ASU scientists unveil new methods to extract reliable truth from imperfect data
Members of the ASU team involved in the data research are (from left) Zachary Hendrix, Ayush Saurabh, Steve Presse and Pedro Pessoa. Photo courtesy of Pedro Pessoa
Scientists at Arizona State University are advancing how they interpret complex, imperfect data — challenging long-standing assumptions in fields ranging from imaging to cellular biology.
In two papers published in Nature Communications and Proceedings of the National Academy of Sciences, the team introduces fundamentally new approaches to solving “inverse problems,” where researchers work backward from noisy observations to uncover the underlying reality.
Together, the studies highlight a central issue in modern science — extracting trustworthy insights from incomplete or indirect data — and demonstrate how new methods can make those insights significantly more reliable.
Rethinking image clarity: A physics-based approach to de-blurring
In the Nature Communications study, Professor Steve Presse and his team from ASU’s School of Molecular Sciences and Department of Physics introduce a new method for de-convolving blurry images. This is an essential task in fields such as microscopy, astronomy and medical imaging.
“This sounds like something people would have done before, except interestingly, existing de-blurring methods were not based on the physics of how we know the data is collected,” Presse said. “Thus, existing methods, used widely for over 50 years, made up features that don’t exist. We have addressed that by integrating what we know about data collection to make sure to de-blur in a scientific way. In doing so, we can learn how reliable things are that we deduce from the data.”
“The new method explicitly incorporates the physics of data collection into the de-blurring process,” said Zachary Hendrix, Presse’s doctoral student and the paper’s first authorThe other authorOther authors on this paper include Juan Andreas Martinez, Vincent Vandenbroucke and Frank Delvigne from the Terra Research and Teaching Centre in the Gembloux Agro-Bio Tech at the University of Liege in Gembloux, Belgium.s on the paper are: Peter T. Brown, Tim Flanagan, Douglas P. Shepherd and Ayush Saurabh.. “By grounding the reconstruction in how images are truly generated, the approach not only produces more accurate results but also enables scientists to assess the reliability of the features they observe.”
Hendrix also explains that the conventional tool for this purpose, Richardson-Lucy deconvolution, cannot provide statistical confidence in its outcomes and, paradoxically, degrades images by amplifying noise and introducing spurious structures unless the process is halted manually.
“To overcome these fundamental limitations, we have developed DeBayes: a statistically rigorous, physics-informed deconvolution framework that reconstructs only the details of an image that its associated instrument can resolve, while simultaneously quantifying the remaining uncertainty," Hendrix said.
The method directly accounts for the microscope's optics and the camera's noise statistics when generating many candidate reconstructions consistent with the data. By combining these reconstructions, DeBayes not only recovers features substantiated by the experiment but also provides researchers with uncertainty maps that reveal where the mean reconstruction is well supported and where the data are less informative.
"More generally, our work replaces ad hoc guesswork with a principled and reliable methodology for computational imaging — one that sharpens and de-noises microscope images, thereby enhancing researchers' ability to analyze faint, complex biological structures and supporting advancement in other imaging-intensive disciplines.
"We demonstrate this with low-light imaging data of mitochondrial networks within HeLa cells, recovering high-contrast structures without the high-frequency artifacts frequently observed in results from traditional and recent deconvolution methods,” Hendrix said.
This shift marks an important step toward more trustworthy imaging, where conclusions drawn from visual data are better aligned with physical reality.
Decoding cellular memory: A new lens on stress response
In a complementary study published in PNAS, the researchers tackle a long-standing challenge in biology: understanding how cells respond to stress over time.
Generally, physics describes systems in the "forward" direction — starting with a model and predicting what data it should produce. However, real scientific discovery works in reverse. Researchers observe noisy, incomplete data and attempt to infer the hidden processes that generated it.
This inverse problem is especially difficult when studying the physics of biological systems, where “randomness” — or in statistics terminology, “stochasticity” — and history both play critical roles.
“We start from observations and try to infer the underlying processes that produced them,” said Pedro Pessoa, postdoctoral student in the Presse lab and first author on this paper. “These inverse problems are far more difficult. The challenge becomes especially severe in biology, where randomness plays a central role and the mathematical tools needed for inference are often unavailable.”
The study focuses on protein production in dividing cells, specifically examining how yeast cells activate a stress-response gene known as glc3.
A key complication arises from cellular inheritance: When cells divide, they pass proteins to their offspring. This means that the proteins observed in a cell may not reflect current activity, but rather a legacy of past generations.
“This changes how the data should be interpreted: If inheritance is ignored, protein production can appear stronger or more persistent than it really is,” Pessoa said.
Using a novel simulation-based inference framework powered by advanced neural network models, the researchers overcame this challenge. Their method allows them to infer protein production dynamics even when traditional mathematical tools fail.
The findings reveal a striking insight: What appears to be sustained gene activity is often misleading. In reality, much of the observed protein is inherited, not newly produced. When this cellular “memory” is properly accounted for, the data show that gene activation is actually rare but the proteins generated in these rare events survive many generations.
“More broadly, the study highlights a common scientific problem: Simulation is often easy, but inference is hard. Our results show that it is now possible to tackle these difficult inverse problems without throwing away the very biological features that make them interesting in the first place,” Pessoa said.
A broader impact: Making science more trustworthy
Both studies underscore a common theme: While simulating data from known models is often straightforward, inferring the underlying processes from real-world data is far more difficult and more prone to error.
By developing methods that respect the physical and biological realities of how data are generated, the ASU team is helping to close this gap. Their work enables scientists to extract more accurate insights without discarding the complexity that makes real systems meaningful.
Ultimately, these advances not only deepen scientific understanding but also strengthen the reliability of scientific conclusions — an essential step in building trust in data-driven discovery.
More Science and technology
What the origins of stone sculptures can tell us about an ancient Andean civilization
By studying ancient statues in South America, a group of researchers from Arizona State University have gleaned further insights into the cosmological and political power of an ancient civilization.…
How smarter water management can reduce arsenic in rice
Rice is a staple food for more than half of the world's population. But unlike other major grains, it naturally absorbs arsenic — a known carcinogen — because of the unique way it is grown.For years…
AI companionship poses risks for teen development, study shows
As teenagers increasingly turn to artificial intelligence chatbots for advice about friendships, family conflicts and romantic ties, researchers are raising concerns that the technology could disturb…