A modular Bayesian framework for inferring transmission networks from polyclonal infections, with application to Plasmodium falciparum

This paper introduces a modular Bayesian framework, exemplified by the Plasmotrack software for *Plasmodium falciparum*, that reconstructs directed transmission networks from polyclonal infections by accommodating multiple genetic sources and unobserved parents to estimate key public health metrics.

The Gist

Imagine trying to figure out who passed a secret note to whom in a crowded classroom, but with two major twists: first, the notes are written in a code that changes slightly every time it's copied; and second, some students aren't just holding one note—they are juggling several different notes at once, each coming from a different classmate.

This is the challenge scientists face when trying to track how diseases like malaria spread.

The Problem: The "One Note, One Source" Myth
Most existing tools for tracking disease spread are built like a simple relay race. They assume that if Student B gets sick, they got it from exactly one Student A, who got it from Student Z, and so on. They also assume the "note" (the genetic code of the germ) stays mostly the same as it passes down the line.

But in the real world, especially with diseases like malaria, tuberculosis, or HIV, this assumption often fails. A person can get infected by multiple different sources at the same time. It's like Student B receiving a stack of notes from three different people simultaneously. The old tools get confused by this "polyclonal" mess and can't draw an accurate map of who infected whom.

The Solution: A Modular Detective Kit
The authors of this paper built a new, flexible detective kit called a "modular Bayesian framework." Think of it as a smart, adaptable puzzle solver.

Instead of forcing the data to fit a simple "one-to-one" story, this new system allows for complex stories:

• Multiple Parents: It can figure out that a patient was infected by a combination of sources.
• Missing Pieces: It acknowledges that some "parents" (infectors) might not be in the data set at all (like a student who left the classroom before the notes were collected).
• Plug-and-Play Design: The system is "modular." Imagine a Lego set where the core brain is the same, but you can swap out the "legs" depending on the disease. For malaria, you attach a specific "malaria leg" that understands how malaria genes mix. For a different disease, you could swap in a different leg without rebuilding the whole machine.

The Test: Plasmotrack
To prove this works, the authors built a specific version of their kit for malaria called Plasmotrack. They fed it data from targeted genetic tests (like taking a snapshot of the malaria genes in a patient's blood).

They ran a massive simulation where they created a fake world of malaria spread with known rules. Even when the simulation was tricky and the genetic data didn't perfectly match the rules (a bit like a blurry photo), the system was still able to:

1. Correctly guess the average number of people one infected person went on to infect.
2. Accurately estimate how many infections came from "outside" sources (people not in the study).
3. Draw the correct lines showing who likely infected whom with high precision.

The Bottom Line
This paper introduces a new way to map disease transmission that doesn't get confused when a patient has multiple infections at once. It successfully reconstructed the "who-infected-whom" network for malaria using genetic data, even when the data was messy. The software, Plasmotrack, is now available for others to use and adapt for their own disease-tracking needs.

Technical Summary

Technical Summary: A Modular Bayesian Framework for Inferring Transmission Networks from Polyclonal Infections

Problem Statement
Current methods for reconstructing infectious disease transmission networks using genetic data are predominantly designed for scenarios where each infection is characterized by a single dominant source, a single representative haplotype, and genetic divergence driven primarily by mutation along transmission chains. These assumptions fail to accommodate polyclonal infections, where a single host carries multiple genetically distinct clones and recipient infections may arise from contributions by multiple sources. This limitation is critical for pathogens such as Plasmodium falciparum (malaria), tuberculosis, HIV, and various parasitic infections, where polyclonality is common. Consequently, existing tools cannot accurately estimate public-health quantities derived from transmission networks—such as importation rates, source-sink structures, and differential onward transmission across intervention strata—when the underlying data involves complex, multi-clone infections.

Methodology
The authors propose a modular Bayesian framework designed to estimate directed transmission networks among sampled cases while explicitly accommodating polyclonal data. The core innovation lies in its flexibility regarding parentage: an infection may be inferred to have no sampled parent, a single parent, or multiple parents (including sources outside the observed panel).

The framework operates through two primary components:

1. Pathogen-Specific Modules: These supply likelihoods over candidate parent sets. They connect to the shared inference engine to generate posterior distributions.
2. Shared Inference Engine: This component aggregates the likelihoods to yield:
   • Marginal directed edge probabilities.
   • Posterior mean out-degree.
   • Inclusion probabilities for unobserved parents.

To demonstrate this framework, the authors developed Plasmotrack, a specific implementation for Plasmodium falciparum utilizing targeted amplicon sequencing data. Plasmotrack integrates three specific technical components:

• A per-locus allele-mixture transmission likelihood to handle the complexity of mixed clones.
• An amplicon genotyping error model to account for sequencing artifacts.
• Data augmentation techniques to explicitly model and infer unobserved parents (sources not present in the sampled panel).

Results
The framework was evaluated using simulations generated from a biologically informed generative model. Notably, these simulations tested the framework under conditions where the inferential per-locus allele-mixture likelihood was misspecified, representing a realistic scenario where the model does not perfectly match the generative process.

Under these conditions, Plasmotrack demonstrated:

• Recovery of Aggregate Network Summaries: The model successfully recovered key metrics, including the mean out-degree and the mean inclusion of unobserved sources.
• High Precision and Recall: The framework maintained high accuracy in detecting specific directed transmission events despite the model misspecification.
• Modularity: The architecture allows other pathogens to reuse the same modular composition by simply substituting the transmission and observation likelihoods, without altering the core inference engine.

Significance and Claims
The paper claims to provide a necessary methodological advancement for molecular surveillance in the context of polyclonal infections. By moving beyond the "single-source, single-haplotype" assumption, the framework enables the estimation of transmission network functions that were previously difficult to observe directly. The authors position Plasmotrack not as a universal solution for all pathogens, but as a validated proof-of-concept for P. falciparum that establishes a reusable template for other infectious diseases. The availability of the software and documentation under an open-source license is presented as a resource for the broader community to adapt these methods to their specific pathogen contexts.