ARCADIA Reveals Spatially Dependent Transcriptional Programs through Integration of scRNA-seq and Spatial Proteomics

ARCADIA is a novel generative framework that integrates unpaired single-cell RNA sequencing and spatial proteomics data by aligning modality-specific archetypes to reconstruct tissue architecture and uncover spatially dependent transcriptional programs without requiring cell barcode pairing or direct feature correspondence.

The Gist

Imagine you are trying to understand a bustling, complex city. You have two different ways of looking at it, but both have major flaws:

1. The "Dissociated" View (scRNA-seq): Imagine you take a photo of every single person in the city, but you rip them out of their neighborhoods and dump them all into one giant, featureless warehouse. You can see exactly what job they do, what they are wearing, and what they are thinking (their genes). But you have no idea where they live, who their neighbors are, or how the neighborhood influences them.
2. The "Neighborhood" View (Spatial Proteomics): Now imagine you have a satellite map of the city. You can see exactly where every person is standing and who their neighbors are. You can see the "vibe" of the neighborhood. However, your camera is old and blurry; it can only tell you a few basic things about the people (like "wearing a red shirt" or "holding a tool"), but it can't read their minds or see their full personality (their full genetic program).

The Problem: Scientists want to combine these two views to understand how a person's job (genes) changes based on where they live (spatial context). But current tools are like trying to match the warehouse photos to the satellite map by looking for people wearing the exact same red shirt. If the satellite map doesn't have a "red shirt" sensor, or if the person in the warehouse isn't wearing one, the matching fails.

The Solution: ARCADIA
The authors of this paper built a new tool called ARCADIA. Think of ARCADIA as a brilliant translator and architect that doesn't need to match specific items (like red shirts) to connect the two worlds.

Here is how it works, using simple analogies:

1. Finding the "Extreme Personalities" (Archetypes)

Instead of trying to match every single person, ARCADIA looks for the most "extreme" examples of each type of person in both datasets.

• In the warehouse, it finds the "Ultimate Chef," the "Ultimate Athlete," and the "Ultimate Artist."
• In the satellite map, it finds the "Ultimate Chef" (maybe someone surrounded by food stalls) and the "Ultimate Athlete" (someone surrounded by a gym).
• The Magic: Even if the satellite camera can't see the chef's apron, it can recognize the context of the chef. ARCADIA says, "Okay, the 'Ultimate Chef' in the warehouse must correspond to the 'Ultimate Chef' in the satellite map, even if we can't see the exact same details." These "Ultimate" examples are called Archetypes.

2. Building a Shared Map (Alignment)

Once ARCADIA has identified these extreme archetypes in both worlds, it uses them as "anchor points" to build a shared map.

• It draws a line between the Warehouse Chef and the Satellite Chef.
• It draws a line between the Warehouse Athlete and the Satellite Athlete.
• Once these anchors are set, it fills in the rest of the map. It can now guess where the "Average Chef" from the warehouse belongs on the satellite map, simply because they are close to the "Ultimate Chef" anchor.

3. The "Dual Brain" (Deep Learning)

ARCADIA uses two artificial brains (called Variational Autoencoders) that talk to each other.

• Brain A looks at the warehouse photos (RNA).
• Brain B looks at the satellite map (Proteins).
• They are trained to agree on the "vibe" of the city. If Brain A sees a cell that looks like it's in a "Tumor Neighborhood," and Brain B sees a cell in a "Tumor Neighborhood," they lock hands and say, "Yes, these are the same place."

What Did They Discover?

When they applied this to real human tissue (specifically the tonsils, which are part of the immune system), ARCADIA revealed secrets that were previously hidden:

• The "Location Matters" Effect: They found that B-cells (immune soldiers) act completely differently depending on which "neighborhood" they are in. Deep inside a " Germinal Center" (a training camp), they are busy mutating and learning. On the outskirts, they are acting as guards or messengers.
• T-Cell Exhaustion: They saw that T-cells only become "exhausted" (tired and giving up) when they are stuck in specific, crowded neighborhoods. If they were just looking at the cells in isolation, they would have missed this crucial link between where the cell is and how it behaves.

Why is this a Big Deal?

Previous tools were like trying to solve a puzzle by matching the shape of the pieces. If the pieces didn't fit perfectly (because the data was different), the puzzle broke.

ARCADIA is like solving the puzzle by looking at the picture on the box. It understands the story and the structure of the data, allowing scientists to mix and match different types of biological data without needing them to be perfectly identical. It lets us finally see how the "neighborhood" shapes the "person," helping us understand diseases like cancer and autoimmune disorders much better.

In short: ARCADIA is the ultimate translator that connects the "Who am I?" (Genetics) with the "Where am I?" (Location), revealing that in biology, context is everything.

Technical Summary

1. Problem Statement

The integration of single-cell RNA sequencing (scRNA-seq) and spatial proteomics (e.g., CODEX, MIBI-TOF) is critical for understanding how tissue microenvironments shape cellular phenotypes. However, current integration methods face significant limitations:

• Loss of Context: scRNA-seq provides high-resolution transcriptomic data but loses spatial organization due to tissue dissociation.
• Limited Scope: Spatial proteomics captures spatial context but is restricted to limited protein panels, constraining the inference of full transcriptomic programs.
• Data Availability Constraints: Existing integration methods often rely on strong linkage (paired measurements with shared cell barcodes, e.g., CITE-seq) or feature-level correspondence (requiring known gene-protein mappings).
• The Gap: There is a lack of robust methods for weak-linkage integration (where datasets are from separate studies with no shared barcodes or direct feature overlap) that can simultaneously preserve spatial neighborhood information and infer transcriptomic programs.

2. Methodology: ARCADIA Framework

The authors propose ARCADIA (ARchetype-based Clustering and Alignment with Dual Integrative Autoencoders), a generative framework designed for weak-linkage, cross-modal integration without assuming direct feature-to-feature correspondence.

Core Architecture

• Dual Variational Autoencoders (VAEs): ARCADIA employs two separate VAEs, one for scRNA-seq data and one for spatial proteomics data. These encoders project high-dimensional inputs into latent spaces.
• Archetype-Based Alignment: Instead of aligning individual cells or features, ARCADIA identifies archetypes—convex combinations of cells representing extreme phenotypic states (vertices of a polytope).
  • Learning Archetypes: Principal Convex Hull Analysis (PCHA) is used to decompose each modality's data into $k$ archetypes.
  • Cross-Modal Alignment: Archetypes are aligned across modalities by minimizing the cosine distance between their cell-type composition profiles. This establishes a biological correspondence based on shared cell-type distributions rather than specific gene-protein pairs.
• Anchor-Guided Regularization: Cells dominated by a single archetype (high match weight) serve as "anchors." A matching loss function penalizes the distance between corresponding anchors in the latent spaces of the two modalities, ensuring high-confidence alignment.

Spatial Feature Engineering

• Neighborhood Features: For spatial proteomics, the model constructs a spatial graph (using physical coordinates) to compute neighborhood-mean features (average protein expression of the 15 nearest neighbors).
• Input Representation: These neighborhood features are concatenated with cell-specific protein markers to create spatially aware vectors, allowing the model to learn how local microenvironments influence cell states.

Objective Functions

The total training objective combines four components:

1. Evidence Lower Bound (ELBO): Ensures faithful reconstruction of input data (ZINB for RNA, Gaussian for proteins) while regularizing latent geometry.
2. Structure Preservation Loss: Aligns the affinity matrices of cell-type centroids in the latent space with those in the archetype space, preserving known biological relationships.
3. Cross-Modal Maximum Mean Discrepancy (MMD): Minimizes the distributional distance between cell types across modalities in the latent space, encouraging mixing without feature-level alignment.
4. Anchor-Guided Loss: Uses softplus penalties to enforce that high-confidence anchor cells from different modalities are close in the latent space, while non-matching cells remain distant.

3. Key Contributions

• Novel Integration Paradigm: ARCADIA is the first method to integrate scRNA-seq and spatial proteomics using archetype-based alignment without requiring paired barcodes or gene-protein mappings.
• Spatially Aware Deep Learning: It explicitly incorporates spatial neighborhood information into the proteomic encoder, enabling the model to learn how spatial context drives transcriptional programs.
• Bidirectional Translation: The framework supports:
  • Imputing spatial niche labels for scRNA-seq cells.
  • Reconstructing gene programs for spatial proteomics cells based on their inferred niches.
• Scalability: The model is designed to run efficiently on standard hardware (e.g., NVIDIA T4 GPUs) even for large datasets (~180k cells).

4. Results and Evaluation

Datasets

1. Semi-Synthetic CITE-seq Dataset: A ground-truth dataset created by splitting a paired CITE-seq dataset into disjoint RNA and protein sets, with artificial spatial grids assigned to protein cells.
2. Human Tonsil Dataset: Independent scRNA-seq and CODEX datasets from human tonsils (the only publicly available experimentally paired dataset from separate studies).

Performance Benchmarks

• Spatial Separation: ARCADIA significantly outperformed state-of-the-art weak-linkage methods (MaxFuse, scMODAL) in recovering spatially dependent cell states.
  • k-Sep (k-nearest neighbor separation): ARCADIA achieved 0.666 (vs. 0.403 for MaxFuse) on the semi-synthetic data, indicating superior separation of cell neighborhoods (CNs) within cell types.
  • CN F1 Score: ARCADIA achieved 0.979, vastly outperforming competitors (~0.55), demonstrating its ability to predict spatial locations from transcriptomic data.
• Cell Type Integration: ARCADIA maintained high accuracy in cell type clustering (ARI Score ~0.97) comparable to other methods, proving it does not sacrifice cell identity for spatial integration.
• Computational Efficiency: Training on ~180k cells took ~1.8 hours with peak GPU memory usage of only 3.2 GB, demonstrating high scalability.

Biological Discoveries (Human Tonsil)

ARCADIA successfully reconstructed known tissue architecture and revealed novel spatially dependent transcriptional programs:

• B-Cell Maturation: Identified a gradient where B cells in central germinal centers (CN 3) expressed proliferation and hypermutation markers (e.g., BCL6, AICDA), while peripheral B cells (CN 7, 9) expressed plasma-cell and antigen-presentation markers.
• T-Cell Exhaustion: Revealed that CD8+ T cells in specific niches (CN 8) co-expressed cytotoxic and exhaustion markers (e.g., TIGIT, KLRG1), whereas others in different niches maintained stem-like or effector states.
• T-Follicular Helper (Tfh) States: Distinguished Tfh-like states in specific niches (CN 1) versus terminally exhausted phenotypes in others (CN 4).

5. Significance

ARCADIA addresses a critical bottleneck in spatial biology by enabling the integration of high-dimensional transcriptomics with spatially resolved proteomics without the need for expensive, paired experimental designs.

• Mechanistic Insight: It allows researchers to map how specific microenvironmental niches (spatial context) drive specific transcriptional programs (phenotypic states), such as T-cell exhaustion in tumor microenvironments or B-cell maturation in lymphoid organs.
• Generalizability: By relying on composition-level priors (cell-type distributions) rather than feature-level links, ARCADIA can be applied to any combination of scRNA-seq and spatial proteomics datasets, regardless of the specific markers used.
• Future Impact: The framework provides a generative bridge for bidirectional modeling, potentially enabling the prediction of full gene expression profiles from limited spatial protein panels, thereby overcoming the throughput and resolution limitations of current spatial technologies.