TSvelo: Comprehensive RNA velocity by modeling cascade of gene regulation, transcription and splicing

TSvelo is a novel RNA velocity framework that utilizes interpretable neural ordinary differential equations to model the cascade of gene regulation, transcription, and splicing, thereby overcoming the limitations of existing methods to precisely capture complex gene dynamics and infer unified latent time across single cells and multi-lineage datasets.

The Gist

The Big Picture: Taking a "Time-Lapse" of Life Inside a Cell

Imagine you have a camera that can take a picture of a single cell. Usually, scientists can only take one photo at a time. This is like looking at a single frame of a movie. You see a cell, but you don't know if it's about to grow, shrink, turn into a different type of cell, or die. It's a static snapshot.

RNA Velocity is a clever trick scientists use to turn that single photo into a movie. By looking at two different "versions" of the cell's genetic instructions (mRNA)—the "raw draft" (unspliced) and the "final edited copy" (spliced)—they can guess which way the cell is moving in time.

However, existing methods for making this movie have been like trying to predict the plot of a complex drama by looking at just two characters in a noisy, crowded room. They often get confused, mix up the actors, or miss the plot twists.

TSvelo is a new, super-smart director that fixes these problems.

---

The Problem: Why Old Methods Struggle

Think of the cell's life as a busy highway.

1. The Noise: The data is messy. It's like trying to hear a whisper in a rock concert.
2. The Short Time: The process of turning "raw draft" mRNA into "final copy" happens incredibly fast. It's like a blink of an eye.
3. The Mix-up: Existing methods look at each gene (each actor) independently. They don't realize that actors talk to each other. If Actor A (a Transcription Factor) shouts at Actor B, Actor B changes their behavior. Old methods miss this conversation.
4. The Branching Roads: In complex tissues (like a developing brain), cells split into different paths (lineages). Old methods often get lost at the fork in the road, unable to tell which path a cell is taking.

The Solution: TSvelo's "Master Script"

The authors created TSvelo (Transcription-Splicing velocity). Instead of looking at one gene at a time, TSvelo looks at the entire cast of genes simultaneously.

Here is how it works, using an analogy:

1. The "Cascade" of Regulation (The Chain of Command)

Imagine a factory assembly line.

• Step 1 (Transcription): A manager (Transcription Factor) gives an order.
• Step 2 (Unspliced): The workers start building the product (raw mRNA).
• Step 3 (Splicing): The product gets polished and packaged (mature mRNA).

Old methods only looked at Step 2 and Step 3, trying to guess Step 1. TSvelo models the entire chain. It knows that the speed of the workers depends on what the manager is shouting. By including the "manager's voice" (gene regulation), it can predict the flow of traffic much more accurately.

2. The 3D Map vs. The 2D Flat Map

Imagine trying to separate a pile of mixed-up colored marbles on a flat table (2D). Red and blue marbles might look like they are mixed together.

• Old Methods: They look at the marbles on the flat table. They get confused.
• TSvelo: It lifts the marbles into a 3D space. Suddenly, the red marbles float up, and the blue marbles sink down. The "mix-up" disappears.
• The Magic: TSvelo adds a third dimension called "Transcriptional Rate" (how fast the gene is being turned on). This extra dimension helps separate cells that look identical on a flat map but are actually at different stages of their life.

3. The "Time Traveler" (Latent Time)

TSvelo doesn't just guess the time; it calculates a unified timeline for the whole cell.
Imagine a train station with many tracks (lineages). Old methods might try to time each train separately, leading to confusion when tracks merge. TSvelo acts like a central control tower. It assigns a single "pseudotime" to every cell, knowing exactly where it is on the journey, whether it's on the track to becoming a neuron or a muscle cell.

4. The "Detective" (EM Algorithm)

TSvelo uses a smart guessing game called the Expectation-Maximization (EM) algorithm.

• Guess 1: "I think this cell is at time 5."
• Check: "Does the math fit the data?"
• Adjust: "No, the data looks like time 6. Let's adjust the rules (parameters) of the factory."
• Repeat: It does this thousands of times until the "movie" perfectly matches the "photos."

What Did They Prove?

The team tested TSvelo on six different datasets, acting like different "test drives":

1. The Pancreas: Watching cells turn from ducts into insulin-producing cells. TSvelo saw the path clearly where others saw a blur.
2. The Blood: Tracking how blood stem cells become red blood cells. It correctly identified the "boss" genes (like KLF1) that tell the cells what to do.
3. The Brain: A complex map where cells split into neurons, astrocytes, and oligodendrocytes. TSvelo successfully navigated the branching roads where other methods got stuck.
4. The LARRY Dataset: A massive dataset tracking blood cell lineages. TSvelo correctly identified the different "families" of cells and their specific development paths.

The Takeaway

TSvelo is like upgrading from a blurry, black-and-white security camera to a 4K, 3D, AI-powered surveillance system for cells.

• It connects the dots between who is giving the orders (regulation) and what is being built (splicing).
• It creates a 3D map so cells don't get lost in the crowd.
• It handles complex branching paths (like a developing brain) without getting confused.

This means scientists can now predict cell fate with much higher confidence, helping us understand diseases, development, and how to potentially reprogram cells to heal the body.

Technical Summary

1. Problem Statement

Current RNA velocity methods, which infer cell fate and dynamics from single-cell RNA sequencing (scRNA-seq) data by modeling the ratio of unspliced (immature) to spliced (mature) mRNA, face several critical limitations:

• Inaccurate Phase Portraits: Existing methods often fail to capture correct gene dynamics due to data sparsity, high noise, the short time scale of splicing, and the mixing of different cell types in 2D unspliced-spliced phase portraits.
• Lack of Regulatory Integration: Most methods treat genes independently, ignoring underlying transcriptional regulatory networks (e.g., Transcription Factor (TF) interactions). While some methods incorporate regulation, they often neglect splicing signals, preventing a unified model of transcription and splicing.
• Interpretability Issues: Recent deep learning-based approaches (e.g., latent space embeddings) improve flexibility but sacrifice the interpretability of gene-level parameters (transcription, splicing, and degradation rates), which is crucial for biological insight.
• Multi-lineage Complexity: Handling large-scale datasets with complex, multi-lineage differentiation trajectories remains a significant challenge for current models.

2. Methodology: TSvelo Framework

TSvelo proposes a comprehensive mathematical framework that integrates gene regulation, transcription, and splicing into a single, highly interpretable model using Neural Ordinary Differential Equations (ODEs).

Core Mathematical Model

TSvelo models the dynamics of a gene $g$ using a system of ODEs:

1. Unspliced Dynamics: $\frac{du_g(t)}{dt} = \alpha_g(t) - \beta_g u_g(t)$
2. Spliced Dynamics: $\frac{ds_g(t)}{dt} = \beta_g u_g(t) - \gamma_g s_g(t)$

Where:

• $u_g(t)$ and $s_g(t)$ are the abundances of unspliced and spliced RNA.
• $\beta_g$ and $\gamma_g$ are gene-specific constant splicing and degradation rates.
• $\alpha_g(t)$ (Transcription Rate): Unlike previous methods that assume constant or simple time-dependent rates, TSvelo models $\alpha_g(t)$ as a function of Transcription Factor (TF) expression. It uses a linear model with a ReLU activation:
  $$\alpha_g(t) = \text{ReLU}\left(\sum_{i \in TFs(g)} w_{gi} \cdot s_i(t)\right)$$
  Here, $w_{gi}$ represents the regulatory weight of TF $i$ on target gene $g$, derived from prior knowledge (ENCODE and ChEA databases).

Key Technical Components

• Unified 3D Phase Portrait: By introducing the transcriptional rate ($\alpha$) as a latent variable, TSvelo extends the traditional 2D (unspliced-spliced) phase portrait to a 3D space (transcription-unspliced-spliced). This helps disentangle mixed cell states that overlap in 2D projections.
• Neural ODE Solver: Since analytical solutions for the coupled ODE system with time-varying $\alpha_g(t)$ are difficult, TSvelo employs a Neural ODE solver to estimate parameters and simulate trajectories.
• EM Optimization Framework: TSvelo uses an Expectation-Maximization (EM) algorithm to iteratively optimize:
  • E-step: Assigns a global latent time (pseudotime) to each cell via grid search to minimize the distance between observed data and model predictions.
  • M-step: Updates the ODE parameters (transcription weights, splicing/degradation rates) using gradient descent (Adam optimizer) while respecting the sparsity constraints imposed by the TF-target prior.
• Multi-lineage Handling: The method detects initial cell states based on unspliced-spliced delays, segments lineages using PAGA, and models each lineage independently before merging results.

3. Key Contributions

1. Unified Cascade Modeling: TSvelo is the first method to jointly model the full cascade of gene regulation $\to$ transcription $\to$ splicing within a single interpretable ODE framework.
2. High Interpretability: Unlike "black-box" deep learning velocity methods, TSvelo provides explicit, biologically meaningful parameters for transcription rates, splicing rates, and degradation rates, as well as TF-target regulatory weights.
3. 3D Dynamics Representation: The introduction of the transcription rate as a third dimension significantly improves the separation of cell states and the accuracy of phase portrait fitting.
4. Robustness in Complex Scenarios: The method demonstrates superior performance in handling noisy data, complex gene dynamics (e.g., non-monotonic patterns), and multi-lineage differentiation tasks.

4. Results

TSvelo was evaluated on six scRNA-seq datasets, including two multi-lineage datasets, and compared against state-of-the-art methods (scVelo, Dynamo, UniTVelo, CellDancer, Multivelo, etc.).

• Pancreas Dataset (Differentiation):
  • TSvelo achieved higher kNN classification accuracy for cell-state separation in the 3D phase portrait compared to 2D portraits.
  • It successfully modeled complex gene dynamics (e.g., MAML3, ANXA4) where baseline methods failed due to overlapping cell clusters in 2D space.
• Gastrulation Erythroid Dataset:
  • TSvelo showed the highest velocity consistency, in-cluster coherence, and cross-boundary correctness.
  • It correctly captured counter-clockwise dynamics in genes like HSP90AB1 and identified key regulatory relationships (e.g., KLF1 regulating HBA-X, ALAS2, and GYPA), revealing time-delay patterns consistent with biological knowledge.
• Mouse Brain (Multi-omics/ATAC+RNA):
  • Compared to Multivelo (which uses ATAC-seq data), TSvelo (using only scRNA-seq) produced more consistent velocity streams for Radial Glia to Intermediate Progenitor differentiation.
  • It correctly modeled non-monotonic expression patterns (e.g., BASP1) that Multivelo failed to capture due to noise in ATAC data.
• Multi-lineage Datasets (Dentate Gyrus & LARRY):
  • TSvelo successfully segmented and modeled three distinct lineages in the dentate gyrus dataset, whereas scVelo struggled and CellDancer failed to capture specific immature granule cell trajectories.
  • On the LARRY dataset (hematopoietic differentiation), it accurately traced lineage-specific gene expression patterns (e.g., PYGL in neutrophils vs. MS4A3 in progenitors).

5. Significance

TSvelo represents a significant advancement in single-cell trajectory inference by bridging the gap between regulatory network modeling and RNA kinetics.

• Biological Insight: By explicitly modeling TF-driven transcription, it allows researchers to infer not just where cells are going, but why (via regulatory drivers) and how (via kinetic rates).
• Robustness: Its ability to handle noise and complex multi-lineage structures makes it suitable for large-scale, real-world biological datasets where previous methods often fail.
• Future Directions: The framework sets a foundation for incorporating context-specific chromatin accessibility data and modeling variable splicing/degradation rates, further enhancing the precision of cell fate prediction.

In summary, TSvelo offers a more reliable, interpretable, and comprehensive approach to RNA velocity, enabling deeper understanding of gene regulatory mechanisms driving cell differentiation.