Gene Gradients Reveal Directed Structural Connectivity Across Species

Scientists have developed a new way to figure out the direction of signal flow in the brain using genetic information and MRI scans. By combining how genes are expressed in different brain regions with how brain activity spreads, they can predict which way white matter pathways are pointing. This method works in worms, mice, and monkeys. It even helps explain the hierarchical organization of the human brain.

The core challenge lies in a fundamental limitation of modern neuroimaging. We can map the "roads" of the brain—the white matter tracts that connect different regions—using diffusion MRI (dMRI) tractography. However, these maps are undirected. They tell us that a connection exists between Region A and Region B. They cannot tell us if the signal travels from A to B or B to A. This is akin to having a map of highways that shows which cities are connected but fails to indicate which direction traffic is flowing. Without knowing this directionality, we cannot truly understand how information is processed.

The blind spot in tractography

Current efforts to map human structural connectivity (SC) rely heavily on dMRI tractography. This technique excels at measuring the strength of connections. However, it fails to capture their orientation. While invasive techniques like tracer-based studies can map directional pathways in mice or primates, they are not feasible for human subjects. This leaves a massive gap in our understanding of the "directed connectome." This is the complete wiring diagram that specifies which neurons transmit to others.

Without directionality, we lose the ability to distinguish between "sources" (regions that send signals) and "sinks" (regions that receive them). This distinction is vital for understanding how the brain organizes itself into hierarchies. These hierarchies move from basic sensory processing to complex, high-level cognition. The authors argue that we need a framework to estimate this hidden directionality using non-invasive, biologically grounded data.

Linking genes to signal flow

To solve this, the authors introduce a structure-function computational model. It treats gene expression as a compass for structural directionality. The mechanism follows a structured pipeline to turn undirected connections into a directed graph:

1. Extracting Gene Gradients: The researchers first derive "gene gradients." These are the principal components (mathematical summaries of variation) of regional gene co-expression patterns. Think of these as topographic maps showing how genetic programs shift across the brain.
2. Defining Asymmetry: These gradients create a node-level asymmetry metric ($G_w$). A positive value biases connections away from a region, making it a source. A negative value biases them toward the region, making it a sink.
3. Transforming the Connectome: This metric performs a similarity transform on the existing undirected structural connectivity matrix ($C$). This produces a predicted directed SC ($\tilde{C}$) .

4. Optimizing via the Lyapunov Equation: To ensure the predicted directionality is physically plausible, the authors fit the model to empirical functional connectivity data. They use the Lyapunov Equation. This is a mathematical tool describing the stationary covariance (the stable, long-term relationship between signals) of a linear stochastic process. By minimizing the error in this equation, they find the optimal gene weights .

The model also incorporates a higher-order network diffusion (HONeD) operator. This accounts for multi-step connections rather than just immediate neighbors. This allows the model to simulate how signals diffuse through the network over time.

Validating across the phylogenetic tree

The researchers tested the model's ability to recover "ground truth" directionality. These are cases where the actual direction of connections was already known through invasive methods. The paper reports that the model successfully predicted directed edges across three different species:

• C. elegans (nematode): The model achieved a correlation of $r = 0.70$ with actual neuron-to-neuron synaptic directionality [Figure 2a]. This indicates a strong match between predicted and real connections.
• Mouse: The model showed an $r = 0.57$ correlation with tracer-based directionality [Figure 2c].
• Macaque: The model yielded an $r = 0.46$ correlation [Figure 2e].

In humans, where ground truth is unavailable, the authors used "test-retest fingerprinting." This ensures the results are reliable across different scans of the same person. They found that using $k=5$ gene gradients provided the best stability. This setup achieved an Intraclass Correlation Coefficient (ICC) of $0.51 \pm 0.03$ . This score suggests the model produces reproducible results for individual subjects.

The model also revealed striking biological patterns. In humans, the predicted directionality matched established knowledge regarding the bias of sensory areas toward the thalamus [Figure 3d]. Furthermore, the model's "angular flow" (AF)—a new measure of directed functional flow—strongly correlated with the principal functional connectivity gradient ($r = 0.80$) [Figure 5f]. This suggests that the brain's functional hierarchy may emerge from the source-sink organization of signal flow through its physical structure.

Constraints and biological trade-offs

Despite these successes, the authors highlight several important caveats. First, the human directed SC presented here is a model-based prediction. It is not a direct anatomical measurement. While the results are consistent with known biology, they should be viewed as sophisticated hypotheses.

Second, the model's performance in macaques was notably lower than in simpler organisms. The authors suggest this may be due to sparse or incomplete macaque gene expression maps. Low spatial resolution in existing datasets may also play a role [Figure 2e].

Finally, there is a mathematical trade-off in the "angular flow" measure. Because the model relies on a linear diffusion framework, it may face challenges. It might struggle to distinguish between an active signal sender and an inhibitory influence. An inhibitory influence might effectively reverse the perceived flow. Practitioners using AF to map causality must remember its nature. It represents an aggregate "tilt" of signal movement through the network. It is not a strictly pairwise causal link between two specific regions.

The verdict: A new lens for connectomics

The authors have provided a powerful, biologically constrained framework for recovering the "hidden" directionality of the human brain. By anchoring mathematical diffusion models in the reality of gene expression, they have bridged the gap between molecular biology and macroscale neuroimaging.

For researchers looking to move beyond simple "road maps" toward true "traffic flow" models, this approach is highly promising. The ability to derive directed functional flow (AF) from undirected structural data is a scalable way to study brain hierarchy. While the model is not yet a replacement for direct anatomical tracing, it serves as a robust, non-invasive proxy. It makes the invisible directions of the human connectome visible.

Figures from the paper