Generative models of the human connectome

Richard F. Betzel, Andrea Avena-Koenigsberger, Joaquín Goñi, Ye He, Marcel A. de Reus, Alessandra Griffa, Petra E. Vértes, Bratislav Misic, Jean-Philippe Thiran, Patric Hagmann, Martijn P. van den Heuvel, Xi-Nian Zuo, Edward T. Bullmore, Olaf SpornsNeuroImage 2016 · DOI

TL;DR

This paper introduces a family of generative models to understand the rules governing the human brain's wiring (connectome). By combining geometric constraints with topological features like "homophily" (connections between similar regions), they developed a model that accurately reproduces key aspects of individual human connectomes. This approach also reveals how these wiring rules change across the lifespan, suggesting a dynamic rebalancing of brain organization.

What Problem Does This Paper Address?

The human connectome, a map of neural connections, is crucial for brain function. However, the underlying principles that dictate how this intricate network is formed and organized remain largely unknown. While previous research suggested that minimizing wiring cost (favoring shorter connections) plays a role, it couldn't fully explain the observed network topology. This paper aims to uncover the "generative rules" that shape the connectome by building and testing computational models that can generate synthetic brain networks mimicking real ones.

Methods

The core of this work is the development and application of a generative model for brain networks. This model builds synthetic networks by adding connections probabilistically. The probability of a connection forming between two brain regions (nodes) is determined by a formula that combines two factors:

  1. Geometric Wiring Rules: This component, represented by $E(u, v)^\eta$, favors connections based on the Euclidean distance between regions. The exponent $\eta$ controls whether shorter ($\eta < 0$) or longer ($\eta > 0$) connections are preferred.
  2. Non-Geometric Topological Rules: This component, represented by $K(u, v)^\gamma$, captures relationships between nodes beyond simple distance. The paper explores various definitions for $K(u, v)$, including:
  3. Preferential Attachment: Nodes with more connections are more likely to form new ones.
  4. Degree Assortativity: Nodes with similar numbers of connections tend to connect.
  5. Homophily (Matching Index): Nodes with similar connectivity profiles (i.e., similar neighbors) are more likely to connect.

The model's parameters ($\eta$ and $\gamma$) are optimized to minimize an "energy" function, which quantifies the discrepancy between the synthetic and observed connectomes across several graph-theoretic measures (degree, clustering, betweenness centrality, and edge length distributions).

Datasets

The study utilized several datasets:

  • Primary Dataset: 40 healthy adult participants (ages 18-40) scanned at the University Hospital Center and University of Lausanne (CHUV), Switzerland.
  • Replication Datasets:
    • Human Connectome Project (HCP) dataset (N=214 participants).
  • Nathan Kline Institute (NKI) dataset (N=126 participants, ages 7-85).
  • Additional Datasets for Robustness Checks: Various subsets and processed versions of the CHUV, HCP, and NKI datasets were used to test the model's sensitivity to factors like network density, hemisphere inclusion, and different weighting schemes.

Key Insights and Ideas

  • Geometry Isn't Enough: A purely geometric model (favoring short connections) fails to capture the full complexity of the human connectome, particularly its clustering and the presence of long-distance connections.
  • Homophily is Key: The most successful generative models incorporate "homophily," where regions with similar connection patterns tend to connect. This suggests that the brain's wiring isn't just about proximity but also about functional similarity.
  • A Trade-off in Wiring: The best-fitting model balances geometric constraints (penalizing long-distance connections) with homophilic attraction, creating a more realistic network structure.
  • Lifespan Reorganization: The parameters of the generative model change significantly with age, indicating that the rules governing brain wiring are not static but evolve throughout life. Specifically, the penalty on long-distance connections weakens with age.

Key Results

  • The Matching Index (MI) model, which combines geometric constraints with homophily, significantly outperformed other models, including the pure geometric model, in generating synthetic networks that closely matched observed connectomes.
  • The MI model achieved low energy values and reduced discrepancies across key graph-theoretic measures (e.g., KSk, KSb, KSc, KSe) compared to other models. For instance, it reduced discrepancies in degree and clustering coefficient sequences compared to the geometric model.
  • When applied to a lifespan dataset (ages 7-85), the model revealed that the geometric parameter ($\eta$, penalizing long-distance connections) decreased in magnitude with age, while overall energy ($E$) and specific KS statistics ($KSe$, $KSb$) increased, suggesting a less constrained, more "random" organization in older brains.

Limitations and Future Work

  • Model Simplification: The study acknowledges that the generative models are simplified representations of a complex biological process. While the MI model is the best fit among those tested, it's possible that other unexplored wiring rules or combinations could yield even better results.
  • Data Acquisition Artifacts: The use of diffusion MRI and tractography for inferring connections is prone to false positives and negatives, which could influence the results.
  • Node Definition and Scale: The choice of how to define network nodes (parcellation of the cortex) and the scale at which the connectome is best described remain open questions. Future work on data-driven parcellations is suggested.
  • Lifespan Dynamics: While the study shows age-related changes in model parameters, it did not explicitly model the growth and development of the connectome. Future work could explore more complex models that incorporate system-specific details for this purpose.
  • Alternative Energy Functions: The study used a specific "energy" function to evaluate model fit. Exploring alternative energy functions could reveal different optimal models.

Summary updated

BibTeX
@article{betzel2016generative,
  author = {Richard F. Betzel and Andrea Avena-Koenigsberger and Joaquín Goñi and Ye He and Marcel A. de Reus and Alessandra Griffa and Petra E. Vértes and Bratislav Misic and Jean-Philippe Thiran and Patric Hagmann and Martijn P. van den Heuvel and Xi-Nian Zuo and Edward T. Bullmore and Olaf Sporns},
  title = {{Generative models of the human connectome}},
  journal = {NeuroImage},
  year = {2016},
  doi = {10.1016/j.neuroimage.2015.09.041},
}
Figure