A single-cell and spatial atlas of prostate cancer reveals the combinatorial nature of gene modules underlying lineage plasticity and metastasis

This study presents a comprehensive single-cell and spatial atlas of prostate cancer from 128 patients that reveals how combinatorial gene modules drive lineage plasticity and metastasis, identifies distinct tumor microenvironment features, and introduces a transformer-based foundation model for automated cell-state classification.

The Gist

Imagine the prostate gland as a bustling city. In a healthy city, the buildings (cells) have specific jobs: some are the "administrators" (luminal cells) that follow strict rules and produce a specific report (PSA), while others are the "construction workers" (basal cells) or "specialized maintenance crews" (club cells).

Now, imagine Prostate Cancer not as a single invading army, but as a chaotic city where the buildings are losing their job descriptions, changing their architecture, and sending out confusing signals.

This paper is like a massive, high-definition city map created by scientists who interviewed over half a million individual "citizens" (cells) from 128 different patients. They used advanced technology to see exactly what each cell was doing, where it was located, and how it was changing.

Here are the key discoveries from this "city map," explained simply:

1. The "Imposter" Buildings (Lineage Plasticity)

In a healthy prostate, cancer usually starts in the "administrators" (luminal cells). But the researchers found that in many patients, these cells are playing "imposter."

• The Analogy: Imagine an accountant (luminal cell) suddenly deciding to dress up and act like a construction worker (basal) or a specialized maintenance crew (club).
• The Finding: In early-stage cancer, you see a lot of these "imposters." They are confused and trying on different identities. However, as the cancer gets worse and spreads (metastasis), these imposters disappear. The cancer cells stop trying to be different things and just become aggressive, fast-growing machines. The "confusion" is actually a sign that the cancer is still trying to find its footing, but once it spreads, it becomes very focused on one goal: survival and growth.

2. The "Broken ID Cards" (Loss of Identity)

Healthy prostate cells have a clear ID card with a list of 8 specific "keywords" (genes) that say, "I am a prostate cell."

• The Analogy: Think of these keywords as a uniform. A healthy worker wears the full uniform.
• The Finding: The researchers found that cancer cells are losing pieces of their uniform. Some lose one piece, some lose five, and some lose almost all of them.
  • In early cancer, the cells are still wearing most of the uniform, but it's frayed.
  • In advanced cancer, the uniform is almost gone. The cells have become "lineage infidel" (unfaithful to their origin).
  • Why it matters: This "fraying" happens before the cancer looks dangerous under a microscope. It's a molecular warning sign that the cell is losing its way.

3. The "Specialized Spy Networks" (Metastasis)

When cancer spreads to other parts of the body (like the bone, brain, or liver), it doesn't just copy-paste itself. It adapts.

• The Analogy: Imagine a spy sent to a new country. To survive, they must learn the local language and customs.
  • Bone Metastasis: The cancer cells start acting like bone cells, using "osteomimetic" (bone-mimicking) signals to blend in.
  • Brain Metastasis: They turn on "neuro-migratory" genes, acting like neurons to navigate the brain.
  • Liver Metastasis: They switch to "erythroid-like" (blood cell-like) programs.
• The Finding: The cancer isn't just one thing; it's a shape-shifter that changes its software to fit the environment it lands in.

4. The "Neuroendocrine" Chameleon (NEPC)

Some prostate cancers transform into a very aggressive type called Neuroendocrine Prostate Cancer (NEPC).

• The Analogy: This is like a chameleon that completely changes its skin color and texture to become a completely different animal.
• The Finding: The researchers discovered that NEPC isn't just one single type. It's a mix-and-match of different "programs." Some cells turn on a "HES6" program (a new discovery), while others turn on different switches.
• The Good News: They built a new, sharper "detector" (a 18-gene signature) that can spot these dangerous chameleons much better than old methods, even when they are hiding in early stages.

5. The "City Guards" (Immune System & Race)

The study looked at the "police force" (immune cells) protecting the city.

• The Analogy: In some neighborhoods, the police are calm and organized. In others, they are in a state of high alert and inflammation.
• The Finding: The researchers found a significant difference based on ancestry. Patients of African American descent had a much higher presence of Th17 cells.
  • What are Th17 cells? Think of them as "firefighters" who are great at putting out small fires but can sometimes cause a lot of smoke and inflammation.
  • The Implication: These "firefighters" were found in higher numbers in African American patients, even in early-stage cancer. This suggests that the immune environment in these patients is more "inflamed," which might explain why prostate cancer can be more aggressive in this population.

6. The "Hidden Neighbors" (Schwann Cells)

Using a super-high-resolution camera (spatial transcriptomics), they found a rare type of cell called Schwann cells (which usually wrap around nerves).

• The Analogy: These are like the electricians who usually only work on the power lines.
• The Finding: They found these electricians hanging out right next to the cancer buildings, and even inside the tumor. They seem to be having a secret conversation with the cancer cells, possibly helping the cancer spread along the nerves (perineural invasion).

7. The "AI City Planner" (PCformer)

Finally, the researchers built a new Artificial Intelligence tool called PCformer.

• The Analogy: Imagine a super-intelligent city planner who has read every single building permit and blueprint in the city.
• The Function: This AI can look at a messy, confusing cell and instantly say, "Ah, this is a tumor cell," or "This is a T-cell," with 94% accuracy. It automates the boring work of sorting cells, allowing doctors and scientists to focus on finding cures.

The Big Picture

This paper tells us that prostate cancer is not a simple, static enemy. It is a dynamic, shape-shifting ecosystem.

• It starts with cells losing their identity.
• It adapts to new environments when it spreads.
• It interacts differently with the immune system depending on the patient's background.
• And it hides rare, helpful neighbors (like Schwann cells) that might be helping it grow.

By mapping all of this, the scientists have given us a much clearer "GPS" for navigating prostate cancer, which should lead to better tests to catch it early and smarter drugs to stop it from changing its shape.

Technical Summary

1. Problem Statement

Prostate cancer (PCa) is characterized by profound molecular and cellular heterogeneity, which drives therapeutic resistance, metastasis, and poor clinical outcomes. While bulk RNA sequencing has identified broad subtypes, it fails to resolve the complex cellular diversity within the tumor microenvironment (TME) or the continuous spectrum of lineage plasticity in cancer cells.
Key challenges addressed in this study include:

• Limited Metastatic Data: Previous single-cell studies lacked sufficient metastatic samples across different organ sites to draw robust conclusions about organ-specific adaptations.
• Lineage Plasticity: The mechanisms by which luminal cells lose identity (lineage infidelity) and transition to neuroendocrine or other states are not fully understood at single-cell resolution.
• Ancestry Disparities: There is a lack of high-resolution data explaining immune and stromal differences between African American (AFR) and European American (EUR) patients.
• Annotation Bottlenecks: Existing methods for cell-type annotation in PCa are often inconsistent or lack a unified framework for the diverse cell states observed.

2. Methodology

The authors constructed a comprehensive, harmonized single-cell atlas and developed a deep learning foundation model.

• Data Integration:
  • Integrated 128 patients across 14 datasets (12 published, 2 unpublished).
  • Cohort: 95 localized, 10 castration-resistant (CRPC), and 23 metastatic CRPC (mCRPC) samples from 7 organ sites.
  • Scale: ~560,492 cells annotated into 9 major populations (epithelial, immune, stromal, endothelial) and further sub-clustered.
• Computational Framework:
  • Gene Module Discovery: Used Celda (Bayesian co-clustering) to identify 166 co-expression gene modules.
  • Gene Program Association Studies (GPAS): Applied Linear Mixed Effect (LME) models to associate modules with clinical phenotypes (e.g., metastatic site, Gleason score) while correcting for study-level batch effects.
  • Tumor Identification: Combined inferCNV (copy number variation) with a custom XGBoost classifier to distinguish malignant from non-malignant epithelial cells.
  • Spatial Transcriptomics: Utilized 10x Genomics Xenium 5K platform on FFPE samples to validate cell states (basal-like, club-like, Schwann cells) in situ.
  • Deep Learning Model (PCformer): Developed a transformer-based foundation model trained on >500,000 cells. It uses masked language modeling (MLM) on ranked gene expression sequences to learn prostate-specific gene co-expression patterns, followed by fine-tuning for cell-type classification.
• Validation:
  • Multiplex immunofluorescence (mIF) on Tissue Microarrays (TMAs) for Th17 and Schwann cell validation.
  • Independent validation of the 18-gene Neuroendocrine signature in the SU2C bulk RNA-seq dataset.

3. Key Contributions & Results

A. Tumor Cell States and Lineage Plasticity

• Basal-like and Club-like Populations: Identified distinct tumor subpopulations expressing basal (TP63+) and club (LTF+, LCN2+) signatures. These were prevalent in localized disease (8–9% of epithelial cells) but nearly absent in metastatic samples, suggesting progressive loss of these states during disease advancement.
• Luminal Lineage Infidelity: Luminal cancer cells do not simply lose identity; they exhibit combinatorial erosion of luminal gene modules (e.g., KLK3, AR, NKX3-1).
  • High-grade (Gleason 10) and advanced-stage tumors showed significantly lower "luminal infidelity scores" (fewer expressed luminal modules).
  • This infidelity is driven by specific somatic alterations (e.g., SPOP, CHD1, APC mutations) and correlates with disease severity.
• PSA Heterogeneity: PSA expression (KLK2/3) is not binary. Cells were categorized into four states: Depleted, Down-regulated, Baseline, and Amplified.
  • "Kallikrein-Depleted" cells often co-express basal markers (TP63) and are enriched in radiotherapy-treated samples.
  • "Kallikrein-Amplified" cells show high AR signaling but are rare in high-grade metastatic disease.

B. Metastatic Adaptation and Organotropism

• Universal vs. Site-Specific Programs:
  • Universal: Metastatic cells broadly upregulate cell-cycle modules (G1/S and G2/M) regardless of the organ site.
  • Organ-Specific:
    • Bone: Enrichment of osteomimetic signaling (e.g., LGALS1, hematopoietic regulators).
    • Brain: Induction of neuro-migratory genes (EPHA4, MAP1B).
    • Liver: Erythroid-like transitions and glutamatergic signaling.
• Immune Evasion: Metastatic lesions show a loss of MHC Class I and PD-L1 associated modules, correlating with an immunosuppressive TME (enriched Tregs, reduced CD8+ T cells).

C. Neuroendocrine Prostate Cancer (NEPC) Landscape

• Combinatorial NE Programs: NEPC is not monolithic. The study identified two "core" modules (one containing canonical markers like SCG2/3, and a novel HES6 module) and two "supporting" modules.
• Continuum of Differentiation: NEPC cells exist on a continuum. Intermediate NE states were detected in localized disease, suggesting molecular plasticity precedes histological transformation.
• Improved Signature: Developed a refined 18-gene NE signature that outperforms previous signatures (Beltran, Dong) in distinguishing NEPC from non-NEPC in bulk RNA-seq data (AUC > 0.95).

D. Tumor Microenvironment (TME) and Ancestry

• AFR vs. EUR Disparities:
  • Th17 Enrichment: African American patients showed a significant enrichment of pro-inflammatory Th17 T-cells (marked by RORC, CCL20, KLRB1) compared to EUR patients.
  • Spatial Context: Spatial transcriptomics confirmed Th17 cells colocalize with B-cells in tertiary lymphoid-like structures.
  • HSP+ T-cells: A stress-associated CD8+ T-cell population was enriched in AFR patients, potentially linked to technical platform differences or biological stress.
• Schwann Cells: Identified a rare Schwann cell population in the prostate TME.
  • These cells interact with tumor cells via GDF15–TGFBR2 signaling, potentially driving perineural invasion.
  • Spatial data showed Schwann cells surrounding nerve structures and interacting with tumor glands.

E. PCformer: A Foundation Model

• Performance: PCformer achieved 94% accuracy and a macro F1-score of 0.94 on a hold-out set for classifying 12 prostate cell types.
• Generalizability: Successfully annotated independent datasets with high precision, particularly for immune and stromal cells. It struggles slightly with transitional epithelial states (basal/club/tumor) due to their continuous nature, highlighting the biological complexity of plasticity.

4. Significance

• Unified Atlas: Provides the largest and most diverse single-cell resource for prostate cancer, bridging the gap between localized and metastatic disease.
• Mechanistic Insight: Demonstrates that lineage plasticity is a combinatorial process of gene module erosion rather than a binary switch, offering new targets for early detection of aggressive clones.
• Clinical Translation:
  • The 18-gene NE signature offers a robust tool for identifying NEPC in bulk sequencing, crucial for treatment selection (e.g., avoiding androgen deprivation in NEPC).
  • Identification of Th17 enrichment in AFR patients provides a molecular basis for health disparities and suggests potential immunotherapeutic targets.
  • PCformer automates cell annotation, standardizing analysis for future studies and clinical applications.
• Therapeutic Implications: Highlights specific organ-adaptive pathways (e.g., osteomimetic in bone) that could be targeted to prevent metastatic colonization.

In summary, this study moves beyond static cell-type classification to define the dynamic, combinatorial gene programs driving prostate cancer progression, offering a foundational resource for understanding tumor evolution, metastasis, and ancestry-specific disease biology.