#/install/index). Normative modeling was implemented via the PCN toolkit version 0.29 (https://github.com/amarquand/PCNtoolkit). To clone this repositiry, run: git clone https://github.com/kmanoli/NormCerebellum.git

The scripts can be run on a standard desktop computer, however, we recommend running these analyses (especially segmentation algorithms and normative models) on a cluster to parallelize computations.