Joint Inference of Phylogeny and Genomic States

39 concepts23 papers1 proposal

In genomic biology, we are often tasked with solving two problems simultaneously: identifying the precise genetic states of individual samples (such as mutations, copy numbers, or genotypes) and reconstructing the evolutionary tree (phylogeny) that connects them.

Traditionally, these tasks have been treated as a sequential pipeline. First, raw sequencing data is processed using statistical heuristics to call mutations or copy number variants. Second, these discrete calls are fed into a tree-building algorithm to map out their evolutionary history.

However, this sequential approach suffers from a fundamental flaw: error propagation. Biological sequencing data is notoriously noisy. Single-cell sequencing, for instance, is plagued by technical artifacts like allelic drop-out, false negatives, and low coverage. If an early-stage pipeline misidentifies a mutation due to noise, that error is permanently locked in, distorting the downstream phylogenetic tree. Conversely, forcing a tree topology onto flawed mutation calls can introduce spurious evolutionary events that never occurred.

To break this bottleneck, a powerful paradigm has emerged: joint phylogenetic inference. By treating both the evolutionary tree and the underlying genomic states as latent variables within a single, unified probabilistic framework, researchers can leverage the mathematical dependency between history and state to resolve ambiguities that are impossible to untangle in isolation.

The Bidirectional Flow of Information

The core reason joint inference works is that it establishes a bidirectional flow of information between the raw data and the evolutionary model. Instead of a one-way pipeline, the emerging tree structure acts as a statistical regularizer for the mutation calls, while the raw data continuously refines the tree topology.

Sequential Pipeline (Error Propagates):
Raw Data ──> Mutation Calling ──> Tree Reconstruction

Joint Inference (Mutual Refinement):
Raw Data <───> [ Joint Probabilistic Model ] <───> Tree Topology

As Yuan et al. demonstrate, avoiding separate deconvolution and tree-building steps yields a far more integrated and accurate representation of clonal evolution. When these tasks are performed simultaneously, the shared ancestry of related entities provides a powerful mathematical constraint.

Kuipers et al. showed that integrating tree reconstruction directly into the mutation calling process yields a vastly superior balance of precision and recall compared to sequential methods. In their framework, the model allows the evolving topology to guide the interpretation of noisy sequencing reads. Similarly, Kang et al. utilized a probabilistic graphical model to simultaneously estimate tree topology, branch lengths, and cell genotypes. Because information flows both ways, the phylogenetic structure helps filter out stochastic sequencing noise, while the raw read counts directly optimize the tree structure.

This joint optimization is not limited to single trees. In scenarios involving reticulate evolution—such as recombination or hybridization—Matsen demonstrated that multiple related trees can be reconstructed simultaneously by enforcing topological distance constraints between them. This joint constraint prevents experimental noise in individual genomic segments from introducing spurious evolutionary differences.

Phylogenetic Imputation: Using Ancestry to Fill in the Blanks

Why does an evolutionary tree make data cleaner? The answer lies in the principle of shared ancestry: closely related nodes are highly likely to share specific biological attributes. If technical errors or low sequencing coverage obscure the signal in a particular sample, the states of its evolutionary neighbors can be used to "fill in the blanks." This mechanism is known as phylogenetic imputation.

    [Ancestral Node: Mutation Present]
               /         \
              /           \
 [Sister Node]             [Target Node: Zero Coverage]
 (Mutation Observed)        (Imputed: Mutation Present)

In viral surveillance, missing data frequently complicates the identification of closely related strains. Pipes et al. showed that leveraging a global evolutionary tree allows for the accurate imputation of missing nucleotides in wastewater sequencing reads, achieving error rates comparable to or lower than typical sequencing errors.

In cancer genomics, where single cells often exhibit zero coverage at critical genomic sites, the tree serves as a robust filter. Kuipers et al. demonstrated that by analyzing the mutation status of ancestral or sibling cells within an inferred tree, a model can confidently call mutations that would otherwise be lost to technical noise.

This shared signal also helps distinguish true biological events from random artifacts. For example, Kang et al. showed that a genomic deletion observed across multiple related cells provides high-confidence evidence of a heritable change, allowing the model to validate true deletions and dismiss isolated technical drop-outs. This collaborative analysis across related sequences is equally effective in immunology; Ralph demonstrated that simultaneous inference on groups of related B-cell receptor sequences allows models to effectively differentiate between inherited germline nucleotides and stochastic somatic hypermutations.

Resolving Physical and Structural Dependencies

Joint inference is particularly vital when analyzing genomic features that are physically or biologically dependent on one another. A classic example is the relationship between single-nucleotide variants (SNVs) and copy number alterations (CNAs).

The observed frequency of a point mutation in a sequencing sample is directly dependent on the local genomic copy number. If a chromosome segment is duplicated or deleted, the ratio of mutated to normal alleles shifts dramatically. As Liu et al. point out, analyzing these features in isolation creates severe ambiguities when estimating subclone prevalence. By integrating SNVs and CNAs into a unified statistical framework, the physical dependency between local copy numbers and point mutations is explicitly modeled, resolving subclonal architecture with much higher fidelity.

To exploit this, the authors of these frameworks use different genomic signals as statistical anchors for one another:

  • Point Mutations as Anchors for Copy Numbers: Sollier et al. developed a joint approach that utilizes total read counts for copy number estimation and variant allele frequencies for point mutations. By correlating frequency shifts with depth variations, the model uses point mutations as statistical anchors to distinguish true copy number changes from stochastic coverage fluctuations, enabling the detection of complex events like copy-neutral loss of heterozygosity.
  • Copy Numbers as Anchors for Point Mutations: Conversely, when dealing with ultra-low coverage single-cell data where SNV signals are incredibly sparse, Weber et al. demonstrated that the large-scale, robust signals of copy number profiles can be used to guide and constrain the reconstruction of fine-grained, SNV-based evolutionary relationships.

This principle extends to global sample characteristics like tumor purity, ploidy, and cell fractions. Yau et al. showed that treating tumor purity and ploidy as latent variables to be inferred simultaneously—rather than applying sequential heuristic corrections—prevents the propagation of errors to final genomic calls. Similarly, Cmero et al. utilized a Bayesian mixture model to simultaneously estimate mutated copy numbers and the fraction of cells harboring a specific structural variant, accounting for the tight statistical interdependence of these variables. In clinical settings, Erdmann-Pham et al. demonstrated that jointly estimating tumor purity and absolute copy number profiles using parsimony-based modeling effectively distinguishes low-frequency malignant signals from experimental noise.

Anchoring Multi-Modal and Multi-Region Data

Finally, joint inference provides a mathematical bridge to combine entirely different data modalities or spatial regions, using the shared evolutionary history as the common denominator.

When analyzing multi-region biopsies, Antonello notes that treating multiple samples from the same patient as independent observations fails to leverage their shared history; incorporating joint evolutionary modeling across distinct samples constrains copy number estimates based on common ancestral lineages.

This anchoring effect is beautifully illustrated when combining bulk sequencing (which offers high-depth, robust mutation frequencies but lacks cellular resolution) with single-cell sequencing (which offers clear cellular resolution but suffers from high error rates).

Bulk Sequencing Data ────────┐
(Provides Temporal Ordering) │
                             ├─> [ Joint Likelihood Model ] ──> High-Fidelity Tree
                             │
Single-Cell Sequencing Data ─┘
(Provides Branching Structure)

Malikić et al. developed a joint likelihood model that integrates these two data types. The bulk data provides a robust temporal ordering of mutations based on frequency, which anchors the noisy branching signals found in single-cell data. By using the strengths of one modality to anchor the weaknesses of the other, the joint model achieves a high-fidelity reconstruction that neither data type could produce alone. Following a similar logic, Jun et al. integrated bulk DNA sequencing with single-cell RNA sequencing (scRNA-seq). By modeling raw read counts directly within a joint Bayesian framework, their method accounts for biological noise like bursty expression and mono-allelic transcription, successfully linking specific genetic lineages to functional gene expression profiles.

By moving away from rigid, step-by-step pipelines and embracing unified probabilistic models, joint phylogenetic inference transforms evolution from a mere downstream output into a powerful structural prior that cleans, validates, and illuminates noisy genomic data.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_joint_phylogenetic_inference_7e912816,
  title = {joint phylogenetic inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-phylogenetic-inference/rdr_7e912816ce7d40569bc32cf562f0ebe0},
  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_joint_phylogenetic_inference_7e912816_v1,
  title = {joint phylogenetic inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-phylogenetic-inference/rdr_7e912816ce7d40569bc32cf562f0ebe0/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