PROTOTYPE-BASED CONTINUAL LEARNING FOR SINGLE-CELL ANNOTATION

The paper introduces scEvolver, a prototype-based continual learning framework that enables scalable and accurate single-cell annotation by incrementally refining cell-type representations without revisiting historical data, thereby overcoming catastrophic forgetting and batch biases while revealing context-specific cellular dynamics in complex diseases.

The Gist

Imagine you are a biological librarian trying to organize a massive, ever-growing library of single-cell data. Every day, new books (cells) arrive from different authors (labs), written in different languages (sequencing platforms), and about different topics (tissues).

Your job is to label every book correctly so scientists can find them later. But here's the catch: you can't keep all the old books on your desk. You have limited space, and you can't go back to the library to re-shelve everything every time a new shipment arrives. If you try to learn from the new books without looking at the old ones, you might forget how to label the old ones. This is called "catastrophic forgetting."

This paper introduces scEvolver, a smart new system that acts like a super-librarian who never forgets and keeps getting smarter without needing to re-read the entire library.

Here is how it works, broken down into simple concepts:

1. The "Mental Filing Cabinet" (Prototypes)

Instead of trying to memorize every single cell (which is impossible), scEvolver creates a "mental prototype" for each cell type.

• The Analogy: Imagine you have a "Mental Image" of what a Red Blood Cell looks like. It's not one specific photo; it's the average idea of a red blood cell.
• How it works: When a new cell arrives, scEvolver doesn't ask, "Is this exactly like the 500th red blood cell I saw yesterday?" Instead, it asks, "Does this look like my Mental Image of a Red Blood Cell?"
• The Magic: As new data comes in, it gently updates that "Mental Image" to be more accurate, without erasing the old knowledge. It's like refining your definition of "dog" as you meet more breeds, without forgetting what a dog is.

2. The "Time-Traveling Notebook" (Memory Bank)

To make sure it doesn't forget old cell types when learning new ones, scEvolver keeps a special Memory Bank.

• The Analogy: Think of this as a highlighted notebook. When the librarian learns something new, they don't just throw away the old notes. They keep a few "hard-to-remember" examples in their pocket.
• How it works: When the system learns a new batch of cells, it occasionally pulls out a few "old" examples from its notebook to review. This keeps the old labels fresh in its mind, preventing it from forgetting how to identify rare cell types (like a specific type of immune cell) just because it's busy learning about new ones.

3. The "Universal Translator" (Cross-Platform & Cross-Tissue)

Cells from different labs often look different due to technical noise (like photos taken with different cameras).

• The Analogy: Imagine trying to recognize a friend whether they are wearing a winter coat, a summer dress, or a raincoat.
• How it works: scEvolver learns to ignore the "clothing" (the technical noise from different machines) and focuses on the "face" (the true biological identity). It can take a cell from a kidney, a pancreas, or a tumor, and realize, "Hey, this is still a T-cell," even if the data looks slightly different.

4. The "Spot the Imposter" Detector (Outlier Detection)

Sometimes, a new cell arrives that doesn't fit any known category.

• The Analogy: Imagine you have a mental image of a "Cat." If a Dog walks in, your mental image says, "That doesn't look like a cat at all!"
• How it works: scEvolver measures the distance between the new cell and its "Mental Images." If the cell is too far away from any known prototype, the system flags it as a "New Discovery" or an anomaly, rather than forcing it into a wrong category. This is crucial for finding new disease states.

5. The "Few-Shot" Superpower

Usually, AI needs thousands of examples to learn a new category. scEvolver is amazing at learning with very few examples (like seeing just 5 cells of a new type).

• The Analogy: Most students need to read a whole textbook to understand a concept. scEvolver is like a genius student who can understand a new concept after seeing just a few examples and connecting them to what it already knows.

Why Does This Matter? (The Real-World Impact)

The researchers tested scEvolver on real disease data, specifically looking at inflammatory gut diseases.

• The Discovery: They found a subtle change in gut cells. Some cells were starting to transform into a different shape (metaplasia) to fight inflammation.
• The Result: Because scEvolver could track these tiny, gradual changes in the "Mental Image" of the cells, it spotted this disease progression earlier and more accurately than previous methods.

Summary

scEvolver is a smart, evolving AI system that:

1. Never forgets old knowledge while learning new things.
2. Adapts to new data formats without needing to retrain from scratch.
3. Finds new discoveries by spotting cells that don't fit the mold.
4. Works with very little data, making it perfect for rare diseases.

It turns the chaotic, messy world of single-cell biology into an organized, ever-updating encyclopedia that helps doctors and scientists understand how diseases change over time.

Technical Summary

1. Problem Statement

The rapid expansion of single-cell atlases has created a critical bottleneck in cell-type annotation. Existing frameworks face three major challenges:

• Static Nature & Scalability: Most methods rely on static reference datasets. As new data emerges (from different platforms, tissues, or modalities), models require retraining on the entire historical dataset, which is computationally expensive and often impossible due to privacy regulations preventing data sharing.
• Catastrophic Forgetting: When models are updated with new data sequentially, they tend to lose knowledge of previously learned cell types, a phenomenon known as catastrophic forgetting.
• Batch Effects and Imbalance: Single-cell data suffers from severe batch effects (technical variation across platforms like 10x, CEL-Seq, etc.) and class imbalance (rare cell types vs. abundant ones), leading to biased predictions and poor generalization across tissues and modalities.

2. Methodology: scEvolver

The authors propose scEvolver, a prototype-based continual learning framework designed to incrementally integrate biological knowledge without revisiting historical data.

Core Architecture

• Backbone: Built upon the scGPT foundation model, which is pre-trained on large-scale single-cell data to capture rich biological semantics.
• Parameter-Efficient Fine-Tuning (PEFT): Instead of retraining the entire model, scEvolver uses Low-Rank Adaptation (LoRA) combined with a Mixture-of-Experts (MoE) module. This freezes the pre-trained weights and only updates a small number of trainable parameters, preserving prior knowledge while adapting to new data.
• Shared Latent Space: The model maps cell expression profiles into a shared embedding space where cell types are represented by class prototypes (centroids).

Key Mechanisms

1. Memory-Augmented Prototypes:
   • Instead of storing raw historical data, the model maintains a Memory Bank of class prototypes ($M_p$).
   • As new data arrives, prototypes are updated using a high-momentum rule ($p_c \leftarrow \alpha p_c + (1-\alpha)\bar{p}_c$), ensuring smooth evolution of cell-type representations while retaining historical semantic consistency.
2. Dual-Level Memory Replay:
   • Prototype Memory: Stores the global structure of learned classes.
   • Sample Replay: A buffer of "hard examples" (selected based on high prediction entropy and distance to prototypes) is replayed during training. This focuses learning on decision boundaries and rare cell types, mitigating forgetting.
3. Memory-Augmented Prototypical Proxy Loss (MAPPL):
   • A novel loss function that optimizes the distance between a sample and its memory-augmented prototype (current + historical).
   • It enforces intra-class compactness (pulling samples toward their prototype) and inter-class separation (pushing samples away from other class prototypes), effectively handling class imbalance.
4. Expandable Classification Head:
   • The framework supports class-incremental learning, allowing new cell types to be added to the label space dynamically without modifying the feature extractor or retraining on old classes.
5. Cross-Modal Adaptation:
   • For multimodal data (e.g., ATAC+RNA, ADT+RNA), the model uses adversarial learning and masked token prediction to learn modality-invariant representations, bridging gaps between different sequencing technologies.

3. Key Contributions

• Scalable Continual Learning: scEvolver is the first framework to enable continual single-cell annotation that avoids catastrophic forgetting without requiring access to historical raw data, addressing privacy and storage constraints.
• Prototype-Centric Representation: By treating cell types as evolving prototypes, the model maintains consistent semantics across batches, tissues, and modalities, effectively harmonizing heterogeneous data.
• Robustness to Scarcity: The framework demonstrates superior performance in few-shot settings (using only 5 labeled cells per class) and handles severe class imbalance better than offline baselines.
• Interpretability & Outlier Detection: The distance between a cell and its prototype serves as a metric for cell-state transitions and outlier detection. Cells far from their prototype can be flagged as novel populations or disease-associated states.

4. Results

The authors evaluated scEvolver across multiple benchmarks:

• Cross-Platform & Cross-Tissue: On the PANCREAS (9 batches, 5 platforms) and MYELOID (8 cancer types) datasets, scEvolver achieved the highest Macro-F1 scores (0.9584 and 0.8023, respectively) among online methods, outperforming static baselines and other continual learning approaches. It successfully aligned cells of the same type across different platforms while minimizing batch effects.
• Cross-Modal Integration: On the BMMC dataset (RNA, ATAC, ADT), scEvolver achieved robust alignment across modalities, outperforming scNym and scGPT in latent space coherence and batch correction scores.
• Few-Shot Performance: In few-shot scenarios (5 labeled cells/class), scEvolver improved Macro-F1 scores by 24.5% (PANCREAS) and 11.6% (MYELOID) compared to offline baselines, proving its efficacy with limited supervision.
• Catastrophic Forgetting: Forgetting curves showed that scEvolver maintains stable performance on previously learned batches, whereas other models (like scNym and scGPT) exhibited significant performance drops when new data was introduced.
• Biological Discovery: Applied to inflammatory gut disease data, scEvolver identified metaplastic transitions in epithelial cells, specifically distinguishing canonical Surface Foveolar (SF) cells from SF-like cells. It quantified this transition via prototype distance and linked it to specific gene signatures (e.g., CEACAM7, LCN2) and pathway enrichments (e.g., inflammasome activation).

5. Significance

• Dynamic Reference Atlases: scEvolver enables the construction of "living" reference atlases that evolve with new data, crucial for precision medicine and longitudinal disease studies.
• Privacy-Preserving: By relying on prototype updates and memory replay rather than raw data storage, it facilitates knowledge sharing across institutions without violating data privacy regulations.
• Biological Insight: The framework moves beyond simple classification; the prototype distance metric provides a quantitative measure of cellular state deviation, enabling the discovery of subtle, continuous biological transitions (e.g., disease progression) that discrete clustering might miss.
• Resource Efficiency: By leveraging PEFT and avoiding full retraining, it significantly reduces the computational cost and time required for updating large-scale single-cell models.

In summary, scEvolver represents a paradigm shift from static, one-time training to a dynamic, continual learning approach for single-cell analysis, offering a scalable, privacy-compliant, and biologically interpretable solution for the next generation of cellular atlases.