📄 cell biology

Inferring Cell-Cell Interaction Dynamics from Cell Trajectory Data Using Deep Attention Networks

This paper presents a deep attention network-based model and an accompanying interpretability toolkit that accurately infer and visualize the specific rules governing cell-cell interaction dynamics directly from trajectory data, without requiring prior assumptions about biological mechanisms.

Original authors: Boyle, J., Baker, R. E., Byrne, H. M.

Published 2026-07-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Boyle, J., Baker, R. E., Byrne, H. M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a bustling city where millions of tiny, self-driving cars are zooming around. Sometimes they move in perfect, swirling traffic jams; other times, they scatter like startled birds. In the world of biology, these "cars" are cells, and their movement is the engine behind everything from how a baby grows in the womb to how a tumor spreads or how the immune system fights off an infection. For a long time, scientists have been able to take photos of these cells and track their paths, like watching a time-lapse movie of a busy highway. But knowing where the cars went is different from knowing why they turned, sped up, or stopped. Did they avoid a pothole? Did they follow a friend? Did they get pushed by a stranger? Figuring out the invisible rules that govern these microscopic dance moves has been a massive puzzle.

Enter a new kind of digital detective: a "Deep Attention Network." Think of this not as a robot that guesses, but as a super-observant student who watches thousands of hours of cell traffic footage and learns the rules of the road without being told what they are. Instead of just asking, "Did the car turn left or right?" this new model asks, "Exactly how far did the car move, and in what direction?" By breaking down the movement into tiny interactions with neighbors, this model can peek behind the curtain to see if cells are pulling each other closer, pushing each other away, or simply trying to match their speed with the crowd. It's like finally understanding the secret language of a swarm, revealing whether the cells are acting like a coordinated team or just a chaotic crowd.

The New Detective Tool

In this paper, James Boyle and his team at the University of Oxford introduce a fresh, powerful way to decode cell movement using this "Deep Attention Network." Unlike older methods that tried to guess simple yes-or-no answers (like "will this cell turn?"), their model predicts the entire movement vector—essentially calculating the exact step a cell will take next. They trained this digital brain on real data from a specific type of kidney cell called MDCK cells, which are famous for moving in tight, swirling groups.

The model works by looking at a single "focal" cell and its ten closest neighbors. It asks a simple question for each neighbor: "How much does this specific neighbor influence the focal cell's next move?" The model learns to assign a "weight" or "attention score" to each neighbor. If a neighbor is directly in front, the model might say, "This one matters a lot!" If a neighbor is far away, it might say, "Ignore them." By stacking thousands of these observations together, the researchers created "attention heatmaps." These are colorful maps that show exactly where a cell is "looking" or paying attention. For the MDCK cells, the heatmaps revealed a clear pattern: the cells pay the most attention to the neighbors directly in front of them and the ones immediately behind, while ignoring those to the far sides.

Uncovering the Hidden Rules

But the researchers didn't stop at just seeing who pays attention to whom. They wanted to know how that attention changes the cell's behavior. To do this, they invented a clever trick: they asked the model to imagine a scenario where a specific neighbor suddenly vanished. By comparing where the focal cell would have gone with the neighbor present versus where it would have gone without them, the model could calculate the specific effect of that interaction.

Using this method, they created two new types of maps:

Attraction-Repulsion Heatmaps: These show if a neighbor pulls a cell closer or pushes it away. When they tested this on simulated data (where they knew the rules beforehand), the model correctly identified zones of strong repulsion at short distances and weak attraction at medium distances. When applied to the real MDCK cells, it showed that cells tend to be repelled by neighbors to their side, rear, or immediate front.
Alignment Heatmaps: These check if cells try to match their speed and direction with their neighbors. The model found that MDCK cells do indeed try to align their velocities with neighbors in all directions, but this effect is strongest for those in front or to the side.

Why This Model is Different

The authors are careful to point out that their new model is a significant upgrade over previous versions. Older models tried to predict just the "turning" behavior (left or right), which often led to confusing maps that looked like fans or had strange gaps. The authors suggest these odd shapes weren't real biological rules but were just artifacts of asking the wrong question. By predicting the full movement distance instead of just a turn, and by removing unnecessary mathematical constraints, their new "generalised deep attention network" (gDAN) produces much cleaner, more consistent results.

They also tested their model on simulated data from the famous "d'Orsogna model," where the rules of attraction and repulsion were programmed in by hand. The model successfully recovered these hidden rules, proving it can learn genuine dynamics rather than just memorizing noise. Furthermore, they showed that the model can be extended to handle systems with different types of cells (like leaders and followers), suggesting it could one day help untangle the complex interactions in immune responses or cancer growth.

The Bottom Line

The paper suggests that deep attention networks are a powerful, flexible tool for understanding cell movement. It doesn't claim to have solved every mystery of biology, but it offers a new way to look at the data that is both accurate and easy to interpret. By letting the data speak for itself without forcing it into pre-made boxes, this approach helps scientists see the invisible forces that guide the microscopic world, turning a chaotic swarm of dots into a story of connection, push, and pull.

Technical Summary: Inferring Cell–Cell Interaction Dynamics from Cell Trajectory Data Using Deep Attention Networks

Problem Statement
Cell movement is a fundamental driver of biological processes such as embryonic development, collective migration, tumor invasion, and immune responses. A primary mechanism governing this movement is the interaction between spatially proximal cells, which can occur via short-range forces (adhesion, friction, repulsion) or longer-range mechanisms (chemical signaling, physical force transmission). While advances in live-cell imaging and tracking algorithms have made it possible to collect high-resolution single-cell trajectory data, a significant challenge remains: inferring the specific interaction rules that govern these movements in a manner that is both accurate and biologically interpretable. Existing methods often rely on a priori assumptions about interaction mechanisms or are limited to predicting discrete behaviors (e.g., turning left or right) rather than continuous movement vectors.

Methodology
The authors propose a novel deep learning model based on Deep Attention Networks designed to learn cell–cell interaction dynamics directly from trajectory data without hypothesis-driven constraints.

Model Architecture: The model predicts the movement vector ( $\Delta x$ ) of a "focal cell" over a time interval $[t, t + \Delta t)$ based on its current velocity ( $v_f$ ) and the relative displacement ( $x_j$ ) and velocity ( $v_j$ ) of its $N$ nearest neighbors.
Decomposition: The prediction is formulated as a weighted sum of pairwise interactions:
$\Delta x = \sum_{j=1}^{N} \frac{\Pi(x_j, v_f, v_j) W(\tilde{x}_j)}{\sum_{k=1}^{N} W(\tilde{x}_k)}$
Here, $\Pi$ is a fully connected neural network representing the pairwise interaction function (the effect of a neighbor on the focal cell's movement), and $W$ is a weight function representing the interaction strength (the relative influence of a neighbor based on displacement). The tilde ( $\tilde{x}$ ) denotes coordinates aligned with the focal cell's velocity.
Training: The networks $\Pi$ and $W$ are trained using backpropagation and adaptive moment estimation on trajectory data. The model is "hypothesis-free," allowing the exact interaction dynamics to be learned directly from the data.
Generalized Deep Attention Network (gDAN): The authors distinguish their model (gDAN) from previous "turn deep attention networks" (tDAN). Unlike tDANs, which predict binary turning behavior and impose symmetry constraints, the gDAN predicts the full continuous displacement vector and removes hard symmetry constraints and velocity dependence from the weight function to improve interpretability and robustness.

Key Contributions

Continuous Vector Prediction: The model shifts the predictive target from binary turning behavior to the complete movement vector, enabling the inference of how interactions alter cell movement magnitude and direction, not just turning probability.
Type-Dependent Dynamics: The architecture is extended to handle heterogeneous cell populations by including cell types ( $\tau$ ) as input variables, allowing for the inference of type-dependent interaction dynamics (e.g., leader-follower relationships).
Interpretability Tools: The authors develop a suite of tools to visualize the learned dynamics:
- Attention Heatmaps: Visualize the relative influence (weight) of neighbors based on their position relative to the focal cell.
- Attraction–Repulsion Heatmaps: By comparing predicted movement with and without a specific neighbor, the model calculates an "attraction score" to determine if interactions are attractive or repulsive at specific distances.
- Alignment Heatmaps: By calculating an "alignment score" based on cosine correlation, the model infers velocity alignment behaviors.
Validation: The approach is validated using both experimental data (Madin–Darby Canine Kidney (MDCK) epithelial cells) and simulated data (d'Orsogna and Vicsek models), demonstrating the ability to recover known ground-truth interaction rules (e.g., short-range repulsion, medium-range attraction, velocity alignment).

Results

MDCK Cell Analysis: Applied to MDCK cells, the gDAN reveals that these cells derive most movement cues from neighbors directly in front of them. The model identifies that cells are repelled by neighbors to the side, rear, or immediately in front, while interactions further ahead are more complex.
Comparison with tDAN: The gDAN produces more consistent attention heatmaps for both "bulk" and "edge" cells in the MDCK assay compared to the tDAN. The tDAN exhibited artifacts (e.g., "fans" of attention to the front-left and front-right) likely caused by the binary nature of the turning prediction, whereas the gDAN's continuous prediction target yields results more reflective of underlying biology.
Simulated Data Recovery: When trained on d'Orsogna model data, the gDAN successfully recovers zones of strong repulsion and weak attraction. When trained on Vicsek model data, it correctly identifies the perceptual zone where velocity alignment occurs.
Interaction Potentials: The authors demonstrate that radially averaging the learned weight function can qualitatively recover features of interaction potentials (force vs. distance), suggesting the model can indicate whether an interaction potential is an appropriate mechanistic description.

Significance
The paper claims that this approach represents a highly general methodology for linking cell trajectory data to mechanistic hypotheses. By combining data-driven learning with structural interpretability, the deep attention network serves as a powerful exploratory tool. It moves beyond merely inferring whether cells interact to inferring how these interactions affect movement (e.g., attraction, repulsion, alignment) and allows for the analysis of type-dependent dynamics in complex, multi-cellular systems. The removal of rigid symmetry constraints and the shift to continuous vector prediction are presented as critical steps toward making deep attention networks a reliable, general-purpose tool for analyzing cell movement in diverse biological systems.

More like this