TopicVI: A Knowledge-guided deep interpretable model for resolving context-specific gene programs

TopicVI is a deep interpretable model that integrates prior biological knowledge with data-driven refinement via optimal transport to discover context-specific gene programs, outperforming existing methods in benchmarking and successfully revealing convergent tumor states in glioblastoma.

The Gist

Imagine you walk into a massive, chaotic library where millions of books (cells) are thrown onto the floor. Each book contains thousands of pages (genes) written in a complex code. Your goal is to organize these books into meaningful sections (like "History," "Science," "Fiction") and understand what specific stories are being told in each section.

This is the challenge scientists face with single-cell RNA sequencing. They have data on individual cells, but it's noisy, messy, and often looks like a jumbled mess. Traditional methods try to sort these books by looking at the cover or just guessing based on what they already know. But sometimes, a book looks like "History" but is actually telling a "Science" story because of a specific event (like a disease or a drug).

Enter TopicVI, a new "super-librarian" AI developed by the researchers in this paper. Here is how it works, explained simply:

1. The Problem: The "Rigid" vs. "Wild" Librarian

• The Old Way (Rigid): Some methods rely entirely on a pre-written catalog (prior knowledge). They say, "If a book has these words, it must be History." The problem? If a cell is sick or reacting to a drug, it might use "History" words to tell a "Science" story. The rigid librarian misses this nuance.
• The Other Way (Wild): Other methods ignore the catalog and just group books by whatever words appear together most often. The problem? They might create weird groups that make no sense biologically, like grouping "Cooking" and "Astrophysics" together just because both use the word "heat."

2. The Solution: The "Smart" Librarian (TopicVI)

TopicVI is a hybrid. It acts like a librarian who respects the old catalog but is smart enough to realize the books have changed.

• The "Optimal Transport" Magic: Imagine you have a map of where books should go based on the old catalog. TopicVI uses a special mathematical tool (called Optimal Transport) to gently nudge the books. It asks: "This book looks like it belongs in History, but the data says it's actually acting like Science. Let's move it just enough to fit the new story, without throwing the whole library into chaos."
• The Result: It creates groups that are interpretable (we know what they mean because they connect to known biology) but also flexible (they can discover new, disease-specific patterns that the old catalog didn't know about).

3. Real-World Examples from the Paper

Example A: The Blood Cell Detective (PBMCs)

Think of immune cells in your blood as a crowd of people at a concert. Some are just standing there (naive cells), while others are jumping and dancing (activated cells).

• The Challenge: Activated cells look almost identical to normal ones; they just have a little more "energy."
• TopicVI's Win: While other methods saw a blurry crowd, TopicVI used its "smart librarian" approach to spot the subtle differences. It found a specific group of T-cells that were "dancing" (activated) because of a specific signal (Interferon), and it even found a hidden group of cells that looked like "Others" but were actually specific types of immune cells (macrophages) that no one had noticed before.

Example B: The Brain Map (Spatial Transcriptomics)

Imagine looking at a brain slice where different layers are like floors in a skyscraper.

• The Challenge: The "disease" (like Alzheimer's) is happening on every floor, making it hard to tell which genes are just part of the "floor" (anatomy) and which are part of the "disease."
• TopicVI's Win: The researchers told TopicVI, "Ignore the disease, just tell me about the floors." TopicVI successfully separated the "floor" signals from the "disease" signals. It even refined the "floor" map, realizing that the old catalog had some extra genes that didn't actually belong to that specific floor, making the map much sharper and clearer.

Example C: The Cancer Drug Test (Glioblastoma)

Imagine testing a cancer drug on tumor cells.

• The Challenge: You want to know why the drug works. Does it stop the cell from dividing? Does it trigger self-destruction?
• TopicVI's Win: It didn't just say "The drug worked." It broke the reaction down into specific "stories" (topics).
  • It found that the drug triggered a "Cell Cycle Stop" story.
  • It found a new, mysterious "Stress Response" story (Topic 32) that the old catalog didn't have.
  • The Big Discovery: They realized that patients with a specific genetic mutation (EGFR) couldn't tell this "Stress Response" story, which explained why the drug didn't work for them. This is a huge clue for personalized medicine.

The Bottom Line

TopicVI is like a translator that speaks both "Old Biology" (what we already know) and "New Data" (what is actually happening right now).

Instead of forcing new data into old boxes, or ignoring the old boxes entirely, it builds new, flexible boxes that fit the data perfectly while still making sense to human scientists. This helps doctors and researchers understand complex diseases, find new drug targets, and see the hidden stories inside our cells that were previously invisible.

Technical Summary

1. Problem Statement

Current analytical pipelines for single-cell (scRNA-seq) and spatial transcriptomics rely heavily on cell clustering followed by differential expression analysis and manual interpretation using curated biological priors. This approach faces several critical limitations:

• Context-Specificity Gap: Curated biological databases (priors) are static and often fail to capture context-specific gene programs, particularly in complex disease states or under drug perturbations.
• Interpretability vs. Discovery Trade-off: Existing methods force a choice between prior-guided models (high interpretability but rigid, missing novel context-specific variations) and de novo models (flexible discovery but often biologically uninterpretable and inconsistent across studies).
• Signal Confounding: In spatial transcriptomics and disease contexts, gene expression is driven by multiple overlapping factors (e.g., anatomical location vs. disease state). Traditional methods struggle to disentangle these signals.
• Decoupled Modeling: Most methods treat cell clustering and gene program discovery as independent tasks, ignoring their mutual influence.

2. Methodology: TopicVI Architecture

TopicVI is a deep interpretable model that integrates Variational Autoencoders (VAE), Non-negative Matrix Factorization (NMF), and Topic Modeling (inspired by Latent Dirichlet Allocation) to jointly infer cell clusters and gene topics.

Core Components:

• Cell Representation Module (VAE): Uses a VAE architecture to learn low-dimensional cell embeddings, effectively handling technical noise (dropouts) via a Zero-Inflated Negative Binomial (ZINB) distribution.
• Deep Biclustering Module: Performs soft clustering on cell embeddings using differentiable cluster centers. It jointly models:
  1. Cell Cluster Assignment: Probability of a cell belonging to a cluster.
  2. Cluster-Topic Loading: How active specific topics are within each cluster.
  3. Topic-Gene Weights: The contribution of genes to specific topics.
• Optimal Transport (OT) for Knowledge Guidance: This is the novel core mechanism. Instead of rigidly constraining topics to match prior gene programs (PGPs), TopicVI uses Optimal Transport to align data-derived topics with PGPs.
  • It calculates distances between genes in PGPs and constructed topics based on negative log-probabilities.
  • A Sinkhorn divergence algorithm solves the OT problem, allowing topics to remain consistent with prior knowledge while permitting data-driven refinements (e.g., excluding irrelevant genes or adding context-specific genes).
• Semi-Supervised Learning: The model can operate in unsupervised mode or supervised mode (using focal loss) to explicitly disentangle specific biological variables (e.g., anatomy vs. disease) by treating annotations as labels.

Loss Function:
The total loss combines:

• Likelihood Loss: Reconstructs expression data (ZINB).
• Topic Loss: Includes Optimal Transport loss (aligning with PGPs) and a similarity loss (ensuring topics are distinct).
• Clustering Loss: Includes adaptive loss, deep clustering loss (DCE), and cluster centrality loss (Davies-Bouldin inspired) to prevent cluster collapse.

3. Key Contributions

• Unified Framework: First model to jointly optimize cell clustering and gene topic discovery, leveraging the mutual influence between cell states and gene programs.
• Flexible Knowledge Integration: Introduces an Optimal Transport module that treats prior knowledge as a guide rather than a constraint, enabling the discovery of context-specific modifications to established pathways.
• Signal Disentanglement: Demonstrates the ability to separate confounding biological signals (e.g., anatomical location vs. disease state) through supervised learning modes.
• Rare Cell Identification: Significantly improves the detection of rare cell subtypes compared to state-of-the-art methods.

4. Key Results

A. Benchmarking (Human Lung Cell Atlas - HLCA)

• Performance: TopicVI achieved the highest aggregate score (0.67) across 13 methods, outperforming competitors in bio-conservation (alignment with manual annotations), batch correction, and topic coherence.
• Interpretability: Models using PGP input showed superior overlap with known biological functions. Crucially, TopicVI with PGP input generated topics that explained cell-type variance better than the raw prior programs themselves, proving its ability to refine static knowledge with data.
• Rare Cells: TopicVI achieved the highest Isolated Label F1 score, demonstrating superior ability to identify rare cell subpopulations compared to methods like scANVI and scVI.

B. Application: PBMC Immune Cell Subtypes

• Fine-Grained Resolution: Successfully distinguished closely related myeloid subtypes (e.g., FCGR3A+ vs. CD14+ monocytes) and subdivided CD4+ T cells into naive and activated states based on interferon signaling topics.
• Novel Discovery: Identified a previously undefined "Others" population, reclassifying them as macrophage/neutrophil clusters based on specific topic signatures (apoptosis, myeloid markers).
• Robustness: Outperformed scANVI in the Zheng68k dataset, where marker genes from other datasets failed due to batch effects, highlighting TopicVI's data-adaptive nature.

C. Application: Human Brain Spatial Transcriptomics

• Disentangling Signals: In supervised mode, TopicVI successfully separated anatomical layer signals from disease (Alzheimer's) signals.
• Refinement of Pathways: For a topic related to NMDA receptor activation, TopicVI refined the canonical Reactome pathway by excluding 13 irrelevant genes (67% retention). This refined topic showed significantly higher spatial specificity (F-statistic improvement) than the original pathway.
• Disease Specificity: Supervised training on disease labels successfully isolated apoptosis and mTOR pathways associated with Alzheimer's, distinct from anatomy-driven topics.

D. Application: Glioblastoma (GBM) Drug Response

• Drug-Responsive Clusters: Identified specific tumor subpopulations enriched in etoposide-treated cells, linked to cell cycle regulation (APC/C) and MECP2 pathways.
• Convergent Phenotypes: Revealed that distinct drugs (bortezomib and etoposide) converge on a shared "dual-hit" vulnerability (Topic 12: cell cycle arrest).
• Novel Biomarker Discovery: Identified Topic 32 (de novo), a gene program involving cell cycle, DNA damage, and EGFR co-mutation.
  • Clinical Relevance: High Topic 32 scores correlated with longer progression-free survival in TCGA-GBM patients.
  • Genotype Dependence: This survival benefit was observed only in EGFR-wild-type tumors, suggesting EGFR mutations override this stress-response mechanism.

5. Significance

• Bridging the Gap: TopicVI resolves the long-standing trade-off between biological interpretability and data-driven discovery, enabling researchers to find novel, context-specific gene programs without sacrificing biological grounding.
• Mechanistic Insight: By disentangling confounding factors (anatomy vs. disease) and identifying convergent drug responses, it provides deeper mechanistic insights into complex diseases and therapeutic mechanisms.
• Clinical Translation: The identification of genotype-dependent prognostic markers (e.g., Topic 32 in EGFR-wild-type GBM) highlights the potential for precision medicine applications, such as patient stratification and drug repurposing.
• Scalability: The model is applicable to diverse data modalities (scRNA-seq, spatial transcriptomics, pseudo-bulk) and offers a robust framework for virtual cell modeling and out-of-distribution prediction.