Relaxing the Infinite Sites Assumption in Tumor Phylogenetics

37 concepts24 papers1 proposal

For decades, computational biology relied on a beautiful, simplifying rule to map the history of evolution: the Infinite Sites Assumption (ISA). The ISA posits that the genome is infinitely large, meaning that a genetic mutation will occur exactly once at any specific site and, once acquired, will never be lost or reversed.

Under the ISA, reconstructing an evolutionary tree is a clean, deterministic puzzle known as the perfect phylogeny problem. If two cells share a specific mutation, they must have inherited it from a common ancestor. If a model assumes permanent, unique mutations, drawing the lines of descent is straightforward.

However, in rapidly evolving systems—most notably within the chaotic genomic landscape of cancer tumors—this elegant assumption breaks down. Recent genomic research has shifted toward exposing why the ISA fails and designing new mathematical frameworks that can tolerate the messy reality of mutation loss and recurrence.

The Three Modes of Violation

When we look at the genomes of individual cells within a single tumor, we frequently find patterns that are mathematically impossible under a perfect phylogeny. As demonstrated by Jack Kuipers et al. in an empirical analysis across multiple cancer types, violations of the unique-mutation assumption are not rare anomalies; they occur in the vast majority of tumor datasets.

These violations typically happen through three distinct biological mechanisms:

Infinite Sites Assumption (Perfect Phylogeny):
Ancestral State [0] ───> Mutated State [1] (Permanent)

Violations in Finite-Sites Models:

1. Mutation Loss (Dollo Parsimony):   [0] ───> [1] ───(Deletion/CNA)───> [0]
2. Parallel Mutation (Homoplasy):     Lineage A: [0] ───> [1]  |  Lineage B: [0] ───> [1]
3. Back Mutation (Reversion):         [0] ───> [1] ───(Point Mutation)───> [0]

1. Mutation Loss via Chromosomal Deletions

Cancer cells are notorious for large-scale genomic instability. When a tumor cell undergoes a Copy Number Alteration (CNA) or a large chromosomal deletion, it can physically wipe out an entire genomic region. If that region contained a previously acquired single-nucleotide variant (SNV), that mutation is lost.

In phylogenetic terms, this introduces the Dollo parsimony problem, where a trait can be lost after being acquired. As noted by Johannes G. Reiter et al., assuming that mutations are never lost forces models to misinterpret these large deletions, distorting our understanding of how metastases seed and spread.

2. Parallel Mutation (Homoplasy)

In highly mutable environments or under intense selective pressure (such as chemotherapy), the exact same nucleotide transition can happen independently in two completely different cell lineages. This convergent evolution at the base-pair level is called homoplasy. When two distinct branches independently hit the same genetic target, an ISA-bound model will incorrectly force them into the same branch, inventing a shared ancestry that never existed.

3. Back Mutation

A site can literally revert to its ancestral state through a subsequent point mutation. While rare in stable organisms, Filippo Utro et al. demonstrated that back-mutations are a shared challenge when tracking both intra-patient tumor clones and rapidly mutating viral sublineages, such as SARS-CoV-2.

The Algorithmic Cost of Assuming Perfection

What happens if we ignore these biological realities and force an ISA-compliant algorithm to process real-world tumor data? The model breaks down in predictable, destructive ways.

First, the algorithm overestimates evolutionary complexity. When an identical mutation appears independently due to homoplasy, or when a mutation is lost, standard compatibility tests flag these discrepancies as genetic recombination events. As Qi Zhang et al. showed, this sensitivity causes models to overestimate the minimum number of recombination breakpoints needed to explain the data, mistaking recurrent mutations for structural crossovers.

Second, technical noise compounds the error. Modern tumor phylogenetics relies heavily on single-cell DNA sequencing (scDNA-seq). Single-cell data is notoriously noisy, suffering from a phenomenon called allelic dropout, where one of the two alleles at a heterozygous site fails to amplify, making a mutated site look wild-type (ancestral).

If a model operates under the ISA, it faces an impossible choice when encountering a cell missing a mutation: it must either assume a massive, improbable technical error occurred, or it must distort the entire tree topology to accommodate the missing data. As Kiran Tomlinson et al. evaluated, this rigid constraint prevents standard phylogenetic algorithms from finding any valid evolutionary trees in many real-world bulk and single-cell cancer datasets.

The Solution: Finite-Sites and Loss-Aware Frameworks

To resolve this, researchers have developed a new generation of probabilistic frameworks that replace the infinite sites assumption with a finite-sites evolutionary model.

Traditional Model (ISA)             Finite-Sites Model (e.g., SIEVE / SCIΦN)
      [0]                                     [0]
       │                                     ┌─┴─┐
      [1]                                   [1] [0] (Parallel Mutation)
     ┌─┴─┐                                   │   │
    [1] [1]                                 [0] [1] (Mutation Loss)

By allowing a single genomic position to undergo multiple state transitions over time (e.g., $0 \rightarrow 1 \rightarrow 0$), these models can mathematically evaluate whether a missing mutation is the result of a technical artifact or a true evolutionary event.

Distinguishing Artifacts from Evolution

The core innovation of frameworks like SCIΦN, developed by Jack Kuipers et al., lies in the likelihood function. By incorporating the probabilities of both technical allelic dropout and biological mutation loss into a unified Bayesian framework, the algorithm can look at surrounding context to make an informed bet. If a mutation is missing in a single cell alongside high read coverage, the model infers a biological deletion; if the coverage is zero, it infers a technical artifact.

Similarly, the SIEVE framework introduced by Senbai Kang et al. bypasses binary mutation calls entirely. Instead, it operates directly on raw sequencing read counts, modeling multiple genotype transitions at the same site. This allows it to accurately reconstruct lineages in high-mutation environments where sites mutate more than once.

Moving Beyond Diploid Assumptions

Many older models assumed that cells maintain a stable, diploid (two-copy) genome. However, Senbai Kang et al. expanded this paradigm with DelSIEVE by explicitly pairing single-nucleotide variants with deletion events. By acknowledging that a deletion can physically remove an SNV, the model avoids misinterpreting deletions as rare, complex double-mutation events.

On a broader scale, Etienne Sollier et al. introduced COMPASS to jointly reconstruct trees using both point mutations and copy number alterations simultaneously, capturing complex events like whole-genome duplications that violate traditional monotonic evolutionary assumptions.

Why It Matters

Shifting from the infinite sites assumption to finite-sites modeling is not just a mathematical exercise in error reduction; it fundamentally changes our understanding of cancer progression.

When we use rigid models that forbid mutation loss or recurrence, we get highly complex, fragmented trees that obscure the true trajectories of a tumor. By allowing the tree to breathe—permitting mutations to be lost via deletions or to recur independently via parallel evolution—frameworks like TreeMHN, designed by Jack Kuipers et al., can uncover true mutational exclusivity and recurrent evolutionary trajectories.

Understanding these real trajectories is what ultimately allows clinicians to predict how a tumor will adapt, move, and respond to therapeutic pressure.

Key Foundational Work

Inferring Genome-Wide Mosaic Structure (2009) — contributed 3 concept elements to this direction.
Relationship between the selection pressure and the rate of mutation accumulation (2008) — 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_infinite_sites_assumption_9f411b73,
  title = {infinite sites assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-sites-assumption/rdr_9f411b73db5d47dd9d58807c0994a363},
  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_infinite_sites_assumption_9f411b73_v1,
  title = {infinite sites assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-sites-assumption/rdr_9f411b73db5d47dd9d58807c0994a363/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