CoPhaser: generic modeling of biological cycles in scRNA-seq with context-dependent periodic manifolds

CoPhaser is a versatile, biologically informed variational autoencoder that disentangles context-dependent periodic trajectories from other sources of cellular variability in single-cell RNA sequencing data, enabling the accurate reconstruction and analysis of diverse biological cycles such as the cell cycle, circadian rhythms, and developmental clocks across various tissues and disease states.

The Gist

Imagine you are trying to understand the rhythm of a busy city. You have thousands of people (cells) moving around, each doing different things: some are working, some are sleeping, some are running errands, and some are just sitting in a coffee shop.

Now, imagine that within this chaos, there are specific, repeating patterns. For example, everyone has a 24-hour sleep cycle (circadian rhythm), and construction workers have a 9-to-5 work cycle (cell cycle).

The problem scientists face is that these rhythms don't look the same for everyone. A construction worker in a sunny park looks different than one in a dark basement. A person sleeping in a quiet room looks different from one sleeping in a noisy train station. In biology, these "locations" are called contexts (like different cell types, diseases, or environments).

Previously, trying to map these rhythms was like trying to draw a single, perfect circle on a piece of paper that is being stretched, squashed, and twisted by different people. If you tried to force everyone into that one perfect circle, you'd get a messy, confusing drawing where the sleep cycle of a tired parent looks exactly like the work cycle of an energetic child.

Enter CoPhaser: The "Shape-Shifting Rhythm Detective"

The paper introduces a new tool called CoPhaser. Think of it as a super-smart, shape-shifting detective that can look at a chaotic crowd and say: "Okay, I see the rhythm, but I also see that the rhythm looks different depending on where you are."

Here is how it works, using simple analogies:

1. The Two-Track System (Separating the "When" from the "Who")

Imagine you are listening to a song.

• Track A (The Phase): This is the beat of the song. It tells you when you are in the cycle (e.g., "It's 3:00 AM" or "It's the S-phase of the cell cycle").
• Track B (The Context): This is the instrument playing the song. Is it a piano? A guitar? A heavy metal band? This tells you who is playing and where they are (e.g., "This is a liver cell" or "This is a cancer cell").

Old tools tried to mix these tracks together, making it hard to hear the beat clearly if the instrument was too loud. CoPhaser separates the tracks. It learns the beat (the cycle) and the instrument (the context) at the same time, ensuring that the beat stays consistent even if the instrument changes.

2. The Flexible Rubber Band

Think of a biological cycle as a rubber band.

• In a healthy, simple environment, the rubber band is a perfect circle.
• In a complex environment (like a tumor or a developing embryo), the rubber band gets stretched, squashed, or twisted.

CoPhaser doesn't force the rubber band to stay a perfect circle. Instead, it learns how the rubber band is being deformed by the environment. It says, "Ah, in this specific cancer cell, the 'sleep' part of the cycle is stretched out, but the 'wake up' part is squashed." This allows it to map the rhythm accurately, no matter how weird the environment gets.

What Did CoPhaser Discover?

The authors tested CoPhaser on four different "cities" (biological systems), and it found some amazing things:

• The Cancer City (Breast & Ovarian Tumors):
  In cancer, some cells are "sleeping" (quiescent) and some are "running wild" (proliferating). CoPhaser found that in aggressive breast cancer, the "running wild" cells are actually very active. But more importantly, it helped distinguish between genes that are always loud (bad for the cell) and genes that are only loud during the "running wild" phase. This is like realizing a siren is only loud when the fire truck is moving, not when it's parked. This helps doctors find better drug targets.

  • Spatial Discovery: In ovarian cancer, CoPhaser showed that the "running wild" cells aren't scattered randomly; they form neighborhoods. It's like a dance floor where everyone in one corner is dancing in sync, while the people in the other corner are sitting down.

• The Pediatric Leukemia City:
  In children with leukemia, the disease often comes back (relapse). CoPhaser looked at the cells before and after treatment. It found that the "bad" cells didn't necessarily start multiplying faster to come back. Instead, they changed into a "primitive" state (like a seed that goes dormant) and waited. This explains why some treatments fail: they kill the active cells but miss the dormant seeds.

• The Body Clock (Circadian Rhythm):
  Your body has a 24-hour clock. CoPhaser mapped this in the mouse aorta (a blood vessel). It found that the clock runs differently in different types of cells. The "muscle" cells in the wall have a strong, loud rhythm, while the "lining" cells have a quieter, more subtle rhythm. It's like a symphony where the drums are loud, but the flutes are whispering; CoPhaser can hear both clearly.

• The Monthly Cycle (Menstrual Cycle):
  The lining of the uterus changes every month. CoPhaser mapped this as a smooth, continuous movie rather than a series of still photos. It found that in women with endometriosis (painful tissue growth), the "movie" plays differently—certain scenes (gene expressions) are louder or happen at the wrong time.

• The Embryo Construction Site (Somite Clock):
  When an embryo grows, it builds its spine in segments, like stacking blocks. This happens in a wave. CoPhaser looked at a snapshot of an embryo and reconstructed the traveling wave of the construction crew. It also discovered that the "construction crew" (cell cycle) and the "building timer" (somite clock) are talking to each other, coordinating their steps.

The Bottom Line

CoPhaser is a universal translator for biological time.

Before, scientists had to use different tools for different cycles, and those tools often got confused when cells were in different states (like being sick or healthy). CoPhaser is a single, flexible tool that understands that rhythms change shape depending on the context.

It allows scientists to:

1. See the beat clearly, even in a noisy room.
2. Understand the environment without losing track of the time.
3. Find new patterns in diseases like cancer, revealing how cells hide, synchronize, or malfunction.

In short, CoPhaser helps us stop seeing cells as static dots and start seeing them as dynamic dancers moving to a rhythm that changes with the music of their environment.

Technical Summary

1. Problem Statement

Biological processes such as the cell cycle, circadian rhythms, the segmentation (somite) clock, and hormonal cycles are inherently periodic. In single-cell RNA sequencing (scRNA-seq) data, these processes manifest as continuous trajectories on low-dimensional manifolds. However, existing methods for inferring these cycles face significant limitations:

• Context Dependence: Biological cycles are rarely invariant; their gene expression profiles (amplitude, mean, and phase) are modulated by cellular context (e.g., cell type, disease state, metabolic conditions).
• Confounding Variability: Standard dimensionality reduction techniques (like PCA) or existing autoencoders often fail to disentangle cyclic dynamics from non-periodic cellular heterogeneity. This leads to "confounding," where context-induced shifts are misinterpreted as phase changes, or vice versa.
• Lack of Generalizability: Many current tools are specific to the cell cycle or require synchronized experimental time points (e.g., bulk time-series), making them unsuitable for unsupervised analysis of asynchronous, heterogeneous single-cell data across different biological cycles.

2. Methodology: CoPhaser

CoPhaser (Context-aware Phaser) is a biologically informed Variational Autoencoder (VAE) designed to learn context-dependent periodic manifolds. It decomposes scRNA-seq data into independent periodic and non-periodic sources of variation while preserving interpretability.

Core Architecture

The model utilizes a structured latent space with two distinct components:

1. Periodic Latent Space ($f$-space): Encodes the phase ($\phi$) of the cycle. It uses a Power Spherical distribution on the unit circle to represent the cyclic nature of the data.
2. Context Latent Space ($z$-space): Encodes non-periodic cellular variability (cell type, batch, disease state). It uses a standard multivariate Gaussian distribution.

Generative Model

Gene expression is modeled as a function of the latent phase and context. For a gene $g$ in cell $c$, the expected log-count is modeled as:
$$\log E[X_{cg}] = \mu_g + \Delta\mu_{cg}(z_c) + b_c \cdot \lambda(z_c) \cdot F_g(\phi_c)$$
Where:

• $\mu_g$: Global baseline expression.
• $\Delta\mu_{cg}(z_c)$: Context-dependent mean shift, allowing the baseline expression to vary by cell type/state.
• $b_c$: A binary indicator (from a Bernoulli classifier) distinguishing cycling vs. non-cycling (e.g., G0) cells.
• $\lambda(z_c)$: Context-dependent amplitude scaling, allowing the cycle's amplitude to vary by context.
• $F_g(\phi_c)$: A Fourier series representing the periodic trajectory of the gene as a function of phase.

Key Technical Innovations

• Mutual Information Minimization (MINE): To prevent the context encoder ($z$) from absorbing cyclic signals (and vice versa), CoPhaser employs a Mutual Information Neural Estimator (MINE). This explicitly penalizes the mutual information between the phase ($\phi$) and context ($z$) latent spaces, enforcing statistical independence.
• Dual Encoders: The model uses two parallel encoders. A "rhythmic encoder" processes a seed set of rhythmic genes to infer $\phi$, while a "context encoder" processes highly variable genes to infer $z$.
• Cycling Classifier: A relaxed Bernoulli encoder jointly trains to classify cells as cycling or non-cycling (e.g., G0), preventing non-cycling cells from distorting the phase inference.
• Flexible Initialization: The model can be trained from scratch or initialized with Fourier coefficients from a reference dataset (transfer learning), which is crucial for cycles with low-amplitude signals (e.g., circadian rhythms).

3. Key Contributions

1. Context-Dependent Periodic Manifolds: CoPhaser is the first framework to explicitly model biological cycles as manifolds that deform (shift in mean and amplitude) based on cellular context, rather than assuming a fixed trajectory for all cells.
2. Unsupervised & Generic: It requires no external time labels or prior knowledge of specific phase states (beyond a seed list of rhythmic genes). It is applicable to any periodic system (cell cycle, circadian, menstrual, segmentation).
3. Disentanglement of Proliferation and Identity: By separating phase from context, the model can distinguish between gene overexpression driven by high proliferation rates (phase-specific) versus constitutive overexpression driven by cell identity or disease state.
4. Spatial Resolution: The method extends to spatial transcriptomics, enabling the mapping of continuous cell-cycle phases onto tissue coordinates to reveal spatial synchronization.

4. Key Results

The authors validated CoPhaser across four distinct biological cycles and multiple datasets:

• Cell Cycle (General & Cancer):

  • Robustness: Accurately recovered continuous cell-cycle phases in highly heterogeneous datasets (mouse embryos, RPE1 cells) outperforming PCA, Cyclum, and DeepCycle.
  • Cancer Subtypes: In triple-negative breast cancer (TNBC), CoPhaser identified subtype-specific proliferation rates. It distinguished between genes overexpressed due to high proliferation (e.g., CDC20 in S/G2/M) versus constitutive overexpression (e.g., KRT14), offering insights into therapeutic targets.
  • Pediatric AML: Revealed that relapse is driven by a shift toward primitive, quiescent (G0) states rather than increased proliferation, identifying persistent cancer cells as the source of resistance.

• Spatial Transcriptomics (Ovarian Cancer):

  • Applied to Xenium spatial data, CoPhaser revealed that proliferative cells are not randomly distributed but form spatially synchronized patches (neighborhoods of cells in the same phase). This "proliferative architecture" was invisible to traditional single-marker assays.

• Circadian Rhythms (Mouse Aorta):

  • Successfully inferred circadian phases in single cells with low expression of clock genes.
  • Achieved lower Median Absolute Deviation (MAD) against external time compared to the state-of-the-art method Tempo, particularly in sparse cell types like fibroblasts and macrophages.
  • Identified cell-type-specific dampening of clock genes (e.g., Npas2 silence in endothelial cells).

• Menstrual Cycle (Human Endometrium):

  • Mapped continuous endometrial remodeling, identifying distinct transcriptional dynamics in endometriosis patients compared to controls (e.g., altered PAEP expression).

• Segmentation Clock (Mouse Embryo):

  • Reconstructed traveling waves of gene expression (e.g., Hes7) from snapshot data.
  • Novel Discovery: Revealed a coupling between the cell cycle and the segmentation clock, specifically showing that the G1/S transition probability is modulated by the phase of the segmentation clock (peaking when Hes7 levels rise).

5. Significance

CoPhaser represents a paradigm shift in analyzing biological rhythms in single-cell data.

• Biological Insight: It moves beyond simple phase assignment to provide a mechanistic understanding of how cellular identity modulates biological clocks.
• Clinical Relevance: In oncology, it offers a rigorous way to separate proliferation-driven effects from intrinsic tumor subtype characteristics, potentially identifying better therapeutic targets (e.g., targeting constitutive vs. phase-specific drivers).
• Methodological Advancement: By integrating topological constraints (periodicity) with flexible context modeling, it sets a new standard for unsupervised representation learning in genomics, applicable to any system where cyclic and non-cyclic processes coexist.

The code is open-source, facilitating broad adoption for comparative and integrative analyses of biological rhythms across diverse datasets.