PathFinder: Joint Decompositions of Linked Multimodal Datasets

The human brain is a vast, intricate landscape, and scientists have long sought to map its hidden patterns. To do this, they often gather data from many different angles: scanning the brain's electrical activity, measuring its blood flow, or even looking at the genes that shape it. Each of these methods produces a massive grid of numbers, a snapshot of the brain from a specific perspective. For years, researchers have used mathematical tools to break these grids down, looking for the simple, underlying structures that repeat across the noise. These tools are powerful because they are purely data-driven; they do not need a pre-written script to tell them what to look for, allowing them to reveal hidden connections in complex systems like the brain. However, a significant hurdle has always existed: these tools generally require that all the data being compared shares a common feature, like the same group of people or the same brain regions. If you try to compare the brain of a human with that of a monkey, or if you have data for some people but not others, the standard methods often break down because the grids do not line up.

A team of researchers at the University of Oxford has developed a new approach to solve this problem, a method they call PathFinder. Their work addresses a fundamental limitation in how scientists combine different types of brain data. Imagine trying to solve a puzzle where the pieces come from different boxes, and some pieces are missing entirely. Traditional methods would insist that every piece must fit perfectly into a single, pre-defined frame. PathFinder, by contrast, is more flexible. It realizes that even if the entire collection of data does not share one common dimension, smaller groups within the data might still overlap. As long as there is a chain of connections—a path—that links one piece of data to another, the method can weave them all together into a single, coherent picture. This allows scientists to find common patterns across different species, different types of scans, or even datasets where some information is missing, without forcing the data to fit a rigid structure that it does not naturally possess.

The researchers demonstrated the power of this new method through a series of tests, starting with computer simulations. They created a grid of six different data sets, representing two types of measurements across three different groups. In their test, they deliberately removed two of these data sets, pretending they were missing, and asked the algorithm to figure out what those missing pieces should look like based on the remaining four. The method succeeded, accurately reconstructing the missing data and even outperforming older techniques when the data was noisy. This proved that the method could not only find patterns in what was present but could also predict what was absent, effectively filling in the blanks of a complex puzzle using the logic of the surrounding pieces.

To see if this worked in the real world, the team applied PathFinder to actual brain scans from the Human Connectome Project. In one experiment, they looked at data from a visual task where subjects watched rotating wedges and moving bars on a screen. They had recordings of the brain's activity and the visual patterns themselves, organized into a grid of six different runs. They then hid one specific visual pattern—the counter-clockwise rotating wedge—from the computer. Using only the brain data and the other five visual patterns, PathFinder successfully predicted the missing visual pattern. The result was so precise that the predicted image matched the actual stimulus almost perfectly. However, when they tried the same trick but removed half of the brain data, the prediction failed, showing that the method relies on having enough connected information to build a reliable bridge across the gaps.

The team also showed that PathFinder is a general framework that can do what several existing, more complicated methods do, but in a simpler, unified way. For instance, a common technique called dual regression, used to map brain networks in individual people, can be re-created using PathFinder in a single step. This suggests that PathFinder is not just a new trick, but a broader umbrella under which many previous methods can sit, offering a more flexible way to handle data that doesn't fit neatly into a single box.

Perhaps the most striking application involved the study of loops in the brain that connect the cortex to deeper structures like the thalamus and the basal ganglia. These connections form a cycle, like a ring road, where information flows from one region to the next and eventually back to the start. The researchers used diffusion MRI data from ten human subjects to map the structural connections between these regions. They arranged the data in a way that reflected this circular flow, creating a chain of matrices where each one shared a dimension with its neighbor. By applying PathFinder to this cyclic arrangement, they were able to uncover the three major functional pathways within these loops: the motor pathway, the associative pathway, and the limbic pathway. The method successfully grouped the brain voxels into these distinct circuits, revealing a clear anatomical organization that matched what scientists already knew about these brain systems.

The researchers are careful to note that while the method is powerful, it is not a magic wand that solves every problem instantly. The mathematical nature of the decomposition means that the results can sometimes be ambiguous, requiring extra steps to make them unique and easy to interpret. In their experiments, they used a follow-up process to rotate the results into a more understandable form, which helped them recover the true underlying structures in their simulations and produce anatomically sensible maps in real data. They also acknowledge that the current version of the tool works best when data can be arranged in a two-dimensional grid, though they have already begun thinking about how to expand it to handle more complex, multi-dimensional data and even non-linear relationships.

Ultimately, this work opens the door to comparing brains across species and scales in ways that were previously impossible. Because the method does not require a one-to-one match between every single data point, it allows researchers to use data from one species, such as a monkey, to help interpret data from another, like a human, even if they do not have the exact same measurements for both. It turns the challenge of missing or incompatible data into an opportunity to find shared patterns across the biological world. By providing a way to link disparate datasets through simple, overlapping connections, PathFinder offers a new lens through which to view the complex, interconnected nature of the brain, turning fragmented pieces of information into a unified story of how we think, move, and feel.