Decoupling Lineage and Intrinsic Information in Single-Cell Lineage Tracing Data with Deep Disentangled Representation Learning

This paper introduces DeepTracing, a deep generative framework that effectively disentangles intrinsic transcriptional states from lineage-driven effects in single-cell lineage tracing data, outperforming existing methods in clustering, phylogenetic recovery, and trajectory reconstruction across diverse biological contexts.

The Gist

Imagine you are trying to understand the life story of a family. You have two types of information about them:

1. Their DNA (Lineage): Who their parents are, who their siblings are, and how they are related.
2. Their Personality (State): What they do for a living, their hobbies, and their current mood.

In the world of biology, scientists use a technology called single-cell RNA sequencing to look at individual cells. They want to know two things:

• The Lineage: How did this cell get here? (Is it a child of a specific parent cell?)
• The State: What is this cell doing right now? (Is it a skin cell, a liver cell, or a cancer cell?)

The Problem:
Until now, looking at these cells was like trying to read a family tree and a resume at the same time, but the ink was all mixed up. Existing computer programs could see that cells were related, but they got confused about what the cells were doing. Or, they could see what the cells were doing, but they lost the family tree. It was a "choose one, lose the other" situation.

The Solution: DeepTracing
The authors of this paper created a new AI tool called DeepTracing. Think of it as a super-smart librarian who can take a messy pile of mixed-up books and perfectly sort them into two separate, neat stacks: one stack for "Family History" and one for "Current Jobs."

Here is how it works, using simple analogies:

1. The "Two-Drawer" Filing System

Imagine the AI has a special filing cabinet with two distinct drawers:

• Drawer A (The Lineage Drawer): This drawer only cares about who is related to whom. It uses a special map (called a Gaussian Process) that acts like a GPS for family trees. If two cells are close relatives, they get filed next to each other here, regardless of what they are doing.
• Drawer B (The Intrinsic Drawer): This drawer only cares about what the cell is doing. It ignores the family tree. If two cells are doing the same job (like both being liver cells), they get filed together here, even if they come from totally different families.

2. The "Noise-Canceling" Headphones

Usually, when you try to separate these two things, the "family noise" leaks into the "job noise."
DeepTracing uses a special mathematical trick (called Total Correlation regularization) that acts like noise-canceling headphones. It actively silences the "family" signal when looking at the "job" drawer, and vice versa. This ensures the two stories stay completely separate and clear.

3. The "Three-View" Camera

Once the AI sorts the data, it gives scientists three different ways to look at the cells:

• The Family View: Shows the pure family tree. Great for seeing how a tumor spread from the lung to the liver.
• The Job View: Shows the pure cell types. Great for seeing how a brain cell matures from a baby cell to an adult cell.
• The Combined View: A perfect blend of both. This is the "super-view" that shows the whole picture without the confusion.

Why This Matters (Real-World Examples)

Example 1: The Cancer Detective
In a study of lung cancer, scientists wanted to know how cancer spread (metastasized) to other organs.

• Old Way: The computer saw the cancer cells and the healthy cells mixed together, making it hard to tell which cancer came from where.
• DeepTracing: It successfully separated the "family tree" of the cancer. It could say, "Ah, these cancer cells in the kidney came from this specific group in the liver, while those came from that group." It even found the specific "genetic clues" (genes) that told the cancer how to travel.

Example 2: The Time Traveler
In a study of developing mouse brains, scientists looked at cells from different days of pregnancy.

• Old Way: The computer got confused by the dates. It thought cells from Day 11 were totally different from Day 15, even if they were the same type of cell.
• DeepTracing: It realized, "Wait, the date is just a label. The cell type is what matters." It stripped away the "time" confusion and showed a smooth, continuous movie of how a brain cell grows up, regardless of when it was measured.

The Bottom Line

DeepTracing is a powerful new tool that stops scientists from having to choose between seeing a cell's family history or its current job. It separates the two, giving us a crystal-clear view of how life develops, how diseases like cancer evolve, and how cells change over time. It's like finally getting a high-definition, 3D map of the biological world instead of a blurry, 2D sketch.

Technical Summary

1. Problem Statement

Single-cell RNA sequencing (scRNA-seq) combined with lineage tracing technologies generates multimodal data containing both gene expression profiles and ancestral lineage information (e.g., CRISPR barcodes). However, existing computational methods face a fundamental challenge: the entanglement of intrinsic cellular states and lineage-driven effects.

• The Trade-off: Current visualization and analysis tools (e.g., UMAP, t-SNE, standard VAEs) typically prioritize transcriptional continuity, which obscures the underlying lineage tree structure. Conversely, methods focusing on lineage fidelity often disrupt the smooth gradients of cellular differentiation states.
• Biological Complexity: Cells sharing the same lineage may cluster together due to common ancestry rather than cell-autonomous programs. Conversely, cells with the same lineage may diverge due to intrinsic transcriptional programs.
• Limitations of Existing Tools: Methods like PORCELAN struggle with scalability on large datasets and often fail to preserve continuous expression-defined trajectories while recovering lineage structures. There is a lack of frameworks that can explicitly decouple these two sources of variation to produce interpretable, independent latent representations.

2. Methodology: DeepTracing

The authors propose DeepTracing, a deep generative framework based on a Disentangled Variational Auto-Encoder (VAE) that integrates Gaussian Processes (GPs) and Total Correlation (TC) regularization.

Core Architecture

The model accepts two inputs: the gene expression matrix ($Y$) and lineage information ($X$, e.g., lineage tree or barcodes). It maps these to a structured latent space $Z$ decomposed into two statistically independent components:

1. Intrinsic Embedding ($Z_{P+1:L}$): Captures cell-specific local variation (cell type, state, identity).
   • Prior: Standard Gaussian distribution $\mathcal{N}(0, I)$.
2. Lineage Embedding ($Z_{1:P}$): Captures global constraints imposed by ancestral relationships and developmental history.
   • Prior: Gaussian Process (GP) prior, where the kernel matrix is derived from the pairwise lineage distance between cells (either tree path distance or Hamming distance of barcodes).

Key Technical Innovations

• Layered Latent Space: The latent space is explicitly factorized. The lineage dimensions are constrained by a GP prior to enforce smoothness and adherence to the lineage tree structure, while intrinsic dimensions remain unstructured to capture heterogeneity.
• Total Correlation (TC) Regularization: To ensure the two latent factors are truly independent, the model minimizes the Total Correlation (mutual information) between the intrinsic and lineage embeddings. This is enforced via a penalty term in the loss function, preventing the model from "cheating" by encoding lineage information into the state embedding or vice versa.
• Scalability via Sparse GP: Standard GP inference has $O(N^3)$ complexity, which is prohibitive for large single-cell datasets. DeepTracing employs amortized variational inference combined with sparse GP regression using inducing points. This reduces complexity to $O(bm^2 + m^3)$ (where $b$ is batch size and $m \ll N$ is the number of inducing points), enabling application to datasets with tens of thousands of cells.
• Loss Function: The model optimizes an augmented Evidence Lower Bound (ELBO):
  $$\text{ELBO} = \text{Reconstruction Loss} - \beta_1 \text{KL}(\text{Lineage Prior}) - \beta_2 \text{KL}(\text{Intrinsic Prior}) - \lambda \text{TC Penalty}$$

Output Representations

The model generates three distinct embeddings for downstream analysis:

1. Intrinsic Embedding: Purely reflects cell identity and state, independent of lineage.
2. Lineage Embedding: Preserves the hierarchical tree structure and ancestral relationships.
3. Unified Embedding: A concatenation of both, providing a comprehensive view for tasks like clustering and trajectory inference.

3. Key Results

The authors validated DeepTracing on synthetic data (TedSim) and three real-world biological datasets.

A. Synthetic Data (TedSim)

• Performance: DeepTracing outperformed scVI and original expression data in clustering tasks.
• Metrics: It achieved the highest Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) for both cell-state clustering and lineage subtree recovery.
• Robustness: The model remained stable across varying probabilities of asymmetric cell division ($p_a$), demonstrating robustness to increased data complexity.
• Comparison: Unlike PORCELAN, which failed to produce valid results on the ~4,000 cell simulated dataset due to computational constraints, DeepTracing processed the data efficiently.

B. Mouse Lung Tumor Lineage Tracing (3515_Lkb1_T1)

• Metastasis Separation: In the original expression space, primary and metastatic tumor cells were intermixed. DeepTracing's Unified Embedding cleanly separated primary tumors from metastases (liver, kidney, lymph node).
• Trajectory Recovery: The model successfully reconstructed known metastatic routes, including:
  • Early lymph-node divergence.
  • Liver-to-kidney cross-seeding events (where Kidney metastases K2/K3 shared a common ancestor with Liver L1, while K1 showed mixed ancestry).
• Gene Discovery: Using Canonical Correlation Analysis (CCA) on the unified embedding, the model identified metastasis-associated genes, such as the downregulation of Sftpc (AT2 marker) and upregulation of Clu in metastases.

C. Large-Scale Tumor Dataset (3724_NT_T1)

• Scalability: Successfully processed a dataset with >20,000 cells, a scale where methods like PORCELAN are computationally infeasible.
• Heterogeneity Resolution:
  • The Intrinsic Embedding grouped mesenchymal cells from diverse sites (primary vs. metastatic) into shared clusters based on molecular phenotype.
  • The Unified Embedding organized cells into subtrees that correlated strongly with tumor types, revealing polyclonal seeding and cross-seeding events.

D. Mouse Ventral Midbrain Development (Time-Series)

• Temporal Decoupling: Applied to embryonic midbrain data (E11.5 and E15.5).
• Result: The Intrinsic Embedding merged cells from both time points into a continuous differentiation trajectory (radial glia $\to$ neuroblasts $\to$ neurons), effectively removing the "batch effect" of sampling time.
• Time Embedding: A dedicated latent layer successfully isolated the temporal information, organizing cells strictly by their developmental stage (E11.5 vs. E15.5), proving the model can disentangle time from intrinsic state.

4. Significance and Contributions

• Solving the Lineage-State Trade-off: DeepTracing is the first framework to simultaneously maintain the continuity of expression-defined states and the fidelity of lineage trees, resolving a long-standing limitation in single-cell analysis.
• Interpretability: By explicitly disentangling intrinsic and lineage factors, the model provides biologically interpretable representations. Researchers can now analyze cell state evolution without lineage noise, or lineage history without state bias.
• Scalability: The integration of sparse Gaussian processes and amortized inference makes the method applicable to large-scale, high-dimensional multimodal datasets (20k+ cells), which is critical for modern tumor evolution studies.
• Biological Insight: The model has demonstrated utility in uncovering complex evolutionary patterns, such as cross-seeding in metastasis and continuous developmental trajectories obscured by temporal sampling.
• Generalizability: The framework is adaptable to various lineage tracing modalities (tree-based or barcode-based) and biological contexts (development, cancer, tissue regeneration).

In conclusion, DeepTracing establishes a new standard for analyzing multimodal single-cell data, offering a scalable, interpretable, and mathematically rigorous approach to decoupling the complex interplay between a cell's history and its current state.