DENcode: A model for haplotype-informed transmission probability of dengue virus

The paper introduces DENcode, a robust framework that integrates temperature-modulated epidemiological kernels with phylogenetically informed genetic similarity (using both haplotype and consensus sequences) to estimate probabilistic transmission links between dengue cases, demonstrating its ability to reconstruct informative transmission networks and identify key infection sources using data from Colombo, Sri Lanka.

The Gist

Imagine you are trying to solve a massive, invisible puzzle. The pieces are people who have caught Dengue fever, and the picture you are trying to reveal is who infected whom.

In a crowded city like Colombo, Sri Lanka, thousands of people get sick with Dengue every year. The virus is carried by mosquitoes, but the mosquitoes don't fly very far. So, usually, if you get sick, you got it from a mosquito that bit someone nearby. But here's the problem: mosquitoes don't leave a trail, and people move around. You might get sick in your neighborhood, but you could have been bitten by a mosquito that fed on a sick person who lives 10 kilometers away and traveled there for work.

Traditional methods of tracking this are like trying to solve the puzzle with only the back of the picture pieces. They look at the virus's "family tree" (genetics), but because Dengue is an acute (short-term) infection, the virus doesn't have time to change much. It's like trying to tell two identical twins apart by looking at their baby photos; they look too similar to know who is who.

Enter: DENcode (The Detective's Toolkit)

The authors of this paper built a new digital detective tool called DENcode. Think of it as a super-smart algorithm that acts like a crime scene investigator for viruses. Instead of just looking at the virus, it combines three different types of clues to figure out the most likely story of how the infection spread.

1. The "Time and Temperature" Clue (The Epidemiological Kernel)

Imagine you are trying to figure out if Person A could have infected Person B.

• The Clock: You check their calendars. Did Person A get sick before Person B?
• The Weather: Mosquitoes are like little engines that run on heat. If it's hot, the virus grows faster inside the mosquito. If it's cold, it takes longer. DENcode checks the daily temperature to see if the virus had enough time to mature inside a mosquito and jump from A to B.
• The Map: Mosquitoes are lazy flyers; they usually only travel a few hundred meters. DENcode checks the distance between the two people's homes. If they live next door, it's a strong clue. If they live in different cities, it's a weak clue (unless the sick person traveled).

2. The "Genetic Fingerprint" Clue (The Haplotype Kernel)

This is where the magic happens.

• The Old Way (Consensus): Usually, scientists take a sample of blood and create an "average" picture of the virus. It's like taking a photo of a crowd and blurring it until everyone looks like a generic blob. This is called a consensus sequence. It's too blurry to tell who is related to whom.
• The New Way (Haplotype): DENcode uses a high-powered microscope. It looks at the individual variants of the virus living inside a single person's body. Think of it like looking at a crowd and seeing that Person A has a red hat, a blue scarf, and a green shoe, while Person B has a red hat, a blue scarf, and a red shoe.
  • Because the virus mutates (changes) rapidly inside a human, Person A's virus is a "parent" to the specific variants found in Person B.
  • By comparing these tiny, unique genetic "shoes and scarves," DENcode can say, "Yes, Person A definitely passed the virus to Person B," even if they live far apart.

3. Putting It All Together

DENcode takes the Time/Weather/Distance clues and the Genetic Fingerprint clues and mixes them together in a mathematical blender. It spits out a probability score for every pair of sick people.

• Score 0.9: "I am 90% sure Person A infected Person B."
• Score 0.1: "It's possible, but unlikely."

What Did They Find?

The researchers tested this tool on real data from 90 patients in Sri Lanka. Here is what they discovered:

• The "Super-Spreaders": The tool identified specific people who were the "hubs" of the outbreak. These were the people who, despite not being the first to get sick, were the ones who connected different clusters of infection. They were the bridges that allowed the virus to jump across the city.
• The Power of Detail: When they tried to use the "blurry" average virus data (consensus), the map fell apart. They missed most of the connections. But when they used the detailed "haplotype" data, they found 3 to 4 times more connections. It's like switching from a low-resolution map to a satellite image; suddenly, you can see the roads, not just the country borders.
• The Traveler: In one case, the tool found a strong link between a patient in the capital city (Colombo) and a patient in a city 100km away (Galle). The genetic clues were so strong that the tool knew they were connected, even though the mosquitoes couldn't have flown that far. It correctly deduced that a human must have traveled between the two cities, carrying the virus with them.

Why Does This Matter?

Imagine you are a firefighter trying to stop a forest fire.

• Without DENcode: You are guessing where the fire started and where it might go next based on wind direction alone. You might put out the wrong trees.
• With DENcode: You have a thermal camera and a wind sensor. You can see exactly where the fire jumped, who carried the burning ember, and which specific trees are most likely to catch fire next.

This model helps health officials stop Dengue outbreaks faster. Instead of spraying mosquitoes everywhere, they can target the specific neighborhoods and the specific "super-spreader" individuals who are driving the outbreak. It turns a chaotic mystery into a solvable puzzle.

Technical Summary

1. Problem Statement

Dengue virus (DENV) transmission networks are frequently difficult to resolve due to:

• Sampling Gaps: Incomplete case reporting and unobserved mosquito-mediated transmission.
• Low Genetic Diversity: As an acute infection, DENV does not accumulate significant mutations within a single host or outbreak compared to chronic viruses (e.g., HIV, HCV). Consequently, standard consensus sequences often lack the resolution to distinguish between closely related transmission chains.
• Limitations of Current Methods: Traditional phylogenetics describes evolutionary relatedness but fails to provide explicit, probabilistic transmission links between specific individuals. Furthermore, relying solely on epidemiological data (time and location) is insufficient because infected humans often travel significant distances during the incubation period.

The authors aim to develop a framework that integrates within-host viral diversity (haplotypes) with epidemiological and environmental data to estimate the probability of transmission between specific dengue cases.

2. Methodology: The DENcode Framework

DENcode is a probabilistic modeling framework that estimates the relative likelihood ($P_{ij}$) of vector-mediated transmission from a source case ($i$) to a recipient case ($j$). It combines two primary kernels into a unified probability score:

A. Epidemiological Kernel ($K_{ij}$)

This component models the biological plausibility of transmission based on time, space, and vector dynamics. It calculates the probability that a mosquito biting an infectious human ($i$) survives the Extrinsic Incubation Period (EIP) and subsequently bites a susceptible human ($j$).

• Inputs: Symptom onset dates, residential postcodes, and ambient temperature data.
• Mechanisms:
  • Temporal Window: Accounts for the intrinsic incubation period (human) and EIP (mosquito), which is temperature-dependent.
  • Mosquito Survival: Uses a temperature-dependent mortality rate ($\mu$) to calculate the probability of the mosquito surviving the EIP.
  • Spatial Decay: Applies an exponential decay function based on the geodesic distance between cases, reflecting the limited flight range of Aedes aegypti (~106m).
  • Viremia: Incorporates a normalized viremia curve to weight the infectiousness of the source over time.

B. Genetic Similarity Kernel ($G_{ij}$)

This component penalizes genetic dissimilarity between viral sequences.

• Data Source: Uses haplotype sequences (within-host variants) rather than just consensus sequences.
• Metric: Calculates patristic distances (evolutionary distance) from a Maximum Likelihood phylogenetic tree.
• Hotspot Adjustment: To prevent rapidly mutating regions from skewing results, the model adjusts distances by down-weighting mismatches occurring at known mutation hotspots.
• Aggregation: For cases with multiple haplotypes, distances are aggregated and weighted by haplotype abundance to reflect the within-host viral population structure.

C. Probability Calculation

The transmission probability $P_{ij}$ is calculated as:
$$P_{ij} = \frac{K_{ij} \cdot e^{-\gamma G_{ij}}}{\sum_{j' \neq i} K_{ij'} \cdot e^{-\gamma G_{ij'}}}$$
Where $\gamma$ is a scaling parameter. The denominator normalizes the probabilities so that for any source $i$, the sum of probabilities to all potential recipients equals 1.

3. Key Contributions

1. Haplotype-Level Resolution: DENcode is the first framework to explicitly utilize within-host haplotype diversity for dengue transmission inference, demonstrating that this level of resolution is critical when consensus sequences fail to differentiate transmission links.
2. Integrated Probabilistic Framework: It successfully merges clinical metadata, vector biology (temperature-dependent EIP/mortality), and high-resolution genetic data into a single, reproducible pipeline.
3. Validation and Robustness: The model was rigorously validated using 100 Monte Carlo iterations to ensure stability against parameter fluctuations (mosquito mortality, spatial decay, infectiousness).
4. Ablation Analysis: The study quantified the specific contribution of genetic vs. epidemiological data, proving that removing either component significantly alters the inferred transmission structure.

4. Results

The model was validated using a dataset of 90 dengue infections (DENV1, DENV2, DENV3) from the Colombo Dengue Study (CDS) in Sri Lanka (2017–2020).

• Model Stability:
  • Estimates were highly stable across 100 Monte Carlo runs, with a median credible interval width of <0.001.
  • Top-ranked transmission pairs remained consistent even with parameter variations.
• Ablation Experiments:
  • Removing the genetic component or flattening the epidemiological kernel resulted in significant distributional divergence (Kullback-Leibler divergence of 1.61 and 1.76, respectively), confirming that both data streams are essential.
• Network Analysis:
  • DENV1: Showed a compact network (background endemic transmission) with high connectivity even at strict probability thresholds ($\ge 0.5$).
  • DENV2 & DENV3: Represented outbreak scenarios. These networks were larger but lost significant connectivity at higher thresholds, highlighting the difficulty of resolving transmission in large, rapid outbreaks without high-resolution data.
  • Centrality: Identified specific "super-spreader" or bridging nodes (e.g., Node 761 in DENV2, Node 728 in DENV3) that were critical for linking distinct infection clusters.
• Haplotype vs. Consensus Comparison (Crucial Finding):
  • Edge Loss: Consensus-based networks lost 39–100% of the edges found in haplotype-based networks.
  • Systematic Bias: Consensus networks systematically underestimated transmission probabilities (Mean $P_{ij}$ of 0.21–0.40 vs. 0.60–0.98 for haplotypes).
  • Jaccard Similarity: The overlap between consensus and haplotype networks was poor (<0.15), indicating that consensus data fails to capture the true transmission topology.
• Geographical Validation:
  • The model successfully identified a long-distance transmission event (Node 601 in DENV1) linking Colombo to Galle (100km away), consistent with known human mobility patterns (motorway travel), which would be missed by distance-only models.

5. Significance and Implications

• Improved Surveillance: DENcode provides a robust tool for public health officials to identify transmission clusters and "superspreaders" in real-time, even when genetic diversity is low.
• Value of Haplotype Sequencing: The study provides strong evidence that the extra cost and computational effort of resolving haplotypes (vs. consensus) are justified, as consensus data leads to substantial information loss and misidentification of transmission hubs.
• Targeted Interventions: By identifying high-centrality nodes and long-distance links, the model can guide targeted vector control and travel-based interventions.
• Scalability: While designed for Dengue, the pipeline is adaptable to other mosquito-borne viral infections (e.g., Zika, Chikungunya) with parameter adjustments.

In conclusion, DENcode bridges the gap between phylogenetics and epidemiology, offering a high-resolution, probabilistic approach to reconstructing dengue transmission networks that is superior to traditional methods, particularly when leveraging within-host viral diversity.