Autism Brain Variability

Analysis code for a study testing landscape theory of brain organization in autism using the Human Connectome Project (HCP): does higher autism polygenic score (PGS) predict greater variability in brain network organisation, while preserving global efficiency?

The pipeline integrates three data modalities:

1. Phenotypic — HCP behavioural measures, reduced via confirmatory factor analysis to a social-cognition factor.
2. Genetic — genotype QC, PGS calculation from the Grove et al. (2019) iPSYCH-PGC autism GWAS, and BLUP extension to the full (related) sample.
3. fMRI — resting-state connectivity, consensus community detection, and graph-theoretical landscape analysis (modularity, global efficiency, and their variance as a function of PGS).

Core hypotheses

• Brain–behaviour link. Graph measures of integration and segregation predict social functioning.
• Compensation diversity. High-PGS individuals show greater variability in graph measures of integration and segregation.

These are tested both by group comparison (low / middle / high PGS) and by formal heteroscedasticity tests on the continuous PGS (Breusch–Pagan, White, quantile regression).

Repository layout

code/
  A1_preprocess_phenotypic_data.py    # HCP behavioural + phenotypic cleanup, imputation
  A2_factor_analysis.R                # CFA for the social factor
  A3_evaluate_social_factor.py        # Sanity-check the CFA factor

  B1_plinkQC_genotype_qc.R            # Genotype QC (plinkQC)
  B2_translate_PGS_to_HCP.sh          # SNP harmonization, PCA, PRSice PGS
  B3_select_PGS_threshold.py          # p-value threshold selection (p=0.1, per Grove 2019)
  B4_extend_PGS_with_BLUP.sh          # GCTA BLUP extension to related individuals
  B5_evalute_BLUP_prediction.py       # BLUP evaluation + residualized PGS

  C1_run_univariate_fMRI_prediction.py      # Edge-wise connectivity ~ social / PGS
  C2_find_communities_fMRI.py               # Consensus community detection
  C2b_evaluate_communities.py               # Parcellation resolution tuning
  C3_perform_main_landscape_analysis.py     # Main landscape / graph-theory tests
  C3b_continuous_heteroscedasticity_analysis.py  # Continuous BP / White / quantile tests
  C5_visualize_networks.py                  # Bootstrap ellipses + network figures
  generate_publication_figures.py           # Standalone publication SVGs

  utils/                              # Shared helpers (ID tracking, connectome viz, …)

  Snakefile                           # Workflow definition (see README_SNAKEMAKE.md)
  config.yaml                         # Paths, thresholds, parcellation sizes
  environment.yml                     # Conda environment (BrainComp)
  README_SNAKEMAKE.md                 # Detailed Snakemake usage guide

Scripts are numbered by phase (A / B / C) and execution order; letter suffixes (C2b, C3b) mark companion analyses rather than replacements.

Requirements

Software

• Python 3.11 (see environment.yml)
• R ≥ 4.0 with lavaan, psych, dplyr, plinkQC
• PLINK 1.9 / 2.0
• PRSice-2 (invoked by B2)
• GCTA (invoked by B4, path set in config.yaml)
• Snakemake for workflow orchestration

Install the conda environment once:

conda env create -f environment.yml
conda activate BrainComp

Data

The pipeline expects HCP data and an autism GWAS summary file, organised under project_dir/data/ as configured in config.yaml:

• raw_anonymised/behavioural_data_anonymised.csv — HCP behavioural
• raw_anonymised/phenotypic_data_anonymised.csv — HCP phenotypic (demographics, QC)
• raw_anonymised/movement_data_anonymised.csv — motion summary per subject
• raw_anonymised/subject_ids_anonymised.txt — subject ID list
• raw_anonymised/Neuro_Chip_anonymised.{bed,bim,fam} — HCP Neuro-Chip genotypes
• raw_anonymised/iPSYCH_PGC_ASD_Nov_2017.gz — Grove et al. (2019) GWAS summary stats
• HCP_PTN1200/netmats/ — HCP group-ICA netmats (parcellations 15–300)

The raw HCP data is not included in this repository — access is governed by the HCP data use terms. See HCP data access.

Quick start

Edit config.yaml to point at your project_dir and GCTA install, then:

# Dry-run to inspect the DAG
snakemake --use-conda -n

# Run the full pipeline on 4 cores
snakemake --use-conda --cores 4

# Stop after community-detection tuning (Phase C, C2b)
snakemake --use-conda --cores 4 --until evaluate_parcellations

See README_SNAKEMAKE.md for per-phase targets, cluster execution, logging, and troubleshooting.

Outputs

Outputs are written to project_dir/{results,reports,figures,logs}/:

• results/ — numeric results (CSV): factor scores, PGS residuals, partitions, graph-theory metrics, heteroscedasticity tests, data-retention tracking.
• reports/ — per-step text reports summarising what ran and key statistics.
• figures/ — per-step PNGs. figures/publication/ contains standalone SVGs produced by generate_publication_figures.py.
• logs/ — one log file per Snakemake rule.

Reproducibility notes

• PGS threshold is fixed at p = 0.1 (the discovery-GWAS optimum from Grove et al. 2019) rather than re-tuned in-sample, to avoid overfitting. See the header of B3_select_PGS_threshold.py.
• Community detection uses consensus clustering across n_iterations runs (default 50); the resolution is selected a priori (5–8 communities for resting-state networks) in C2b_evaluate_communities.py. The selected partition is written to a stable filename (C2b_selected_partition.csv) so downstream steps inherit it without repeating the parameter search.
• Landscape analysis (C3, C3b, C5) runs at the single parcellation chosen by C2b — n_nodes is derived from the selected partition. Robustness is assessed across edge thresholds only (0.15 / 0.20 / 0.25) in C3_perform_main_landscape_analysis.py.

Citation

TODO — add paper / preprint citation once available.

License

This project is licensed under the terms of the MIT license.