💻 bioinformatics

CloneVI: A deep generative model for deciphering the relationships between cell state and mitochondrial genotype

CloneVI is a deep generative model that infers clonal structures from single-cell gene expression or chromatin accessibility data to decipher the relationships between mitochondrial DNA mutations and specific cell states in various disease contexts.

Original authors: Hibi, T., Kojima, Y., Hayashi, S., Hisose, H., Mizukoshi, C., Shimamura, T.

Published 2026-11-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hibi, T., Kojima, Y., Hayashi, S., Hisose, H., Mizukoshi, C., Shimamura, T.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside every living cell, there is a complex history written in two different languages. One language is the genetic code, the DNA that acts as a family tree, tracing which cells are related to one another through shared mutations. The other language is the cell's current state, a snapshot of what the cell is doing right now, such as whether it is fighting an infection or dividing to create more tissue. For a long time, scientists could read these two languages separately. They could map the family trees of cancer cells by looking at their DNA, or they could observe what a cell was doing by measuring its activity. But understanding how a cell's family history shapes its current behavior has been difficult because the tools to read both languages at once were often too blunt. They could see the family tree, but they struggled to see how that tree influenced the cell's daily life, or vice versa.

A new tool called CloneVI, developed by researchers at Nagoya University and other institutions, changes this by learning to read both languages simultaneously. The researchers focused on a specific type of genetic marker found in mitochondria, the tiny power plants inside cells. These markers, which are small changes in mitochondrial DNA, act like natural tags that can identify which cells belong to the same family line. The challenge was that these tags are often scattered and hard to find, and the cells carrying them might be doing very different things. The team built a deep learning model that does not just look for the tags to build a family tree. Instead, it first learns to understand the cell's current activity and then asks: given this activity, which family tree is this cell most likely to belong to? This approach flips the traditional method on its head, allowing the model to see how genetic families are distributed across different types of cell behavior.

The researchers tested this idea first with computer simulations, creating artificial data where they knew the exact answers. In these tests, the model successfully reconstructed the hidden family structures, even when different families were mixed together in the same group of cells. It could tell that a cell belonged to one family with high confidence, or that it was in a gray area where two families overlapped, assigning a probability to each possibility rather than forcing a single, rigid choice. This ability to handle uncertainty is crucial because real biological systems are rarely black and white.

When the team applied CloneVI to real data from patients with blood disorders, the results were striking. In one dataset involving cells from a patient with clonal hematopoiesis, a condition where blood cells accumulate mutations, the model identified specific families of cells that were not just defined by their mutations, but by what they were doing. It found that certain mutant families were heavily concentrated in specific types of immune cells, particularly T cells that were actively fighting. The model could predict the presence of these mutations in new cells just by looking at their activity patterns, outperforming older methods that relied on simple averages. It showed that the family history of a cell is deeply tied to its current function, with some mutant families thriving only in specific cellular environments.

The researchers also tested the tool on data from chronic lymphocytic leukemia, a type of blood cancer, using a different method to measure cell activity that looks at how tightly the DNA is packed. Here, the model distinguished between two groups of cells that carried the exact same mutation but had different levels of that mutation inside them. This is similar to having two branches of a family that share a surname but have different amounts of a specific heirloom; the model could tell them apart based on their activity. It linked one of these groups to increased access to a gene called PAX9, which has been associated with the progression of the disease. This finding suggests that the model can connect a specific genetic lineage to a specific biological mechanism driving the illness.

In a large study involving ovarian tumors from five different donors, containing over 100,000 cells, the power of the approach became even clearer. The model identified mutant cell populations that were clustered in specific states, such as immune cells in a state of exhaustion, a condition where they are worn out from fighting cancer. By weighting the analysis based on the likelihood of a cell belonging to a mutant family, the researchers could pinpoint exactly which biological programs were active in those mutant cells. They found that these exhausted mutant cells showed a distinct pattern of activity that was different from non-mutant cells, revealing the specific molecular signature of the disease's progression.

The study demonstrates that by starting with what a cell is doing and working backward to its family history, scientists can uncover relationships that were previously hidden. The model does not force cells into rigid boxes; instead, it maps out a landscape where family lines and cell states overlap and interact. This provides a new way to understand how genetic changes drive the diversity of behavior seen in diseases like cancer. The researchers have made the software available to others, offering a new lens through which to view the complex interplay between a cell's past and its present.

Technical Summary of CloneVI

Problem Statement
Intratumoral heterogeneity arises from both genetic lineages (clones) and non-genetic regulation of cell states. Understanding how clones shape disease requires knowing which cell states are tied to specific lineages. While single-cell multi-omics technologies can simultaneously measure mitochondrial DNA (mtDNA) mutations and gene expression or chromatin accessibility, current analytical approaches typically infer clonal structure solely from mutations and subsequently map cell states to these pre-defined clones. This "genotype-first" approach has a critical limitation: if cells of the same genetic clone exist in diverse states, the specific states characterizing that clone may be obscured. Furthermore, existing methods often treat clones as discrete, binary assignments, failing to capture the continuous nature of cell-state transitions or the probabilistic overlap of clones within specific states.

Methodology: CloneVI
The authors propose CloneVI, a deep generative model based on a Variational Autoencoder (VAE) that integrates genetic and non-genetic heterogeneity. Unlike previous methods, CloneVI infers clonal structure directly from the latent cell state derived from non-genetic data (gene expression or chromatin accessibility), rather than from mutations alone.

The model operates through the following components:

Encoder: Compresses non-genetic input data ( $x_c$ , e.g., gene expression counts or chromatin accessibility) into a low-dimensional latent variable ( $z_c$ ), representing the "cell state."
Clonal Probability Decoder: Estimates the probability ( $\pi_c$ ) that a cell belongs to each of $K$ inferred clones based solely on the latent cell state $z_c$ . This creates a continuous landscape of clonal probabilities over the cell-state space.
Mutation Reconstruction Decoder: Reconstructs the observed mtDNA mutation counts ( $b_c$ ) for each cell. It models the mutation counts using a binomial distribution, where the expected variant allele frequency (VAF) is determined by the clone-specific allele frequencies ( $\phi_k$ ) and the cell's clonal probabilities ( $\pi_c$ ).
Training Objective: The model is trained by maximizing the likelihood of the observed mutation counts while minimizing the reconstruction error of the non-genetic data and the KL divergence of the latent space. A regularization term (group lasso) is employed to prevent the inference of unnecessary clones.
Key Features:
- Continuous Probabilistic Inference: Cells near clone boundaries or in shared states are assigned fractional probabilities rather than forced into a single clone.
- State-Enriched Capture: The model preferentially captures mutant-cell populations that are enriched in specific cell states.
- Heteroplasmy Separation: By modeling allele frequencies ( $\phi$ ) rather than binary mutation status, CloneVI can distinguish populations carrying the same mutation at different heteroplasmy levels.
- Background Separation: The model can decompose allele frequencies into components shared by all clones (background) and clone-specific components.

Key Results
The authors validated CloneVI using simulated data and three real-world datasets:

Simulated Data: Using a PBMC3k-based simulation, CloneVI successfully recovered ground-truth clonal structures in both single-clone and mixed-clone populations. It accurately estimated the dominant clone of each expression cluster and the degree of mixing as continuous probabilities.
MAESTER Data (Clonal Hematopoiesis, scRNA-seq): Applied to data from a patient with clonal hematopoiesis, CloneVI inferred six clones. It linked specific mtDNA variants to distinct cell states, notably identifying that Clones 3 and 5 were associated with high expression of specific T-cell receptor (TCR) variable-region genes ( $TRBV7-9$ and $TRDV2$ ). Crucially, the reconstructed VAFs derived from CloneVI's cell-state-based probabilities predicted observed VAFs in test cells better than a baseline using mean VAFs of cell types. The model preferentially captured mutant populations enriched in specific cell states, showing a strong positive correlation between cell-state enrichment and the model's ability to distinguish mutant from non-mutant cells (AUC).
CLL mtscATAC-seq Data: In chronic lymphocytic leukemia data, CloneVI separated populations carrying the same mtDNA mutation (position 5140) but at different heteroplasmy levels (Clone 1 vs. Clone 3). It also identified Clone 4, associated with a specific variant, as having increased chromatin accessibility around the PAX9 locus, a gene previously linked to CLL prognosis.
Ovarian Tumor Cohort (Multi-donor, mtscATAC-seq): In a large cohort of 108,260 cells from five donors, CloneVI inferred 17 clones. It demonstrated that clones were often confined to single donors and could separate nested clones within a donor. A differential test weighted by clonal probability ( $\pi$ -weighted) on CD8+ T cells revealed transcription factor activities (NFAT, NR4A1, EGR) consistent with T-cell exhaustion, a state in which the mutant cells concentrated. This $\pi$ -weighted approach characterized the cell state of the clone more effectively than a binary mutation-based test.

Significance and Claims
The paper claims that CloneVI provides a probabilistic framework for integrating cell states with mtDNA mutations, shifting the analytical focus from "genotype-first" to "state-first" inference. Its primary significance lies in its ability to:

Identify genetic clones that are specifically associated with cell-state diversity, rather than just describing the genetic structure itself.
Characterize the molecular properties (transcriptional or chromatin) of the cell states in which mutant cells concentrate.
Handle the continuous nature of cell states and heteroplasmy levels, which binary methods often miss.

The authors position CloneVI as complementary to genotype-first methods: while genotype-first methods describe the genetic clonal structure, CloneVI identifies which parts of that structure are associated with specific cell states. The authors note that because clonal probabilities are defined for every cell (including those without covered mutations), the method could potentially be extended to nuclear somatic mutations detected in single-cell data, though this study focuses on mtDNA. The tool is made available as open-source software to facilitate the study of genetic and non-genetic heterogeneity at single-cell resolution.

Technical Summary of CloneVI

More like this