SpaTRACE: Spatiotemporal recurrent auto-encoder for reconstructing signaling and regulatory networks from spatiotemporal transcriptomics data

SpaTRACE is a transformer-based framework that reconstructs dynamic cell-cell communication and gene regulatory networks from spatiotemporal transcriptomics data by modeling time-lagged dependencies without relying on predefined ligand-receptor databases.

The Gist

Imagine your body is a bustling, high-tech city. Inside this city, billions of cells are the citizens. To keep the city running smoothly—whether it's growing a new building (development), fixing a pothole (regeneration), or dealing with a riot (disease)—these citizens need to talk to each other constantly.

Some citizens send out messages (signals), while others listen and change their behavior based on what they hear. This is how your body knows to grow a new brain cell or heal a wound.

For a long time, scientists had two big problems trying to map these conversations:

1. They were looking at a static photo: Most tools only looked at a single snapshot in time, assuming everyone was standing still. But cells are always moving, changing, and evolving.
2. They needed a pre-written phone book: Most tools relied on a "known list" of who talks to whom. If a cell invented a new way to communicate that wasn't in the book, the tools missed it completely.

Enter SpaTRACE. Think of it as a super-powered, time-traveling detective that doesn't need a phone book.

How SpaTRACE Works: The "Time-Traveling Detective"

Instead of taking a single photo, SpaTRACE watches a movie of the cells. It uses a special kind of artificial intelligence (a "spatiotemporal recurrent auto-encoder") that understands two things at once:

• Space: Where the cells are sitting next to each other.
• Time: How the cells change as they move through their life cycle (like a caterpillar turning into a butterfly).

Here is the magic trick: SpaTRACE uses a "predictive guessing game."

Imagine you are watching a row of dominoes fall. If you see the first few dominoes fall, you can predict exactly which one will fall next. SpaTRACE does this with genes. It looks at the current state of a cell and the signals it's receiving from its neighbors, then asks: "Based on what I see right now, what will this cell's genes look like a moment from now?"

• If the guess is right: The model learns that the signal it saw was important.
• If the guess is wrong: It realizes it missed a connection.

By playing this game millions of times, the model figures out the hidden rules of the city without ever needing a pre-written list of rules. It discovers new ways cells talk to each other that scientists didn't know about before.

The Three Superpowers

SpaTRACE solves three specific mysteries that other tools struggle with:

1. The "Who Talked to Whom?" Mystery (Ligand-Receptor Matching)

• The Old Way: "I know Cell A talks to Cell B because they are neighbors." (But maybe they aren't actually talking!)
• The SpaTRACE Way: "I see Cell A send a signal, and I see Cell B change its behavior immediately after. Therefore, they are definitely talking." It connects the sender to the receiver based on the result of the conversation, not just their proximity.

2. The "What Happened Inside?" Mystery (Gene Regulation)

• Once a cell receives a message, it has to decide what to do. Does it grow? Does it divide? Does it die?
• SpaTRACE traces the message from the outside (the signal) all the way to the inside (the genes that turn on or off). It builds a map showing exactly which "switches" inside the cell were flipped by which outside message.

3. The "No Phone Book" Advantage

• Because it learns by watching the movie of time, it doesn't need a database of known interactions. It can find brand new conversations in species (like axolotls) or situations where scientists have never looked before.

Real-World Adventures

The paper shows SpaTRACE solving two real-life puzzles:

• The Mouse Brain Construction Site: Scientists used SpaTRACE to watch how a mouse brain grows. They found specific "construction managers" (signaling molecules) that tell stem cells when to stop being generic and start becoming specific types of neurons. It's like finding the exact blueprint that tells a bricklayer, "Now, build a window, not a wall."
• The Axolotl Super-Healer: Axolotls are famous for regrowing their brains. Scientists used SpaTRACE to watch the healing process day-by-day. They discovered that different "healing crews" show up at different times. One crew arrives early to clean up, and a different crew arrives later to rebuild. This helps us understand how to potentially heal human brains in the future.

The Bottom Line

Think of SpaTRACE as a dynamic, self-learning translator for the language of life. Instead of just reading a dictionary of known words, it watches the conversation unfold in real-time, figures out the grammar, and discovers new words that no one knew existed.

It turns a blurry, static photo of a busy city into a high-definition, 4D movie, showing us exactly how the citizens of our bodies coordinate to build, repair, and sustain life.

Technical Summary

1. Problem Statement

Current computational methods for analyzing cell-cell communication (CCC) and gene regulatory networks (GRNs) face two critical limitations when applied to developmental and regenerative processes:

• Reliance on Curated Databases: Most CCC inference tools (e.g., CellPhoneDB, CellChat) depend heavily on pre-defined ligand-receptor (LR) databases. This restricts their applicability to well-studied species and prevents the discovery of novel, unannotated interactions.
• Static Assumptions: Existing methods typically analyze static gene expression profiles, assuming steady-state conditions. They fail to capture the temporal dynamics and causal dependencies inherent in dynamic biological processes like development and tissue regeneration.
• Disjoint Analysis: CCC and GRN inference are often performed separately, ignoring the joint orchestration of extracellular signaling and intracellular transcriptional regulation.

2. Methodology: SpaTRACE Framework

SpaTRACE is a transformer-based, temporal-causal framework designed to jointly infer intercellular signaling and intracellular regulatory networks from spatiotemporal transcriptomics (ST) data without relying on predefined LR databases.

Core Architecture

The model utilizes an attention-based recurrent autoencoder to model cellular dynamics along pseudotime-sampled trajectories.

• Input: Spatial transcriptomics data including gene expression ($X$), spatial coordinates, cell types, and pseudotime values ($\pi$).
• Encoder:
  • Constructs contextualized gene representations at each time point.
  • Uses a single self-attention layer (without value projections) to preserve interpretability.
  • Generates per-cell embeddings by modulating global gene embeddings with expression-dependent weights derived from the cell's molecular profile (Ligands, Receptors, TFs, Target Genes).
• Decoder:
  • Predicts the future target gene (TG) expression state ($x_{t+1}$) based on recent intra- and intercellular signals ($x_{t-k:t}$).
  • Inputs include current TG state, recent Transcription Factor (TF) abundances, and recent Ligand-Receptor (LR) co-expression signals.
  • The prediction is formulated as a weighted sum of past signals, where weights are derived from attention scores.

Inference Mechanisms

The model leverages the Granger causality principle: if past molecular states improve the prediction of future target expression, a causal regulatory relationship is inferred.

1. Signaling & Regulatory Network Reconstruction:
   • LR-TG & TF-TG: Infers regulatory interactions by analyzing attention scores (dynamic, context-specific) and global embedding similarities (context-invariant affinity).
   • Pathway-Free: Does not require a pre-defined LR database; interactions are learned directly from the data's temporal dynamics.
2. Ligand-Receptor Matching:
   • Aggregates LR-TG scores to identify which ligand-receptor pairs share common downstream target genes, effectively reconstructing binding relationships.
3. Cell-Cell Communication (CCC) Inference:
   • Reconstructs spatially resolved CCC networks by aggregating pairwise effects based on sender ligand abundance, receiver receptor abundance, and the inferred downstream potency of the LR pair.

3. Key Contributions

1. Temporal-Causal Modeling: Integrates spatial context with pseudotime trajectories to capture dynamic intra- and intercellular signaling, moving beyond static correlation-based methods.
2. Pathway-Free Multi-Layer Reconstruction: Enables the de novo reconstruction of multi-layer networks (LR binding, LR-TG signaling, and TF-TG regulation) at single-cell resolution without relying on curated interaction databases.
3. Joint Inference: Simultaneously recovers both GRNs and CCC networks, quantifying per-cell regulatory activity and signaling strength.

4. Results and Evaluation

Synthetic Benchmarks

SpaTRACE was evaluated on two synthetic datasets with known ground truths (one low-noise, one high-noise with perturbed trajectories).

• CCC Inference: SpaTRACE significantly outperformed existing methods (COMMOT, SpaTALK, CellPhoneDB, CellChat) in terms of AUPRC (Area Under the Precision-Recall Curve).
  • Noisy Setting: SpaTRACE achieved 0.831 (Dataset 1) and 0.614 (Dataset 2), substantially higher than the next best method (COMMOT: 0.589 and 0.538).
  • Robustness: Unlike database-dependent methods, SpaTRACE maintained high performance even when LR annotations were perturbed or removed.
• GRN Inference: SpaTRACE achieved the highest AUPRC (0.967 and 0.903) compared to GENIE3, GRNBoost2, and Velorama, demonstrating superior ability to recover causal regulatory edges.

Real-World Applications

1. Mouse Midbrain Development (MOSTA Dataset):
   • Validation: Evaluated using Early Precision Ratio (EPR) against curated databases. SpaTRACE achieved an EPR@100 of 9.87 for LR binding and 50 for GRN reconstruction, indicating a 50-fold enrichment of true interactions in top predictions compared to random chance.
   • Biological Insights: Identified stage-specific signaling programs driving radial glial cell (RGC) differentiation into neuroblasts (NeuB) and glioblasts (GlioB). Key interactions included PTN–EPHB1, PTN–CD44, and TGFB3–LRP6, which showed strong activity in progenitor states but diminished as cells differentiated.
2. Axolotl Brain Regeneration:
   • Applied to a dataset covering 2–60 days post-injury.
   • Identified stage-specific signaling dynamics (e.g., VIM-associated signaling at 10–15 DPI, PSAP-associated at 15–20 DPI).
   • Discovered candidate interactions (e.g., GPI–ENO1, VIM–ENO1) not currently annotated in standard mouse/human LR databases but supported by functional evidence in STRING.

5. Significance

• Overcoming Database Bias: By learning interactions directly from spatiotemporal data, SpaTRACE enables the discovery of novel signaling pathways in understudied species or contexts where curated databases are incomplete.
• Dynamic Resolution: It provides a mechanistic link between extracellular cues and intracellular transcriptional responses over time, offering a more accurate model of development and regeneration than static snapshots.
• Interpretability: The attention-based architecture allows for direct interpretation of regulatory weights as putative driver-target effects, facilitating biological hypothesis generation.
• Robustness: The framework demonstrates stability across different trajectory inference tools and is resilient to noise and missing annotations, making it a robust tool for analyzing complex spatiotemporal datasets.

In conclusion, SpaTRACE represents a significant advancement in computational genomics by unifying spatiotemporal modeling with causal inference, providing a powerful tool for reconstructing the dynamic regulatory landscapes of developing and regenerating tissues.