🧠 neuroscience

When are population connectome harmonics adequate representations of individual brains?

This study demonstrates that population-level connectome harmonics can be robust and practically useful for predicting cognition despite imperfect representation of individual brain structures, suggesting their adequacy depends on the specific scientific question and the type of information being analyzed.

Original authors: Bolton, T., Patel, J., Schoettner Sieler, M., Chan, C. H. M., Van De Ville, D., Hagmann, P.

Published 2026-10-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bolton, T., Patel, J., Schoettner Sieler, M., Chan, C. H. M., Van De Ville, D., Hagmann, P.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is not a single, uniform organ; it is a vast, intricate network of billions of cells connected by trillions of pathways. To understand how this complex system works, scientists often map these connections, creating a structural blueprint known as a connectome. This map shows which regions of the brain are physically linked by white matter cables, forming the physical scaffold upon which thoughts and feelings travel. However, because every person's brain is wired slightly differently, researchers face a difficult choice when studying groups of people. They can either build a unique map for every single individual, which captures every personal detail but makes comparing people nearly impossible, or they can create a single, average map that represents the "typical" brain. This average map allows scientists to compare groups and find common patterns, but it risks smoothing over the very differences that make each person unique. The central question is whether this simplified, population-level map is good enough to stand in for the complex, individual reality of a specific human brain.

A team of researchers set out to answer this question by analyzing brain scans from 875 healthy young adults. They focused on a mathematical tool called connectome harmonics, which breaks down the brain's structural map into a set of fundamental patterns, much like how a musical chord can be broken down into individual notes. These patterns range from smooth, slow waves that span the whole brain to rapid, jagged fluctuations that occur in small, localized areas. The researchers wanted to know if a single set of these patterns, derived from averaging the brains of many people, could accurately describe the brain of a single person who was not part of that original group. They tested this idea along three specific lines: first, whether the average map is stable and reproducible; second, whether it truly represents the unique structure of individual brains; and third, whether using this average map leads to correct conclusions when studying brain function and behavior.

The study revealed that the answer is not a simple yes or no, but rather depends entirely on what the scientist is trying to achieve. When the researchers increased the number of people used to build the average map, the map itself became more stable and consistent. If they built the map from 10 people and then from 50, the resulting patterns were quite different. But as they added more people, up to hundreds, the map settled into a reliable form that did not change much with the addition of new individuals. This means that for studies looking for broad, shared features of the human brain, a reasonably sized group is enough to create a solid, reproducible reference.

However, this stability did not mean the map was a perfect mirror of any single individual. Even with a large group of 700 people, the average map still missed significant details when compared to the unique wiring of a specific person. The researchers found that the average map was particularly good at capturing the large-scale, slow-moving patterns of brain organization, but it struggled to represent the finer, faster details. Crucially, they discovered that a map being stable and reproducible does not automatically mean it is representative of every individual. A map can be very consistent across different groups of people while still failing to capture the specific quirks of a single brain.

Despite these imperfections, the researchers found that the average map was surprisingly useful for practical applications. When they used the population-based map to analyze brain activity and predict cognitive abilities, the results were often just as good as, and sometimes even better than, using a map built from that specific person's own brain. This was especially true for the faster, more detailed patterns of brain activity. It appears that the average map acts as a kind of filter, smoothing out the noisy, irrelevant variations in an individual's brain that might confuse the analysis, while preserving the meaningful signals related to how the brain works. In fact, for predicting cognitive performance, the population-based map outperformed the individual map at high levels of detail.

The study also examined how the way scientists build these average maps affects the results. They tested different methods, such as deciding whether to include every possible connection between brain regions or only the ones that appear in most people. They found that the choice of which connections to include had a much larger impact on the results than how they calculated the strength of those connections. The overall shape and layout of the network mattered more than the specific numbers assigned to the links. This suggests that the structure of the brain's wiring is the most critical factor in creating a useful reference map.

Ultimately, the research concludes that there is no single "perfect" population map that works for every scientific question. Instead, the adequacy of such a map depends on the specific goal of the study. If a researcher needs to understand the precise, unique wiring of a single brain, a population average may not be sufficient. But if the goal is to understand general principles of brain organization or to predict behavior across a group, a population-level map is not only adequate but often superior. The study demonstrates that a representation can be useful and scientifically valid even if it does not perfectly match every individual it describes. The key is to choose the right tool for the job, recognizing that a simplified, shared view of the brain can reveal truths that a hyper-detailed, individual view might obscure.

Technical Summary: When are Population Connectome Harmonics Adequate Representations of Individual Brains?

Problem Statement
Network neuroscience increasingly utilizes Graph Signal Processing (GSP) to analyze brain organization, representing the structural connectome (SC) as a graph and functional activity as a signal evolving on that graph. A fundamental methodological choice arises in subject populations: should connectome harmonics (the eigenvectors of the graph Laplacian) be derived from a population-level SC (averaged across subjects) or from individual-specific SCs?

Population-level harmonics offer a common spectral coordinate system facilitating group-level comparisons but primarily encode shared structural organization rather than individual anatomical scaffolds. Conversely, individualized harmonics preserve subject-specific structure but yield non-comparable spectral spaces. While population representations are widely used, their adequacy for representing individual brains remains unclear. Specifically, it is unknown whether a population representation is robust (reproducible across samples), representative (faithfully capturing individual structural organization), and practically useful (yielding valid conclusions in downstream functional analyses). Furthermore, the relationship between these properties and their dependence on sample size, connectome construction methods, and graph frequency are not well-defined.

Methodology
The authors analyzed multimodal neuroimaging data from 875 healthy young adults from the Human Connectome Project (HCP). The study employed a nested sampling framework to construct population-level structural connectomes across a broad range of sample sizes ( $N \in \{5, \dots, 700\}$ ).

Connectome Construction: Four population-level SC construction schemes were evaluated by crossing two edge selection strategies (AllEdges vs. ConsensusEdges) and two weight definition strategies (AllValues vs. NZValues).
Harmonic Generation: Connectome harmonics were generated via eigendecomposition of the symmetric normalized graph Laplacian for both population-level and subject-specific SCs.
Three-Dimensional Evaluation: The study assessed population harmonics along three complementary dimensions:
- Robustness: The consistency of harmonic bases estimated from non-overlapping population samples of equal size. Metrics included harmonic similarity, required projection window, and Laplacian reconstruction error.
- Representativeness: The ability of population-level harmonics to capture the structural organization of independent individuals (left-out subjects). Metrics included harmonic similarity, spectral bandwidth, leakage entropy, and subspace similarity.
- Practical Utility: The consequences of replacing individualized harmonics with population-level ones in downstream functional analyses. This included graph filtering (low- and high-pass), functional error, second-order moment error, functional connectivity (FC) error, and the prediction of cognitive performance.
Statistical Analysis: Convergence was assessed using linear mixed-effects models across sample size contrasts. Relationships between the three dimensions were quantified via Spearman correlations and Fisher-transformed correlations. Power-law models were fitted to estimate the sample size ( $N_{90}$ ) required to achieve 90% of modeled convergence.

Key Contributions

Decoupling of Properties: The study establishes that Robustness, Representativeness, and Practical Utility are distinct properties that do not necessarily co-vary. A representation can be highly reproducible (Robust) without being fully representative of individual structures, and imperfect individual correspondence does not necessarily preclude practical utility.
Frequency-Dependent Dissociation: The authors demonstrate a dissociation between structural fidelity and predictive utility. While population-level harmonics showed lower structural mismatch than individual ones at high frequencies, they resulted in higher functional errors (second-order moments and FC). Paradoxically, high-frequency FC derived from population-level harmonics predicted cognition more accurately than individualized harmonics, suggesting population bases may act as a form of regularization.
Construction Sensitivity: The study identifies that the choice of network topology (edge selection) exerts a significantly greater influence on Robustness, Representativeness, and Practical Utility than the definition of edge weights.
Sample Size Trajectories: The research maps the convergence trajectories of these properties, showing that Practical Utility stabilizes at smaller sample sizes than average Representativeness, which in turn stabilizes earlier than the precise resolution of stable inter-individual differences.

Key Results

Robustness: Increases monotonically with sample size. Significant and tangible improvements persist across all investigated sample size increments (up to $N=350$ ). Local spectral subspace correspondence stabilizes earlier ( $N_{90} \approx 263$ ) than one-to-one harmonic correspondence ( $N_{90} \approx 770$ ) or Laplacian reconstruction ( $N_{90} \approx 526$ ).
Representativeness: Mean Representativeness stabilizes rapidly at small sample sizes ( $N < 50$ ), with little tangible improvement thereafter. However, the stability of inter-individual differences (the ability to consistently rank subjects by how well they are represented) improves progressively with larger sample sizes, converging only at very large $N$ .
Practical Utility: The functional consequences of using population-level harmonics are largely insensitive to population sample size once $N$ $N$ exceeds small thresholds. Crucially, the choice of graph frequency dictates the outcome:
- Low Frequencies: Population and individual bases yield nearly identical cognitive prediction performance ( $R^2 \approx 0.089$ ).
- High Frequencies: Population-level bases outperform individualized bases in predicting cognition ( $R^2 = 0.175$ vs. $0.139$), despite higher structural mismatch and functional errors.
Construction Effects: Topology (AllEdges vs. ConsensusEdges) is the dominant factor. Consensus topology increases Robustness but decreases Representativeness. It also significantly alters the differential utility of low- vs. high-frequency filtering.

Significance and Claims
The paper argues that the adequacy of a population-level connectome representation cannot be determined by a single metric or a universal sample size threshold. Instead, adequacy is defined by the specific scientific question and the information the analysis must preserve.

Context-Dependent Adequacy: If the goal is to interpret the harmonic basis itself, Robustness is paramount. If the goal is to represent individual structural organization, Representativeness is central. If the goal is downstream functional analysis or prediction, Practical Utility is the deciding factor.
Reproducibility vs. Fidelity: The findings demonstrate that population representations can be reproducible without being fully representative, and useful despite imperfect individual correspondence.
Methodological Guidance: The authors propose a framework where researchers should align their choices of population size, connectome construction, and graph frequency range with their specific inferential objectives. For instance, studies focusing on individual deviations may require larger samples to stabilize subject ordering, whereas studies focused on group-level functional connectivity may find smaller populations sufficient for practical utility.
Limitations: The authors modestly note that their findings are based on a homogeneous cohort of healthy young adults and that the relationships observed may differ in heterogeneous or clinical populations. They also acknowledge that "Representativeness" is defined relative to estimated individual harmonics, which are themselves not ground truth.

In conclusion, the paper posits that there is no universally optimal population representation. The "correct" population-level connectome is the one that preserves the specific properties required for the intended scientific application, whether that be high reproducibility, high individual fidelity, or optimal downstream predictive power.

Technical Summary: When are Population Connectome Harmonics Adequate Representations of Individual Brains?

More like this