Multi-type branching inference on contact trees with application to COVID-19

Decoding the Outbreak: Beyond Genomic Sequences

Epidemiologists rely on reconstructing how a disease moves through a population to predict its future trajectory. Traditionally, this has meant looking at pathogen genomes—analyzing the subtle mutations in a virus's DNA to build a family tree of transmission. However, this genomic approach is often limited by the speed of sequencing and the availability of genetic data. A more direct way to map an epidemic is through contact tracing, where investigators build trees based on reported transmission events between individuals.

The challenge is that these contact trees are inherently messy and incomplete. Current mathematical models for analyzing these trees often assume "homogeneous mixing." This is the idea that any infected person is equally likely to infect any susceptible person in the population. This ignores a fundamental biological reality: as an individual infects their contacts, they deplete their own pool of susceptible neighbors. This "saturation effect" means an individual's potential to spread the disease naturally drops as they move through their personal contact network. Until now, models have struggled to reconcile this changing transmission potential with the fragmented, partially observed data typical of real-world contact tracing.

The Failure of Homogeneous Mixing

Most existing tree-based likelihood methods, such as the multi-type birth-death (MTBD) models used in phylodynamics (the study of evolutionary processes using genetic data), operate under the assumption that transmission rates depend only on a fixed "type" or state of an individual. While powerful, these models rarely capture how an individual's transmission potential changes dynamically as they exhaust their local contacts.

In a real-world network, an individual does not have infinite transmission opportunities. If a person has ten contacts and has already infected three of them, their remaining capacity to drive the epidemic is significantly reduced. Previous attempts to incorporate this, such as pairwise coalescent models, stratified the population by contact number. However, they still assumed an individual transmits at a rate proportional to their total contacts, regardless of how many are already infected. This oversight creates a blind spot. It fails to account for the "contact depletion" that slows down local clusters. This can lead to biased estimates of how fast a disease is truly spreading.

Modeling Saturation via Augmented States

To resolve this, Okolie et al. developed a likelihood framework that operates directly on transmission trees. The core innovation is the use of an augmented state to track an individual's changing status. Instead of just recording if someone is infected, the model characterizes each node by a pair $(i, k)$. Here, $k$ represents the individual's total downstream degree (the total number of contacts available for onward transmission). Meanwhile, $i$ represents the number of those contacts that have already been infected.

The mechanism works through a stochastic Susceptible-Infected-Recovered (SIR) process governed by several layers of logic:

1. Dynamic Transmission Rates: The effective infection rate for a node is not a constant $\beta$, but rather $(k - i)\beta$. As $i$ increases, the rate at which the individual can cause new infections automatically scales down.
2. Closed-Form Probabilities: The authors derive ordinary differential equations (ODEs) to calculate two critical values. First, they find the probability that a specific branch of the tree (a clade) goes entirely unobserved by health officials. Second, they find the probability density that a branch produces exactly one observed case of a specific type.
3. Handling Random Networks: Not everyone has the same number of contacts. Therefore, the model treats the degree $k$ as a random variable drawn from a distribution. It uses "degree-averaged" terms. This ensures that when a new person is infected, the model correctly accounts for the fact that their specific number of contacts is unknown.

These calculations allow the model to traverse the tree from the "tips" (the observed cases) back to the "root" (the first case). This process calculates the total likelihood of seeing that specific pattern of infections and timings.

Validating Accuracy and Robustness

The authors tested the framework using two distinct approaches: highly controlled simulations and real-world clinical data. In the simulation studies, the model demonstrated remarkable precision. When tasked with recovering the basic reproduction number ($R_0$)—the average number of secondary cases produced by a single infection—the model achieved an estimate of $6.04$ [Table 1]. This was calculated for a simulation where the true value was $6.0$.

Crucially, the researchers investigated how "resolved" a tree needs to be to yield good results. In a "fully resolved" tree, every single branching event and timing is known. In a "partially resolved" tree, much of the internal structure is hidden. This reflects the reality of imperfect contact tracing. The paper finds that while $R_0$ is incredibly robust, estimating the actual contact degree ($k$) requires more detail. Specifically, the authors report that at least half of the internal branching times must be observed to recover the contact structure reliably .

The framework was finally applied to the first wave of COVID-19 in Karnataka, India. Analyzing a cohort of 2,401 individuals, the model produced a posterior mean $R_0$ of $2.60$ [Table 4]. This result was consistent with independent studies of the same outbreak. It proves that the model can extract meaningful epidemiological intelligence from noisy, incomplete data.

Limitations of the Static Tree Assumption

While the model is a significant leap forward, it is not a perfect mirror of reality. The framework relies on several simplifying assumptions:

• Static Topology: The model assumes a "static" contact tree. This means the connections between people are fixed. In reality, human social networks are dynamic. People form new connections and break old ones throughout an infectious period.
• Fixed Degrees: It assumes an individual's contact degree $k$ remains constant. If a person's social activity changes significantly after they are diagnosed, the model's accounting of their "remaining" transmission potential may be skewed.
• No Reticulate Patterns: The model is built on a tree structure. This assumes each infection has exactly one source. Real-world outbreaks often feature "reticulate" patterns. These occur when an individual might be infected by multiple sources or where "outside infections" occur that are not captured in the local contact trace.

The Verdict: A New Baseline for Trace-Based Inference

The research presented by Okolie et al. is a definitive "yes" for anyone looking to move beyond genomic-only epidemiology. By successfully integrating the physics of contact depletion into a formal likelihood framework, the authors have provided a principled way to turn raw contact-tracing logs into rigorous statistical estimates.

The model is particularly valuable because it identifies a crucial trade-off. You do not need a perfect, fully-mapped contact network to understand the intensity of an epidemic ($R_0$). However, you do need a relatively high-resolution tree if you want to understand the social architecture ($k$) driving it. For practitioners dealing with the inevitable gaps in public health data, this distinction is vital. The code is available at https://github.com/Austine316/Multi-type-branching-inference-on-contact-trees.

Figures from the paper