Experimental Time Points Guided Transcriptomic Velocity Inference

CellDyc is a semi-supervised learning framework that integrates experimental time points with transcriptomic velocity inference to accurately reconstruct cellular trajectories and reveal intrinsic temporal dynamics, outperforming existing methods across diverse biological contexts.

The Gist

Imagine you are trying to understand a complex story, like a movie, but you only have a few scattered snapshots taken at random moments. You have a photo of a baby, a photo of a teenager, and a photo of an adult. You know the order they were taken (Time 1, Time 2, Time 3), but you don't know exactly how the baby turned into the teenager, or what specific steps happened in between.

This is the challenge scientists face with single-cell RNA sequencing. They take "snapshots" of thousands of individual cells at different times to see how they change (like a cell turning from a stem cell into a muscle cell). But because the process destroys the cell to take the picture, they can't watch the cell grow in real-time. They only have the still photos.

Existing tools try to guess the movie by either:

1. Ignoring the timestamps: Trying to guess the story just by looking at how similar the photos look (which often gets the plot twisted).
2. Only looking at the gaps: Assuming the change only happens between the snapshots, missing all the tiny, fast movements that happen during the time between photos.

Enter CellDyc, a new tool created by researchers at Nanjing University. Think of CellDyc as a super-smart AI director that can watch your scattered photos and reconstruct the entire, smooth movie, filling in every missing frame with high precision.

Here is how CellDyc works, using some everyday analogies:

1. The "Gene Clock" (The Internal Watch)

Imagine you have a pile of photos from a race, but the clocks on the photos are broken or missing. However, you know that the runners' muscles get more tired and their sweat increases as the race goes on.

• Old methods try to guess the time just by looking at the runners' positions.
• CellDyc builds an internal "Gene Clock." It looks at the specific genes inside the cells (the "sweat" and "muscle fatigue") and realizes, "Ah, this gene is turning on, which means we are at minute 15 of the race." It creates a perfect, continuous timeline based on the biology itself, not just the experimental schedule.

2. The "Local Neighborhood" (The Crowd Surfing)

Imagine you are at a crowded concert. You want to know which way the crowd is moving.

• Old methods might look at the crowd at 5:00 PM and then at 6:00 PM and draw a giant arrow from one group to the next. This misses the fact that people were dancing, jumping, and moving in circles during that hour.
• CellDyc looks at your immediate neighbors. It asks, "Who is standing next to you? Are they slightly older or younger in the process?" By comparing a cell to its closest neighbors and checking their "time stamps," CellDyc can figure out the exact direction the cell is moving right now. It's like watching the crowd flow in real-time rather than just comparing two static photos.

3. Cleaning Up the Mess (The Noise Filter)

Sometimes, the data is messy. Maybe the "time stamps" on the photos are blurry or slightly wrong (like a clock that runs fast or slow).

• CellDyc is like a noise-canceling headphone for biology. Even if the experimental time labels are a bit fuzzy, CellDyc uses the "Gene Clock" to correct the errors. It realizes, "The clock says it's 2:00, but the biology says it's actually 2:15," and it adjusts the movie accordingly.

What Did They Discover?

Using this new "movie director," the scientists found some cool things that other tools missed:

• The Delayed Commute (Glioblastoma): In brain cancer, they found that the immune system's "police cars" (monocytes) were trying to get to the scene to fight the cancer, but the cancer's "traffic jam" (immunosuppressive environment) was making them drive much slower than normal. CellDyc could measure exactly how much slower.
• The Reversing Train (Blood Cells): In developing blood cells, they found a section of the track where the cells seemed to be moving backward or taking a detour before moving forward again. Other tools thought the train was just stuck; CellDyc saw the complex maneuver.

Why Does This Matter?

Think of CellDyc as a universal translator for time. It can take messy, incomplete data and turn it into a clear, high-definition movie of how life works at the cellular level.

• For Doctors: It helps them understand how diseases progress and how fast cells are changing, which could lead to better treatments.
• For Scientists: It works with almost any existing tool, meaning they don't have to throw away their old software; they just plug CellDyc in to get better results.

In short, CellDyc takes the "stuttering" snapshots of cellular life and turns them into a smooth, continuous film, revealing the hidden secrets of how our bodies grow, heal, and sometimes get sick.

Technical Summary

1. Problem Statement

Time-series single-cell RNA sequencing (scRNA-seq) allows for the longitudinal tracking of biological processes, but the destructive nature of the technology results in discrete, static snapshots rather than continuous data. Reconstructing continuous cellular dynamics from these snapshots is a fundamental challenge. Existing methods suffer from specific limitations:

• Pseudotime and RNA Velocity methods (e.g., scVelo, cellDancer): These typically ignore explicit experimental time points, relying instead on transcriptomic similarity or splicing kinetics. While they offer high resolution, they are sensitive to noise, parameter selection, and often fail to capture correct global directionality without prior assumptions.
• Optimal Transport (OT) methods (e.g., Waddington-OT, Moscot): These leverage experimental time points to ensure global directionality but treat time points as static probability distributions. Consequently, they fail to reconstruct high-resolution, instantaneous cellular dynamics within time points, often conflating long-term trends with short-term fluctuations.

The core problem is the lack of a framework that can simultaneously leverage explicit experimental time labels for global directionality while resolving fine-grained, instantaneous transcriptomic velocities within those time points.

2. Methodology: CellDyc

The authors propose CellDyc, a semi-supervised deep learning framework designed to infer transcriptomic velocity and an intrinsic "gene-embedded time."

Core Architecture

CellDyc utilizes a semi-supervised learning paradigm that integrates explicit temporal information with local gene expression similarity. The architecture consists of:

1. Linear Embedding Layer: Projects input gene expression vectors into a low-dimensional latent space. The first dimension is explicitly designated as the Gene Clock ($\hat{c}$), representing the gene-embedded time.
2. Shared Backbone: A fully connected network processes the latent representation.
3. Dual-Head Decoders:
   • Non-steady-state Probability Head: Predicts the probability of a gene being in a dynamic (non-steady) regime.
   • Expression Trend Head: Predicts the directionality of expression changes (up or down).

Training Strategy (Semi-Supervised)

The model is trained using two coupled loss functions:

1. Temporal Pairwise Ranking Loss ($\mathcal{L}_{time}$):
   • Instead of regressing directly on discrete time points, the model learns pairwise rank relationships between cells from different experimental time points.
   • It enforces directional consistency between the predicted time differences and the ground-truth experimental time differences, ensuring the model learns a continuous timeline that respects the experimental order.
2. Trend Alignment Loss ($\mathcal{L}_{trend}$):
   • This loss supervises the prediction of instantaneous trends.
   • It utilizes a neighborhood-based temporal smoothing approach. For a given cell, the model computes a "local directionality" by weighting the transcriptomic differences between the cell and its neighbors based on their relative temporal orientation (e.g., neighbor is in the future vs. past).
   • This allows the model to learn instantaneous dynamics even when explicit time labels are coarse or noisy.

Transcriptomic Velocity Prediction

The final transcriptomic velocity vector ($\frac{d\mathbf{x}}{d\hat{c}}$) is derived by combining:

• Direction: The predicted trend vector (sign of change).
• Magnitude: Calculated as the ratio of local gene expression variance to local gene-embedded time variance within a cell's neighborhood. This scales the velocity to reflect the actual rate of change relative to the inferred biological time.

Gene Clock Module

A unique feature is the Gene Clock, a linear combination of genes learned to maximize the preservation of temporal rank constraints. It serves as an intrinsic coordinate system, recovering a "real-world" biological time that is embedded within the transcriptome, independent of noisy experimental labels.

3. Key Contributions

• Novel Framework: Introduction of CellDyc, the first method to seamlessly integrate experimental time supervision with local transcriptomic similarity to infer instantaneous velocities.
• Gene-Embedded Time: The concept and implementation of an intrinsic temporal coordinate system derived from gene expression, which can be recovered even from noisy or inferred time labels (e.g., pseudotime).
• Robustness to Sampling Density: Demonstrated ability to maintain high accuracy even with sparse sampling (e.g., 3 time points), whereas OT-based methods degrade significantly under such conditions.
• Seamless Integration: The output (velocity vectors) is compatible with downstream tools like CellRank, PAGA, and dynamo, allowing for immediate application in fate mapping and trajectory analysis.

4. Results and Validation

The authors validated CellDyc across simulated and real-world datasets:

• Simulated Data (scDesign3):
  • CellDyc achieved a Pearson correlation of 0.98 with ground-truth time and accurately recovered velocity vectors.
  • It outperformed OT-based methods (Waddington-OT, Moscot) in recovering trajectory direction, especially at lower sampling densities (3 time points).
• C. elegans Embryogenesis:
  • Using a dataset with known lineage trees, CellDyc recovered high-resolution embryo time (Pearson $r=0.72$) from only 3 sparse experimental time points.
  • It accurately recapitulated ground-truth velocities, whereas OT methods showed minor local inaccuracies.
• Zebrafish Embryogenesis:
  • CellDyc successfully reconstructed developmental trajectories from 12 time points and demonstrated robustness when trained on subsets (6 or 3 points).
  • It outperformed CytoTRACE and other velocity methods in discriminating late developmental stages.
• Glioblastoma Monocyte Differentiation (Zman-seq):
  • Applied to noisy, asynchronous in vivo timestamping data. CellDyc successfully delineated two distinct differentiation trajectories (pro-inflammatory vs. immunosuppressive).
  • Key Biological Insight: Quantitatively demonstrated that the immunosuppressive microenvironment delays monocyte differentiation compared to the control, a finding obscured by raw experimental time labels.
• Mouse Gastrulation (Erythroid Maturation):
  • Resolved temporal heterogeneity, revealing that self-renewal capacity declines progressively and that different lineages have distinct rates of state change.
  • Identified a reversed trajectory segment in Blood Progenitor 2 cells, which other methods missed.
• Cell Reprogramming:
  • In a mouse fibroblast reprogramming dataset, CellDyc correctly identified initial and terminal states and computed accurate fate probabilities using CellRank.
  • In contrast, scVelo and OT methods failed to identify correct initial/terminal states automatically due to incorrect velocity directionality.
• Generalization to Inferred Time:
  • CellDyc was tested using pseudotime and CytoTRACE time as weak supervision. It successfully "rectified" the distortions in these inferred labels, recovering a gene-embedded time that correlated better with experimental ground truth than the input labels themselves.

5. Significance

• Bridging the Gap: CellDyc resolves the trade-off between the global reliability of OT methods and the local resolution of RNA velocity methods.
• Biological Discovery: By providing a rigorous, data-driven temporal reference, CellDyc enables the detection of subtle biological phenomena, such as microenvironment-induced delays in differentiation and lineage-specific temporal heterogeneity, which were previously inaccessible.
• Methodological Flexibility: The framework's ability to work with experimental time, inferred pseudotime, or noisy timestamps makes it a versatile tool for diverse experimental designs, including those lacking precise temporal annotations.
• Foundation for Future Work: The authors suggest that the "transcriptome-centric" assumption of CellDyc can be extended to multi-omics data (e.g., CITE-seq, spatial transcriptomics) to infer multi-modal velocities and integrate spatial cues.

In summary, CellDyc represents a significant advancement in computational biology, offering a robust, semi-supervised solution for deciphering the continuous dynamics of cellular processes from discrete time-series snapshots.