A new study from the University of Colorado Anschutz Medical Campus combines advanced microscopy and machine learning to reveal how the complete loss of a single gene changes what cells look like in people with Neurofibromatosis type 1 (NF1).
NF1 is a genetic disorder that affects 1 in every 3,000 people. Mutations in the NF1 gene disrupt the production of neurofibromin, a protein that helps control cell growth, especially within Schwann cells which support the nervous system. Without neurofibromin, these cells can grow uncontrollably, leading to tumors.
Nearly all individuals with NF1 develop benign skin tumors that affect appearance and quality of life. About 30–50% also develop plexiform neurofibromas (pNFs), painful and disfiguring tumors that grow along nerves. Even after surgery, half of these tumors return within 10 years. In 8–15% of patients, pNFs can become malignant peripheral nerve sheath tumors (MPNSTs), a rare and aggressive form of cancer with low five-year survival rates.
Currently, only two FDA-approved drugs exist to treat these tumors, and they don’t work for many patients or for the MPNST form of these tumors. That’s why CU researchers are digging deeper into what changes inside Schwann cells when neurofibromin is missing, with the goal of developing more targeted and preventive treatments.
Jenna Tomkinson, a researcher in the Way Lab at the University of Colorado School of Medicine, led a team that included collaborators from iNFixion Bioscience, Massachusetts General Hospital and Harvard Medical School. The researchers used high-resolution microscopy and machine learning to compare healthy Schwann cells with those missing the NF1 gene. They began with a technique called Cell Painting, which uses fluorescence stains to highlight key parts of the cell, like the nucleus, mitochondria, endoplasmic reticulum and the cell’s skeleton (F-actin). This allowed them to visualize and measure subtle structural differences.
“This was a full effort across industry and academia!” said Tomkinson. “Our collaborators at Infixion Bioscience collected the Cell Painting images that our team in the Way Lab utilized to extract features and develop the machine learning model for.”
Before analyzing the images, the team processed the data using a tool called CellProfiler, which removed blurry or poorly lit images. Then, they used it again to outline key parts of each cell and extract over 2,300 measurements, including size, shape, texture, and granularity.
To manage this massive dataset, they used custom-built tools developed within the Way Lab with the support of Dave Bunten from the software engineering team:
Because NF1-mutant cells tend to divide more rapidly, the team also removed dividing cells to focus on meaningful structural differences.
In the end, they had a high-quality dataset of 20,680 cells:
The team trained a binary logistic regression model to detect differences caused by the NF1 mutation using the selected morphology features. Specifically, it detected differences between Schwann Cells with no NF1 mutation (wildtype) and Schwann Cells with complete loss of NF1 (null). The model performed well:
This showed the model could reliably spot subtle morphological patterns linked to NF1.
But when tested on a new set of Schwann cells made using different lab methods, the model’s accuracy dropped to 50%, no better than random guessing. The issue? The new cells had subtle differences caused by how they were made, not by NF1 itself. These non-biological factors, like cloning techniques, different cell sources, or CRISPR editing, confused the model.
“When using immortalized cell lines, any differences in protocol can influence the morphology,” Tomkinson added. “But, by training our model with many different conditions, we can better detect the underlying biological signal.”
The team retrained the model using both sets of cells. Performance improved, showing that the model could learn to ignore irrelevant differences when given a broader range of examples.
Why It Matters
This study shows that combining detailed cell imaging with machine learning is a powerful way to study NF1. It’s sensitive enough to detect subtle changes in cell structure, but it also needs diverse data to stay accurate.
"Morphology is a powerful tool that can be applied to any disease model,” explained Tomkinson. “It is something that can be supplemental to the more standard wet lab practices to get results that weren’t attainable before. I hope to see more scientists apply this method to their projects and see what can be discovered!”
The next steps from this work will include collecting data from more Schwann Cells from different patients, which will minimize the lab-based issues the model struggled with.
“Now that we have validated the assay in NF1, we hope to use an approach called high-throughput drug screening to identify potential therapeutic agents that precisely reverse the disease phenotype and make NF1-deficient Schwann cells look healthy,” added Gregory Way, PhD, MS, assistant professor of biomedical informatics, CU School of Medicine.
These findings lay important groundwork for better diagnostics, more reliable lab models, and future drug discovery for NF1.