Cluster Tracker

The heat map represents the total number of introductions into a region. Once you click a region it will turn purple indicating it is now the focal region for the map. All the other county and state regions will become colored in terms of the number of introductions that come from that region, with respect to the focal region. The coloration can be changed from the default log-fold enrichment, to aid in spotting introductions from less densely populated areas, or raw cluster count, where the most populous regions will likely have the highest raw count.

To understand how the heat map is connected to the table clusters, once you have selected the focal region, the table will now be filtered for that region. The heat map does not represent a confidence score of the introductions, from other regions, just the number of introductions coming from these other counties or states. In the rows on the table, a "Best Potential Origins" column has a name or names corresponding to the regions in the map. The heat map, in essence, is a global view of introductions, rather than a representation of any one specific cluster in a row. In essence, it shows the sum of the potential of all the clusters from each region into the focal region.

The shading of the map with our log-fold enrichment scale is an attempt to mitigate sampling biases resulting from larger populations, higher case rates, increased sequencing, or other factors that are not specific to geography.

Intuitively, the less sequencing is performed in a region, the less likely it will be possible to detect sequences from that region when they are an introduction source to another region. In order to attempt to compensate for this statistical bias the log-fold enrichment calculations of introductions between regions is computed and displayed as the default view. It is possible to switch the view to raw cluster count, however, that will always color sources of high population darker.

The log-fold enrichment calculation is described in the Methods section of the paper describing Cluster Tracker.

In this formula, Iab is introductions from region A to region B, Ixx is introductions from any region to any region, Iax is introductions from region A to any other region, and Ixb is introductions from anywhere to region B. This computation can remove biases in rates of detected introduction which would apply to any pair of regions, but requires many regions to be computed as points of comparison. Log10fold enrichment is used to color the map on Cluster Tracker when a state is selected. Often the coloring has a very strong correlation with geographic distance, which also makes intuitive sense. It is worth stressing that while log-fold enrichment may reveal spatial relationships, it does not reflect the absolute importance of a region as a source or sink of viral transmission.

You can enter multiple terms to search the table of clusters, provided each term is separated by a comma - do not use a space. For instance, if you were looking for rows mentioning Hawaii and Texas you could enter "Hawaii,Texas" and click "Search" and find all the rows in the table that have both of those terms.

If you wanted to further narrow the results, you could add more terms such as a variant term or dates. For instance, if you were curious about rows that included both November and December dates for the rows that include mention of Hawaii, Texas, and the variant XBB you could add  XBB followed by a comma, and then the date 2022-12 for December followed by a comma and 2022-11 for November.

A search such as "Hawaii,Texas,XBB,2022-12,2022-11" will identify the rows in the table where each of these terms exists. Clicking the "Search" button does a Boolean "AND" match by default, however, it can be changed to the "OR" option to expand results. For instance, if you change to the "OR" search and put in many dates, such as "2022-12,2022-11,2022-10,2022-09" you would find any rows that mention any one of those dates once you click "Search".

A "Search Columns:" box allows the removal of columns to further restrict the matched results. For instance, if you searched the term "County" and restricted the "Search Columns:" box to only "Cluster ID" you would filter all California County level information (i.e., remove any clusters from other states). Using the "Region" column rather than "Cluster ID" would provide similar results. Note you can achieve a similar result by clicking the top "Show CA State Introductions" and then click the state of California on the map -while this would not include the county-level metadata.

The Advanced Filter Options are an even better way to drill down on cluster dates, as well as narrow down on specific cluster sizes and growth scores.

Advanced Filter Options

Additional "Cluster Date:" and "Cluster Size:" and "Growth Score:" filter options allow drilling down on clusters that are more recent and potentially more concerning. For instance, perhaps you want to only filter the clusters of 3 or more samples that have been collected in the last month.

First you could click your region of interest on the map, for instance San Francisco County. Or us the above "County" search restricting the "Search Columns:" box to only "Cluster ID" to see all counties. Then you click the side arrow next "Advanced Filter Options" to reveal date, size, and score range options. You can just put a "From: mm/dd/yyyy" date representing the last 30 days. If you click the calendar icon, a pop-up will allow clicking dates rather than typing them. Now in the "Cluster Size:" filter option only put "3" in the "From:" box, leaving the other box empty. Now click "Search" in the top right. If you have no results, move the date further back -often there is a lag in sequencing data entering into the pipeline.

Each time you change the inputs, be sure to click "Search" to have the filter applied. You can apply a "To: mm/dd/yyyy" to limit the date range, and "To:Max" for cluster size. The Growth Score is a calculation derived from both date and sample size. You can use it to further filter, or start filtering samples with Growth Score once you get a feeling of how it highlights more significant clusters.

When looking at a phylogenetic tree, it is important to know that even if two branches are close to each other, the samples may or may not have a relationship to each other. It can be hard to say where a variant may have come from and where it may be going by just looking at a tree. When possible, additional epidemiological data should drive further investigation, and help clarify possible connections. Yet that does not mean it is impossible to make connections with just a phylogenetic tree.

To understand the relationships for a node in a phylogenetic tree, one normally would desire a binary tree, where each node has exactly two descendants, and there is the opportunity to trace back up the tree to an original state related to a leaf node of interest. In many cases, however, there are polytomies in a tree, where one node will have three or more child subtrees, which is also described as a "multifurcation" leading to situations where it is harder to extract relationships expected in a normal binary tree.

Even with a binary phylogenetic tree, it can still not be clear if perhaps a variant originated in one region and traveled to the other, or if the reverse travel could have equally happened. Similarly it is possible by chance the same variant could have emerged in two different locations and the tree could be demonstrating convergent evolution, where two lineages evolved independently and end up looking related.

With these caveats in mind, however, the data in the phylogenetic tree can still be the first stop to make sense of sequences that are highly similar and appear to be related. Cluster Tracker's mutation analysis utilities, or matUtils, can extract information to label the best potential origins for introductions. The matUtils do this by crunching data about the number of samples on a branch, and the number of mutations between samples, calculating a regional index score to help assign if descendents were inside or outside a focal point. 

By looking at the top phylogenetic tree without regional index scores, one can picture that some samples farther to the right are most likely less related to the top-circled CA sample from California. For instance, the bottom branch with the MA sample from Massachusetts, would intuitively not seem likely to be an origin for the virus in California. Likewise, one probably would guess the two adjacent NY samples from New York could be the best potential origin for the CA virus.

The number crunching by matUtils, on the other hand, which takes in variables Li and Lo to refer to the number of leaves (samples) that are inside or outside when looking at a focal node, and Di and Do, to refer to the total branch lengths to leaf descendents in terms of mutations, actually calculates that the three samples in Virginia as having more weight (0.23 vs 0.15) compared to the two samples from New York. While all hypothetical, these heuristic calculations help input how more time has passed for more mutations to happen, and take into account when more samples exist, there is more likelihood a branch is an introduction source or not for another region.

In these ways, the potential origin calculations can help direct hypothesis generation and prioritization. 

While still a hunch, since more samples are from Virginia and the mutation distance is not too far, compared to say the Massachusetts sample, it may make more sense to start investigating additional epidemiological data from Virginia when looking from the California sample, alongside the potential New York connections. It is important to keep in mind it is equally possible the variant started in one of the other branches and a reverse interpretation is possible, and also to keep in mind regions outside of the US are not used in regional index calculations. This real world data example helps demonstrate how one can interpret potential origins calculations to make sense of the many possible places to start an investigation.