Patches: A Representation Learning framework for Decoding Shared and Condition-Specific Transcriptional Programs in Wound Healing

Patches is a novel representation learning framework that utilizes conditional subspace learning to effectively disentangle shared and condition-specific transcriptional programs in single-cell RNA sequencing data, enabling robust integration and biological insights in complex experimental designs such as wound healing studies involving aging and drug treatments.

The Gist

Imagine you are trying to understand a massive, chaotic orchestra playing a complex symphony. This orchestra represents the cells in your body, and the music they play is their genetic activity (gene expression).

Usually, when scientists study this orchestra, they try to figure out the song by listening to the whole group at once. But here's the problem: the orchestra is playing different songs at the same time depending on who is playing (young vs. old musicians), what instrument they are using (different cell types), and whether they are under the influence of a conductor's special baton (drug treatment).

Existing tools are like bad sound engineers. They either try to mix everything into one muddy track (losing the specific details) or they try to isolate every single instrument so strictly that they lose the harmony of the whole song. They struggle when data is missing or when the "musicians" don't match up perfectly across different experiments.

Enter Patches.

What is Patches?

Think of Patches as a super-smart, magical audio mixer designed specifically for this biological orchestra. Its job is to take the messy, overlapping sound of the cells and separate it into two distinct tracks:

1. The "Universal Rhythm" (Shared Programs): This is the part of the music that everyone plays, no matter their age or the drug they took. It's the fundamental beat of "being a skin cell" or "healing a wound."
2. The "Solo Improvisations" (Condition-Specific Programs): This is the unique flair added by specific conditions. It's the "jazz solo" a cell plays because it's old, or the "rock riff" it plays because it's reacting to a new medicine.

How Does It Work? (The Creative Analogy)

Imagine you are a detective trying to solve a mystery in a crowded room where everyone is wearing a mask.

• The Masks: These are the experimental conditions (Age, Drug, Time).
• The Person Behind the Mask: This is the cell's true identity.

Old methods tried to guess the person by looking at the mask, but they got confused when the masks were different. Patches uses a special technique called "Conditional Subspace Learning."

Think of Patches as having two pairs of glasses:

• Glasses A (The Shared Lens): These glasses blur out the masks. They let you see the common features of the people in the room (e.g., "They are all fibroblasts"). This helps you see the universal rules of healing that apply to everyone.
• Glasses B (The Specific Lens): These glasses focus only on the masks. They let you see exactly how the "Old" mask changes a person's posture compared to the "Young" mask, or how the "Drug" mask changes their mood.

Patches forces these two views to stay separate. It uses a "game" (an adversarial classifier) where one part of the AI tries to guess the mask from the shared view, and the other part tries to prevent it from guessing. This ensures the "Universal Rhythm" stays pure and isn't contaminated by the "Solo Improvisations."

Why is this a Big Deal? (The Real-World Application)

The researchers tested Patches on skin wound healing in mice, looking at two big variables: Aging and Drug Treatment.

1. The Aging Experiment:
Imagine trying to heal a cut. A young mouse heals fast and cleanly. An old mouse heals slowly and often gets scarred.

• Without Patches: Scientists might just see "Old cells are different."
• With Patches: They could see exactly what changed. They found that while the basic "healing rhythm" was the same, the "old" cells had a specific "solo" where they messed up the construction of the skin's scaffolding (the extracellular matrix). They also discovered that a specific gene, Apoe, which controls fat and inflammation, was acting strangely in old cells. This gives doctors a specific target to fix: "Don't just treat the wound; fix the fat metabolism in these old cells."

2. The Drug Experiment:
They tested a drug called Verteporfin.

• Without Patches: It's hard to tell if the drug is helping the "healing rhythm" or just changing the "solo."
• With Patches: They could see that the drug successfully tweaked the specific "improvisation" to speed up healing without breaking the universal rules of the cell.

The "Magic" Features

• Filling in the Blanks: Sometimes, in real life, you can't get data for every combination (e.g., you have "Old + Drug A" and "Young + Drug B," but no "Old + Drug B"). Patches is like a creative writer who can imagine the missing story. It can predict what "Old + Drug B" would look like by combining the "Old" solo with the "Drug B" solo it learned from other groups.
• The "Explainable" Decoder: Most AI models are "black boxes"—they give an answer but don't say why. Patches has a special "linear decoder" (like a clear instruction manual) that tells you exactly which genes are responsible for the changes. It doesn't just say "The song changed"; it says, "The song changed because the violin section (Gene X) started playing louder."

The Bottom Line

Patches is a new tool that helps scientists listen to the complex music of life without getting confused. It separates the universal rules of biology from the specific changes caused by age, disease, or drugs.

By doing this, it helps us understand why wounds heal poorly in the elderly and how to design better drugs to fix it. It turns a chaotic, noisy recording of life into a clear, organized sheet music that doctors and researchers can actually read and use to save lives.

Technical Summary

1. Problem Statement

Single-cell RNA sequencing (scRNA-seq) has revolutionized the study of cell states and transitions in complex biological processes like wound healing. However, existing computational methods face significant challenges in multi-condition experimental designs (e.g., varying age, drug treatment, and timepoints):

• Disentanglement Failure: Most methods struggle to simultaneously separate shared transcriptional programs (common across all conditions) from condition-specific variations (unique to age, treatment, or time).
• Data Limitations: Experimental designs often suffer from missing data, unmatched cell populations across conditions, or destructive sampling (where the same cell cannot be observed at multiple timepoints).
• Interpretability vs. Expressivity: Generative models (like VAEs) offer high expressivity for capturing non-linear biological dynamics but often lack interpretability. Conversely, linear models are interpretable but may fail to capture complex non-linear relationships.
• Bias in Integration: Current integration tools often prioritize shared signals (blurring condition-specific differences) or condition-specific signals (failing to align cells across conditions), making it difficult to study how specific factors (like aging) alter universal repair mechanisms.

2. Methodology: The Patches Framework

Patches is a deep representation learning framework based on a Variational Autoencoder (VAE) architecture, extended with concepts from multi-modal learning and information theory.

Core Architecture

• Hierarchical Latent Space: Patches decomposes a cell's latent identity ($\rho_n$) into two distinct components:
  1. Shared Latent Variable ($z_n$): Captures condition-independent features (e.g., intrinsic cell type identity) that remain invariant across experimental attributes.
  2. Condition-Specific Latent Variables ($w_n$): Encodes group and attribute-specific information (e.g., age, treatment, time). $w_n$ is constructed by concatenating vectors corresponding to individual condition groups ($y_n$).
• Generative Model:
  • The model assumes gene expression counts follow a Zero-Inflated Negative Binomial (ZINB) distribution.
  • The parameters of this distribution are conditioned on the concatenated latent vector $[z_n, w_n]$.
• Disentanglement via Adversarial Learning:
  • To ensure $z_n$ does not leak condition-specific information, Patches employs adversarial classifiers.
  • These classifiers attempt to predict the experimental conditions (e.g., "old" vs. "young") solely from the shared latent space $z_n$.
  • The model is trained to minimize the reconstruction loss while maximizing the classifier's error (minimizing mutual information), effectively forcing $z_n$ to be "condition-agnostic."
• Interpretability Module:
  • Patches offers an optional linear decoder layer. Instead of a non-linear neural network, it uses a single linear layer to map latent representations back to gene expression.
  • This allows the extraction of interpretable coefficients that directly quantify how specific genes contribute to shared vs. condition-specific states.

Cross-Condition Transfer

Patches enables the generation of transcriptomic data for unseen condition combinations. By encoding a cell into the shared space $z_n$ and sampling a new condition-specific vector $w_n$ based on a target label, the model can predict how a cell's expression profile would change under a different condition (e.g., predicting an aged cell's response to a drug).

3. Key Contributions

1. Novel Disentanglement Strategy: Unlike prior methods (e.g., scVI, scANVI) that focus on either shared or supervised condition-specific learning, Patches jointly learns both, explicitly enforcing separation via information-theoretic constraints.
2. Handling Unmatched Data: The framework is robust to experimental designs where cells are not observed across all condition combinations (common in destructive time-series experiments).
3. Dual-Mode Operation: It provides a trade-off between generative power (non-linear decoder for imputation/transfer) and biological interpretability (linear decoder for hypothesis generation).
4. Scalable Multi-Attribute Encoding: It utilizes a factorized encoding scheme for condition groups (e.g., Age + Time + Treatment) rather than a massive one-hot vector of all combinations, preventing the "curse of dimensionality" in complex designs.

4. Results

The authors validated Patches using synthetic data and two real-world skin wound healing datasets (Vu et al. and Mascharak et al.).

Synthetic Benchmarks

• Attribute Separation: On synthetic data, Patches successfully separated shared and condition-specific signals, whereas scVI failed to model condition-specificity and scANVI failed to capture shared structure.
• Swiss Roll Dataset: Patches demonstrated the ability to "unroll" non-linear structures (Swiss roll) while maintaining clear separation between binary conditions, validating its non-linear expressivity and disentanglement capabilities.

Real-World Applications

• Aging and Wound Healing (Vu Dataset):
  • Applied to 26,966 cells from young and aged mice.
  • Findings: Identified subtle, condition-specific changes in fibroblasts and keratinocytes associated with aging.
  • Key Genes: Discovered age-specific dysregulation in Apoe (lipid metabolism/inflammation), Scube3, and Cilp (ECM organization), as well as impaired stem cell maintenance markers (Krt15, Lef1) in aged keratinocytes.
• Drug Treatment (Mascharak Dataset):
  • Analyzed the effect of verteporfin treatment on wound healing in mice (2,284 cells).
  • Findings: Successfully captured transcriptional shifts associated with drug treatment despite low sample sizes and imbalanced cell counts.
  • Transfer Accuracy: Patches achieved high accuracy in "transferring" gene expression profiles from one timepoint to another (e.g., predicting Day 14 from Day 7), outperforming scANVI and matching "Oracle" versions of scVI and scGen (which had access to ground-truth shift information).

Quantitative Performance

• Separability: Patches achieved superior separation of condition-specific clusters (measured by KNN accuracy, cASW, and ARI) compared to scVI, scANVI, MultiGroupVI, and scINSIGHT.
• Reconstruction: It maintained competitive reconstruction accuracy (RMSE, Pearson correlation) with scVI, proving that adding disentanglement constraints does not sacrifice generative fidelity.

5. Significance and Impact

• Mechanistic Insight: Patches moves beyond simple clustering to reveal how specific factors (like aging) modulate universal biological programs. It distinguishes between genes that are always active (shared) and those that are only active under specific stressors (condition-specific).
• Therapeutic Discovery: By identifying specific gene programs altered in aging or drug responses (e.g., ECM remodeling in aged fibroblasts), Patches highlights potential biomarkers and therapeutic targets for improving wound healing in elderly populations.
• Experimental Design: The framework is particularly valuable for studies with missing data or destructive sampling, allowing researchers to infer counterfactuals (e.g., "What would this young cell look like if it were old?") without needing matched samples.
• Future Directions: The authors suggest extending Patches to spatial transcriptomics and cross-species comparisons (mouse to human) to bridge the gap between model organisms and clinical applications.

In summary, Patches provides a robust, interpretable, and flexible framework for decoding the complex interplay between universal cellular mechanisms and condition-specific adaptations, offering a new standard for analyzing multi-condition scRNA-seq data in regenerative medicine and aging research.