AI and ‘Ramanomics’ could eliminate a major obstacle to studying living cells

BUFFALO, N.Y. – Fluorescent dyes have long been used in biological research to identify and visualize structures within living cells. Although effective, they have several drawbacks, including altering the cells under study, limiting the number of structures that can be examined at once and reducing measurement accuracy.

A team led by University at Buffalo researchers has developed a new method that draws on advances in artificial intelligence and Raman spectroscopy to overcome the limitations of dye‑based imaging.

The approach combines AI with “Ramanomics,” a UB-pioneered optical technology that measures the biochemical makeup of cells without altering them. Rather than relying on fluorescent labels, which are dyes that bind to specific cellular components and glow under specialized lighting, it identifies cellular structures by their unique biochemical signatures.

“This cross‑discipline collaboration applying AI to Ramanomics revealed for the first time that organelles in living cells can be identified noninvasively without dye labeling,” said Paras N. Prasad, PhD, SUNY Distinguished Professor and executive director of UB’s Institute for Lasers, Photonics and Biophotonics. “This opens new possibilities for identifying biological markers of disease, discovering new medicines and advancing more personalized approaches to medicine.”

By removing the need for fluorescent dyes, the method gives researchers a clearer view of how cells function, how disease affects them and how they respond to treatment.

“When you add dye to a cell, you’re introducing a foreign element that can disrupt its natural behavior and influence the very measurements you’re trying to make,” said Varun Chandola, PhD, associate professor in the Department of Computer Science and Engineering at UB. “Having the ability to see inside a cell without physically altering it is a major step forward. It allows us to capture biology as it naturally occurs, providing a more accurate picture of how cells really function.”

Prasad and Chandola are both lead research investigators for a study, published in June in ACS Omega, a peer‑reviewed scientific journal of the American Chemical Society, describing the technology.

As part of their research, the team used Raman spectroscopy to capture the distinct biochemical “fingerprints” of four cellular organelles. Those fingerprints were then used to train several machine learning models to recognize each organelle. Once trained, the AI was able to analyze new Raman measurements and determine where in the cell they were collected without relying on fluorescent labels.

Of the machine learning models tested, a neural network performed best, correctly identifying the organelles with about 90% accuracy.

“We were highly encouraged by the test results,” Chandola said. “Once the model receives the measurement data, it does the rest. It can recognize cellular structures from their biochemical signatures, making it possible to examine cells in their natural state.”

The approach also offers practical advantages. It simplifies the workflow by eliminating the need to prepare and apply fluorescent dyes, reduces experimental artifacts and improves measurement sensitivity, helping researchers collect cleaner, more reliable data.

Studying living cells without disrupting them could give researchers new insights into how diseases emerge and evolve. It may make it easier to observe the effects of disease on cells, monitor treatment responses and identify molecular changes linked to serious conditions such as cancer and metabolic disorders.

“If you're trying to understand how a disease progresses or how a drug interacts with a cell, you need to measure those changes over time without damaging the cell,” Chandola said. “Our approach gives researchers that ability while preserving the biological information they’re trying to capture.”

The team is now working to pair its AI models with emerging Raman imaging systems that use quantum light sources. They believe these technologies can dramatically accelerate data collection and increase resolution, creating the larger datasets needed to train stronger machine‑learning models.

“Today, you often have to wait up to three minutes to collect a reliable signal,” Chandola said. “If we can collect that information in about a second instead of several minutes, we'll be able to analyze far more cells at a much larger scale. We're also expanding this approach to explore other aspects of cell biology, including using the model's output to distinguish cancerous cells from healthy ones.”

In addition to Chandola and Prasad, the study’s co‑authors include Andrey N. Kuzmin, PhD, and Alexander Baev, PhD, both of UB’s Department of Chemistry and Institute for Lasers, Photonics and Biophotonics; Artem Pliss, PhD, of the School of Pharmacy at D’Youville University; and Giovana R. Teixeira, PhD, of São Paulo State University (UNESP) in Brazil.

The study was supported by a seed grant from the UB Office of the Vice President for Research, Innovation and Economic Development, for which the researchers expressed their gratitude.