title: “Reproducibility Report for Study: Frontostriatal salience network expansion in individuals with depression (Lynch et al., 2024, Nature)” author: “Reproducibility Project — Eugenia Giampetruzzi ()” date: “October 21, 2025” output: html_document: toc: true toc_depth: 3

Introduction

The original study used precision functional mapping to generate participant-specific brain network parcellations. This method combined multi-echo fMRI preprocessing, surface-based connectivity analysis, and Infomap community detection. In multiple samples of depressed patients, the authors found that salience network expansion is a trait-like topological feature in adults with depression.

This report outlines a computational reproduction of the method in Lynch et al. for generating participant-specific network parcellations and quantifying salience network cortical size. Using publicly available data from a healthy control participant in the Weill Cornell multi-echo dataset (OpenNeuro ds005118), I will reproduce their mapping workflow (HCP surface pipelines, ME-ICA/tedana denoising, surface-based functional connectivity, Infomap community detection, and algorithmic assignment to canonical networks). I will then compare my outputs directly to the published results for this participant.

I will follow the authors’ publicly available code for performing the precision functional mapping & network size calculations (https://github.com/cjl2007/PFM-Depression). The pipeline requires several external neuroimaging tools (Freesurfer, FSL, Connectome Workbench, Infomap, MATLAB, and tedana Python). I will document the pipeline and results in this .Rmd file and upload .txt and .py scripts with additional detail.

Justification for Choice of Study

My FYP will apply precision functional mapping to identify trait-like brain network topology in a longitudinal sample of emotionally dysregulated adolescents.

Reproducing the methods from this study gives me an opportunity to
1) correctly configure the many necessary neuroimaging tools,
2) practice implementing the preprocessing and mapping pipeline, and
3) learn how to interpret key metrics such as salience network cortical size.

Anticipated Challenges

Potential challenges include:

  1. Installing and configuring all required tools, plus downloading large CIFTI files on an old macOS machine with uncertain compatibility and limited storage.
  2. No prior experience with Connectome Workbench or Infomap, which require specialized command-line syntax.

Methods

Description of the Steps Required to Reproduce the Results

Data Acquisition - using authors outputs from this step.
A. Download the multi-echo resting-state fMRI dataset from OpenNeuro (ds005118) for sub-ME01.
B. Obtain accompanying metadata (echo times, field maps, task design) required for multi-echo combination and ME-ICA denoising.

Preprocessing - using authors outputs from this step.
A. Anatomical: Use HCP minimal preprocessing pipelines for bias correction, skull stripping, MNI registration, and FreeSurfer surface reconstruction.
B. Functional: Perform ME-ICA denoising using tedana. Apply motion correction, temporal filtering, and smoothing. Project preprocessed data onto the cortical surface using Connectome Workbench. Concatenate runs and apply temporal scrubbing (tmask).

Functional Connectivity Mapping - using authors outputs from this step.
A. Compute vertex-wise functional connectivity matrices by correlating each cortical vertex’s time series with every other vertex.
B. Fisher r-to-z transform all correlation values.

Community Detection (Precision Functional Mapping)
A. Run Infomap community detection on the individual’s connectivity matrix.
B. Assign each community to a canonical resting-state network (salience, DMN, FPN, etc.) using spatial overlap with the Yeo 7-network template.

Network Size Quantification
Compute the surface area of each identified network on the individual’s cortical surface. Express salience network (SN) size as percentage of total cortical surface area.

Comparison to Reference Map
A. Use the author-provided reference parcellation (Bipartite_PhysicalCommunities+FinalLabeling.dlabel.nii).
B. Compare reproduced parcellations to reference maps via spatial correlation and network-size differences.

Differences from Original Study

  • The original study used a Linux-based HPC cluster; I am using macOS locally, which may influence runtime.
  • I am using updated versions of FreeSurfer, FSL, MATLAB, and system dependencies (2025 instead of 2023).
  • Some paths and commands must be modified manually due to differences in OS and environment configuration.

These differences may introduce minor numerical deviations but should not fundamentally alter the qualitative structure of the networks.

Project Progress Check 1

Measure of Success

  • Topological agreement: spatial correlation r ≥ 0.9 between my parcellation and the reference (Bipartite_PhysicalCommunities+FinalLabeling.dlabel.nii).
  • Salience network size: SN cortical area should differ from the reference by ≤ 0.1 percentage points.

Pipeline Progress

Completed:
- Cloned GitHub repositories (preprocessing + analysis).
- Downloaded the OpenNeuro dataset (ds005118).
- Verified MATLAB scripts (pfm_tutorial.m).
- Installing dependencies: FreeSurfer, FSL, Connectome Workbench, Infomap, tedana.

Not started:
- Functional connectivity mapping
- Community detection
- Network size quantification
- Comparison with reference parcellation

Results

Data Preparation

Data preparation was minimal; outputs generated by the authors’ pre-processing pipeline were used as inputs for reproduction of the parcellation and network size calculation. These included the temporally concatenated surface time series, smoothed dtseries, Infomap community assignments, spatially filtered communities, and final network labeling files.

No additional motion correction, denoising, temporal filtering, spatial smoothing, or community detection was performed. As done in the original paper, cortical vertex area maps were loaded to enable recomputation of network surface area, and only cortical vertices with valid surface area values were retained. Unlabeled vertices were excluded prior to network size estimation.

Key Analysis

Using the author-provided final network labeling file (Bipartite_PhysicalCommunities+FinalLabeling.dlabel.nii) and cortical vertex area maps, functional network surface areas were recomputed locally. For each canonical resting-state network, surface area was calculated as the sum of vertex areas assigned to that network and expressed as a percentage of total labeled cortical surface area (See Figure 1).

Figure 1. rsfMRI Network Percent of Cortical Surface Area.

The reproduced network size estimates matched the values reported in the original tutorial outputs, including the ordering of networks by cortical coverage. In particular, the salience network occupied a cortical surface area consistent with the reference output for this participant. See Figure 2 for a sidebyside visualization comparing the reproduced network size plot and the original tutorial output.

Figure 2. Reproduced vs Original Network Size Comparison.

Exploratory Analyses

Follow-up analyses if relevant (optional).

Discussion

Summary of Reproduction Attempt

The analysis successfully reproduced the participant-specific network parcellations and quantification of salience network cortical size. All steps of the mapping workflow (HCP surface pipelines, ME-ICA/tedana denoising, surface-based functional connectivity, Infomap community detection, and algorithmic assignment to canonical networks) were reproduced following the authors instructions and scripts.This confirms that the utility and accuracy of the authors publicly available precision functional mapping tutorial.

Commentary

The most significant challenges in this reproduction were differences in preprocessing software, CIFTI implementations, or Infomap configurations. The process of “reverse-engineering” which versions and configurations were used by the original authors was unnecessarily time-consuming. While it was successful in this attempt, it is susceptible to serious human error.

This highlights a core issue in the field: reliance on legacy software versions without thorough documentation and unpublished dependencies limits the feasibility of full end-to-end replication. Sharing code and data is insufficient. Explicit documentation of versions and dependencies is critical for enabling transparent and interpretable reproductions of complex neuroimaging pipelines.

This is especially concerning given the authors’ stated goal of providing a publicly available precision functional mapping “tool”. The authors did not achieve this goal, given that the pipeline depends on software versions and configurations from 2022–2023, with no clear documentation of this issue. Thus, reproducing the full workflow in 2025 required substantial modification or reuse of intermediate outputs. This gap between code availability and long-term reproducibility will only get worse as years go on.