Single-cell sequencing (SCS) promised to revolutionize our understanding of cancer by revealing the exact order of mutations in a tumor's history. However, the raw data from these experiments is notoriously "noisy." Because a single cell contains only two copies of DNA, the chemical amplification required for sequencing often fails to capture one of them—a phenomenon called allelic dropout (ADO). This creates a massive missing data problem: a mutation might be present in a cell, but the sequencer simply doesn't see it.
To solve this, researchers have moved away from treating single cells as independent data points. Instead, they use the shared evolutionary history of cells to "fill in the blanks," turning a liability of noise into a problem of probabilistic inference.
The Failure of the Infinite Sites Assumption
Traditional models of evolution often rely on the Infinite Sites Assumption (ISA), which posits that a specific genomic location never mutates more than once. Under the ISA, once a mutation occurs, it is passed down to all descendants forever.
However, work by Jack Kuipers, Katharina Jahn, and Niko Beerenwinkel has shown that this assumption is frequently violated in tumors. In Single-cell sequencing data reveal widespread recurrence and loss of mutational hits, they demonstrated that mutations are often lost (via genomic deletions) or recur independently in different lineages. Distinguishing these biological "back-mutations" from technical artifacts like ADO is the central challenge of the field.
Joint Inference as a Solution to Noise
If you look at a single cell and don't see a mutation, you don't know if the mutation is truly absent or if it was just a victim of ADO. But if you look at a hundred cells that all seem to be closely related, and ninety of them have the mutation, the probability that the other ten also have it (but suffered ADO) becomes very high.
This intuition led to the development of frameworks like SCITE (Tree Inference for Single-Cell Data) and SCIΦ (Single-Cell Mutation Identification via Phylogenetic Inference). These tools don't just "call" mutations and then build a tree; they perform joint inference. They simultaneously search for the most likely tree structure and the most likely set of true mutations, using the tree itself as a prior to correct the errors in individual cells.
Integrating Bulk and Single-Cell Data
While single-cell data provides high resolution for branching, bulk sequencing (sequencing thousands of cells at once) provides much more accurate measurements of how common a mutation is across the whole tumor.
Salem Malikić and colleagues developed B-SCITE (Integrative inference of subclonal tumour evolution from single-cell and bulk sequencing data) to combine these strengths. By using bulk data to "anchor" the frequencies and single-cell data to define the "branches," B-SCITE produces far more reliable evolutionary models than either method could alone.
Beyond Point Mutations
The latest generation of tools has expanded this probabilistic approach to more complex genomic changes. SIEVE and DelSIEVE, developed by Senbai Kang and Nico Borgsmüller, explicitly model both single-nucleotide variants and large-scale deletions. By using raw read counts rather than binary "present/absent" calls, these models can distinguish between a technical "locus dropout" and a biological deletion (DelSIEVE: cell phylogeny model of single nucleotide variants and deletions).
Similarly, tools like Monopogen (Jinzhuang Dou) leverage population-level genetic patterns to improve variant calling in sparse transcriptomic data, allowing researchers to track lineages even when the sequencing depth is extremely low.