Accounting for Phylogenetic Uncertainty in Biological Inference

42 concepts34 papers1 proposal

Many computational biology workflows share a common, hidden vulnerability: they treat evolutionary history as a solved puzzle. Whether predicting protein function, tracking tumor lineages, or delimiting species boundaries, researchers typically begin by computing a single "best" phylogenetic tree from sequence data, and then treat that tree as an immutable ground truth for all subsequent analysis.

This two-step pipeline creates a dangerous bottleneck. As noted across diverse domains—from functional genomics (Engelhardt et al. (2005)) to ancestral sequence reconstruction (Sanz Moreta et al. (2022))—assuming a fixed tree topology ignores phylogenetic uncertainty. If the initial tree reconstruction contains even minor errors, those errors propagate downstream, leading to overconfident and systematically biased biological conclusions.

Why Fixed Trees Break

Phylogenetic reconstruction is highly sensitive to noise, sampling limitations, and model misspecification. Forcing a downstream model to rely on a single point-estimate tree is risky for several reasons:

  • Model Mismatches: The underlying mathematical models used to build trees can be wrong. For instance, Nickle et al. (2007) demonstrated that using mismatched evolutionary substitution models can distort branch length estimations by up to 30% and yield incorrect topologies.
  • Data Noise as "Tree-Breakers": Small errors in data processing can completely alter a tree's structure. In cancer genomics, Reiter (2017) showed that a tiny fraction of misclassified genetic variants can disrupt the logical structure of a tumor phylogeny, leading to fundamentally incorrect conclusions about how metastases spread. Downstream statistical frameworks that map clonal mutations are acutely vulnerable to these upstream topological errors (Kuipers et al. (2021)).
  • Weak Phylogenetic Signals: When sequence data provides weak or conflicting signals, forcing an algorithm to choose one "optimal" tree is an arbitrary exercise (Rasmussen (2012)). In reality, there is often an astronomical number of equally parsimonious evolutionary histories (Dumm et al. (2023)). Selecting just one discards this vital context, causing downstream models to overfit to a potentially spurious topology.

The Cascade of Downstream Errors

When a model treats an uncertain tree as absolute truth, the statistical consequences are severe across multiple applications:

[Noisy Sequence Data] 
       │
       ▼
[Single "Best" Tree]  ◄─── Upstream Errors (Misspecified models, sequencing noise)
       │
       ▼
[Downstream Inference] ◄─── Brittle Propagation (Overfitting, false positives)

In species delimitation, relying on a single fixed tree ignores variations in branch lengths and topologies, causing models to deliver highly inaccurate species counts with misleadingly high statistical confidence (Reid (2012); Kapli et al. (2017)).

In functional genomics, mapping gene presence or structural traits onto an incorrect reference tree creates false positives, misidentifying isolated evolutionary events as shared functional relationships (Li et al. (2018); Using Phylogeny to Improve Genome-Wide Distant Homology Recognition (2007)). Similarly, coalescent models that treat the genealogical tree as a fixed input underestimate the variance in historical population sizes, leading to overconfidence in population dynamics (Cappello et al. (2023)).

Embracing Uncertainty Probabilistically

To solve this problem, a major research direction focuses on integrating over tree space rather than optimizing for a single point estimate. The core intuition is that if we do not know the exact tree, we should average our downstream predictions across all likely trees, weighted by their probability.

Bayesian Integration over Posteriors

Instead of a single tree, Bayesian frameworks treat the evolutionary tree as a latent probability distribution. For example, Reid (2012) implemented a Bayesian approach that integrates over a posterior distribution of topologies and branch lengths, providing a much more robust and cautious assessment of species boundaries. Similarly, Li et al. (2018) demonstrated that treating the underlying tree structure as a probability distribution prevents evolutionary clustering models from overfitting to noisy genomic data.

Compact Ensemble Representations

When the space of possible trees is too vast for simple Markov chain Monte Carlo (MCMC) enumeration, researchers use specialized data structures to track ensembles of likely histories. Zoller et al. (2019) introduced probabilistic hypergraphs to capture the $k$-best gene-to-species tree reconciliations, ensuring that noise does not obscure complex evolutionary events like horizontal gene transfer. To handle massive datasets, Dumm et al. (2023) developed the "history sDAG" (subpartition directed acyclic graph), which uses subhistory swapping to compactly represent millions of equally plausible trees, establishing a formal lower bound for phylogenetic uncertainty.

Joint Modeling and Representation Learning

The most complete shift bypasses the two-stage pipeline entirely by performing joint inference. Instead of generating a tree and passing it to a downstream model, the evolutionary constraints and the downstream tasks are optimized simultaneously.

For instance, Sanz Moreta et al. (2022) developed Draupnir, a deep generative model that embeds a tree-structured Ornstein-Uhlenbeck process directly into a Variational Autoencoder (VAE) prior. This allows the model to learn sequence representations and evolutionary history concurrently, preventing the topological bottlenecks that plague traditional ancestral sequence reconstruction workflows. By treating the tree as a dynamic, learnable component rather than a static constraint, these methods ensure that downstream biological insights remain robust to the inherent noise of evolutionary history.

Key Foundational Work

Protein Molecular Function Prediction by Bayesian Phylogenomics (2005) — contributed 1 concept elements to this direction.
Reconstructing Ancestral Gene Orders Using Conserved Intervals (2004) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_phylogenetic_uncertainty_6d201fdb,
  title = {phylogenetic uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/phylogenetic-uncertainty/rdr_6d201fdb596243f399348ce4f8ed22a0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_phylogenetic_uncertainty_6d201fdb_v1,
  title = {phylogenetic uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/phylogenetic-uncertainty/rdr_6d201fdb596243f399348ce4f8ed22a0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake