RF-PHATE: The AI Visualization Tool Revealing Hidden Patterns in Complex Biological Data

The Core Problem: Human Eyes Can’t See High-Dimensional Data

Our visual system is built for three dimensions. Biological data doesn’t care about that limitation.

Clinical datasets tracking disease progression at the cellular level can contain hundreds of thousands of data points — measurements spanning treatments, outcomes, genetic markers, and cellular states simultaneously. No scatter plot handles that gracefully.

“Biological processes are an example of complex, high-dimensional data,” says Kevin Moon, director of USU’s Data Science and Artificial Intelligence Center. “These datasets include hundreds of thousands of data points on disease progression at the cellular level, along with treatments and clinical outcomes.”

That’s where machine learning-powered visualization tools step in. But until now, even the best options had significant blind spots.

What’s Wrong With Existing Visualization Methods

The current go-to tools — PHATE, t-SNE, and UMAP — are widely used and genuinely useful. But they share a critical weakness.

These unsupervised methods tend to over-emphasize differences between data groups. In doing so, they lose something equally important: how those groups relate to each other structurally.

Supervised alternatives exist, but they don’t fully solve the problem either. The result is that researchers often get visualizations that highlight separation without preserving the underlying biological trajectory — the continuous, meaningful progression from one state to another.

That gap is exactly what RF-PHATE was built to close.

What RF-PHATE Actually Does

RF-PHATE stands for Random Forest-Potential of Heat-diffusion for Affinity-based Trajectory Embedding. The name is dense, but the concept is precise.

It’s a supervised data visualization method that combines the predictive power of Random Forests with PHATE‘s trajectory-preserving geometry. The result is a tool that doesn’t just separate data clusters — it maps how those clusters connect, evolve, and relate to each other across multiple dimensions.

In practical terms, RF-PHATE lets researchers explore relevant data relationships in multidimensional datasets while keeping the structural integrity of those relationships intact. That’s a meaningful step forward for any field dealing with continuous biological processes.

The Multiple Sclerosis Breakthrough

The most striking result in the paper involves multiple sclerosis patient data.

Using RF-PHATE on clinical MS datasets, the research team found evidence supporting a previously suspected but unconfirmed MS subtype. That distinction matters enormously in clinical practice.

“Identifying subtypes is crucial, because MS affects each patient differently, and knowing the specific type guides treatment decisions,” Moon explains.

This isn’t just an academic finding. If RF-PHATE can reliably surface subtype distinctions that other visualization methods miss, it becomes a genuine diagnostic support tool — one that could influence how neurologists stratify patients and select therapies.

Beyond MS: COVID-19, Lung Cancer, and More

The team didn’t stop at multiple sclerosis. They stress-tested RF-PHATE across multiple biological domains.

Additional validation datasets included:

• COVID-19 patient plasma data — tracking immune response patterns across disease severity
• Antioxidant-treated lung cancer cell data — mapping how cellular states shift under treatment conditions

Each dataset presented different structural challenges. RF-PHATE handled them consistently, preserving trajectory relationships that other methods distorted or discarded.

And critically, Moon is clear that biological data is just the starting point.

Why This Matters Beyond Biology

RF-PHATE isn’t a niche tool for computational biologists. Its architecture makes it applicable to any domain with high-dimensional, structured data — finance, materials science, climate modeling, social network analysis.

More importantly for the AI tools ecosystem, Moon highlights two additional use cases that go beyond visualization:

1. Building more interpretable AI models. RF-PHATE’s supervised structure can inform how complex models are designed, making their internal logic more transparent.

2. Analyzing AI models themselves. The same visualization approach that maps biological trajectories can be applied to understand how a machine learning model is actually behaving — a growing priority as AI systems become more consequential.

This positions RF-PHATE not just as a research instrument, but as a potential component in the broader interpretable AI stack.

The Research Team and Institutional Backing

The paper was led by Jake Rhodes (USU Ph.D. ’22, statistics), now an assistant professor at Brigham Young University, with Moon as corresponding author. Co-authors include Adele Cutler (USU), Anhong Zhou (USU), and Wei Zhang (University of Utah).

Funding came from the National Institutes of Health and the IVADO Visiting Scholar Program, with additional national and international collaborators contributing to the work.

The research sits within USU’s DSAI Center, which actively promotes the AI for Science movement — an international initiative using AI and machine learning to accelerate research, analyze massive datasets, and simulate complex systems.

What This Means for AI Tool Adopters

If you’re evaluating AI tools for data analysis or scientific research, RF-PHATE signals something worth paying attention to: the next competitive edge in high-dimensional analytics isn’t just processing speed — it’s structural fidelity.

Tools that preserve the relationships between data points, not just the distances between clusters, will produce more actionable insights. That’s true whether you’re a biotech researcher, a data scientist working with complex behavioral data, or an AI engineer trying to understand your own model’s behavior.

The gap between “visualizing data” and “understanding data” is where RF-PHATE operates. And based on the results published in Nature Computational Science, it’s closing that gap in ways the existing toolkit simply couldn’t.

The best AI tools don’t just show you more data — they show you what the data means. RF-PHATE is a strong early example of what supervised visualization can do when it’s built with structural integrity as the primary design goal. Watch this space.