{"id":29,"date":"2025-10-10T15:36:03","date_gmt":"2025-10-10T15:36:03","guid":{"rendered":"https:\/\/erik.mvolz.com\/?p=29"},"modified":"2025-10-10T16:26:23","modified_gmt":"2025-10-10T16:26:23","slug":"can-we-predict-evolution-from-a-phylogeny","status":"publish","type":"post","link":"https:\/\/erik.mvolz.com\/?p=29","title":{"rendered":"Can we predict evolution from a phylogeny?"},"content":{"rendered":"\n

Summary<\/h1>\n\n\n\n

Our preprint asks whether you can spot tomorrow\u2019s “up-and-coming” lineages by reading patterns already encoded in today\u2019s phylogenetic tree. We introduce a simple statistic– coalescent odds<\/strong>— that scores each lineage\u2019s tendency to spawn descendants. It\u2019s fast, interpretable, and designed for messy real-world surveillance. Here’s the preprint<\/a>. And, here’s the R package. <\/a><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Why this matters<\/h2>\n\n\n\n

Epidemiologists sift through pathogen genomes to catch variants early. But turning millions of sequences into an early-warning signal is hard. We asked a basic question: does the shape of the tree itself contain enough signal to flag which lineages are likely to grow?<\/strong><\/p>\n\n\n\n

This isn’t a new problem, and ours isn’t the first proposed solution. For example, the local branching index (LBI)<\/a> is also a simple statistic which can be readily calculated from a tree, and which is the default statistic reported on nextstrain as a proxy for pathogen fitness, e.g. here<\/a>. There’s also a lot of related work to pickout “clusters” or cryptic population structure<\/a> in phylogenies, and these can be used as the basis for epidemic early<\/a> warning<\/a> signals.<\/p>\n\n\n\n

But we had another look at this problem to see if we can come with a method that checks some boxes:<\/p>\n\n\n\n

    \n
  1. Can we define a statistic that is readily interpretable and based on a classic population-genetic model?<\/li>\n\n\n\n
  2. Can we make it work with messy real-world data that is rife with sampling bias?<\/li>\n\n\n\n
  3. Can we make it scalable to very large real-world datasets?<\/li>\n\n\n\n
  4. Can we fit the model to data w\/o relying on arbitrary hyperparameters (e.g. clustering thresholds)?<\/li>\n<\/ol>\n\n\n\n

    The core idea: coalescent odds<\/h2>\n\n\n\n

    Funnily enough, one of the most useful ways to study phylogenies is to read them ‘backwards’ in time.
    Branches coalesce<\/strong> when two lineages share a recent ancestor. If a lineage tends to sit near lots of recent coalescent events, it may be on a growth trajectory. Or it may just be in a region that is highly sampled! It’s important to have a method that can distinguish between these mechanisms.<\/p>\n\n\n\n

    We define a continuous, heritable “propensity to coalesce”<\/strong> for each lineage. Intuitively, it\u2019s a score for “how likely this lineage is to be the parent of many near-future samples.” From this, we compute coalescent odds (cod’s)<\/strong>-a per-lineage number that can be tracked over time.<\/p>\n\n\n\n

    In principle, this gives us a model-light<\/em> approach, without heavy assumptions about evolution or fitness. Coalescent odds can vary in a flexible way across a phylogenetic tree, and we fit the model in a way that maximises the predictive power of the statistic for future growth.<\/p>\n\n\n\n

    While the basic model is not terribly scalable (certainly not in comparison to methods that use message-passing algorithms), we developed a number of very efficient and accurate approximations that make this applicable to trees with 000’s of samples.<\/p>\n\n\n\n

    What we found (in brief)<\/h2>\n\n\n\n
      \n
    1. Prediction.<\/strong> Given a tree built from genomes available today<\/em>, can cod’s help identify lineages that will be more common tomorrow<\/em>?\n