Disentangling network topology and pathogen spread

María Pérez‐Ortiz, Petru Manescu, Fabio Caccioli, Delmiro Fernández-Reyes, Parashkev Nachev, John Shawe‐TaylormedRxiv 2021 · DOI

Abstract

How do we best constrain social interactions to prevent the transmission of communicable respiratory diseases? Indiscriminate suppression, the currently accepted answer, is both unsustainable long term and implausibly presupposes all interactions to carry equal weight. Transmission within a social network is determined by the topology of its graphical structure, of which the number of interactions is only one aspect. Here we deploy large-scale numerical simulations to quantify the impact on pathogen transmission of a set of topological features covering the parameter space of realistic possibility. We first test through a series of stochastic simulations the differences in the spread of disease on several classes of network geometry. We then aim to characterise the spread based on the characteristics of the network topology using regression analysis, highlighting some of the network metrics that influence the spread the most. For this, we build a dataset composed of more than 9000 social networks and 30 topological network metrics. We find that pathogen spread is optimally reduced by limiting specific kinds of social contact – unfamiliar and long range – rather than their global number. Our results compel a revaluation of social interventions in communicable diseases, and the optimal approach to crafting them.

Introduction

Traditional epidemiological models often rely on the "homogeneous mixing" assumption—the idea that any individual in a population is equally likely to encounter any other. However, human contact is a complex web where structure dictates flow. Current public health policies often rely on "indiscriminate suppression" (lockdowns) because they fail to distinguish between different types of social links. This research seeks to "disentangle" which specific topological metrics actually predict how a pathogen moves through a society, moving beyond blunt instruments toward surgical interventions.

Methodology

The study employed a large-scale numerical simulation approach using a newly generated synthetic dataset of 9,061 social networks. These networks were modeled after Erdos-Renyi (random), Scale-free (hubs), and Small-world (short paths) structures, alongside custom community-based generators to capture realistic social clustering.

On these networks, the researchers ran a compartmental SEIR model (Susceptible-Exposed-Infectious-Recovered), calibrated for SARS-CoV-2 parameters. For each network, 30 different metrics were calculated, categorized into Distance (e.g., path lengths), Connection (e.g., degree distribution), and Spectral (e.g., algebraic connectivity) metrics. Multivariate linear regression was then used to determine which metrics best predicted the total percentage of infected individuals.

The Deceptive Nature of Connectivity Volume

A primary finding of the study is that the total number of contacts (average degree) is a poor predictor of total infection compared to the network's underlying structure. Networks with identical average degrees can result in vastly different outcomes, ranging from 10% to 60% total infection.

Impact of global interactions on infection spread across different average degrees. Even with the same average degree, the presence of global interactions significantly shifts the distribution toward higher infection percentages.

As shown in Figure 1, the impact of "global interactions"—incidental contacts with strangers or people outside one's immediate circle—is particularly pronounced in weakly connected networks. For an average degree of 4, the introduction of global interactions causes a dramatic shift in the probability density of infected individuals, highlighting that "who" you are connected to matters significantly more than "how many" people you meet.

Identifying Key Topological Predictors

The research identified Global Efficiency as the most potent predictor of disease spread. This metric measures how easily a virus can travel between any two random nodes in a graph. While traditional "local" metrics like the number of immediate friends (average degree) had an $R^2$ of approximately 0.47, models using only Global Efficiency achieved an $R^2$ of roughly 0.75.

Correlation between various topological metrics and the percentage of infected individuals. Global efficiency and algebraic connectivity show the strongest and most consistent relationships with pathogen spread.

The scatter plots in Figure 5 demonstrate the varying predictive power of different metrics. While metrics like clustering show a noisy relationship with infection levels, Global Efficiency and Algebraic Connectivity exhibit clear, strong correlations. The full regression model, incorporating multiple topological features, achieved an adjusted $R^2$ of 0.934, explaining nearly all variance in pathogen spread based solely on network topology.

Conclusion and Policy Implications

The study confirms that random-mixing models often used in policy-making tend to overestimate total infections by ignoring realistic social structures. The results suggest that pandemic control should focus on limiting "long-range" or unfamiliar interactions rather than total contact volume.

While metrics like Global Efficiency are challenging to measure in real-time without high-fidelity mobility data, they provide a roadmap for more sustainable public health interventions. Future work should focus on estimating these crucial global metrics to allow for "provincial" or targeted social distancing that preserves local community links while severing the high-efficiency pathways that drive rapid transmission.

Summary updated

Abstract

Abstract How do we best constrain social interactions to prevent the transmission of communicable respiratory diseases? Indiscriminate suppression, the currently accepted answer, is both unsustainable long term and implausibly presupposes all interactions to carry equal weight. Transmission within a social network is determined by the topology of its graphical structure, of which the number of interactions is only one aspect. Here we deploy large-scale numerical simulations to quantify the impact on pathogen transmission of a set of topological features covering the parameter space of realistic possibility. We first test through a series of stochastic simulations the differences in the spread of disease on several classes of network geometry (including highly skewed networks and small world). We then aim to characterise the spread based on the characteristics of the network topology using regression analysis, highlighting some of the network metrics that influence the spread the most. For this, we build a dataset composed of more than 9000 social networks and 30 topological network metrics. We find that pathogen spread is optimally reduced by limiting specific kinds of social contact – unfamiliar and long range – rather than their global number. Our results compel a revaluation of social interventions in communicable diseases, and the optimal approach to crafting them.
BibTeX
@misc{perezortiz2021disentangling,
  author = {María Pérez‐Ortiz and Petru Manescu and Fabio Caccioli and Delmiro Fernández-Reyes and Parashkev Nachev and John Shawe‐Taylor},
  title = {{Disentangling network topology and pathogen spread}},
  year = {2021},
  doi = {10.1101/2021.05.24.21257706},
  howpublished = {medRxiv},
}
Figure