New AI Framework Reads Gene Maps Like Images to Reveal Hidden Cell States

Every cell in the body carries the same genome, yet cells in the heart, brain, and immune system behave in radically different ways because they read different genes at different times. For years, biologists have relied on single-cell transcriptomics to catalog these differences, tallying up how many messenger RNA molecules for each gene sit inside a segmented cell and using those counts to assign the cell to a type or state. But a growing body of evidence suggests that the story does not end with the tally. Where a transcript sits inside a cell—clustered near the nucleus, scattered through the cytoplasm, or gathered at a leading edge—can carry as much biological meaning as how many of them there are. A new computational framework called STARIT, short for Spatial Transcriptomics As Rasterized Image Tensors, now promises to capture that hidden layer of information, and it does so by treating the inside of every cell as an image that a neural network can learn to read.

The framework, described in a study published in PLOS Computational Biology by Dee Velazquez, Caleb Hallinan, Roujin An, Kalen Clifton, and Jean Fan, addresses a fundamental blind spot in imaging-based spatially resolved transcriptomics, a family of technologies that measure gene expression with molecular resolution while preserving the physical layout of tissue. These platforms can record the precise coordinates of thousands of individual mRNA molecules across millions of cells, producing datasets of extraordinary richness. Yet the standard analysis pipeline throws most of that richness away. By collapsing the molecular map into a gene-by-cell count matrix, conventional methods reduce each cell to a bag of genes, discarding the spatial texture that the imaging platforms worked so hard to capture.

STARIT’s central insight is deceptively simple: instead of converting transcripts into counts, convert them into pixels. The method takes all of the mRNA molecules detected within the boundary of a segmented cell and rasterizes them into a standardized, image-based tensor representation. Each gene, or set of genes, becomes a channel of the tensor, and the spatial distribution of transcripts within the cell becomes the pixel values of that channel. The result is a stack of images—one per gene or gene set—that encodes not just what a cell is expressing but where, inside the cell, each transcript resides. Because the representation is a tensor of fixed structure, it plugs directly into the deep learning computer vision models that have transformed fields from medical imaging to autonomous driving.

This design choice matters because it lets researchers borrow the full arsenal of modern computer vision. Convolutional neural networks, which excel at detecting spatial patterns in images, can now be trained to recognize spatial patterns in gene expression. A network trained on STARIT tensors can learn, for example, that transcripts of a particular gene tend to hug the nuclear envelope in one cell state and disperse toward the cell membrane in another. Such patterns are invisible to count-based methods by construction, since both configurations would produce identical tallies. By preserving geometry, STARIT opens a dimension of cellular characterization that has been technically accessible but computationally neglected.

The authors put their framework through a rigorous evaluation using both simulated and real imaging-based spatial transcriptomics data. In simulations, where the ground truth is known by design, STARIT could distinguish transcriptionally distinct cell types with high fidelity, matching the performance expected of conventional approaches. The real test, however, was whether it could go further. Across real datasets, the method not only separated known cell types but also distinguished cell states defined purely by subcellular transcript localization—states that conventional gene count analysis failed to capture. In other words, two cells with the same gene expression profile could be told apart if their transcripts were arranged differently inside them, a capability that no bag-of-genes representation can offer.

The implications reach into some of the most active areas of biology and medicine. Cell state, rather than cell type, often determines how a tissue responds to disease. Tumor cells shift into invasive, drug-tolerant, or proliferative states; immune cells flip between activated and exhausted configurations; neurons remodel their local molecular architecture in response to activity. If the spatial arrangement of transcripts within cells tracks these transitions, then a method that reads that arrangement could detect disease-relevant states earlier and more reliably than count-based classifiers. Subcellular localization is also intimately tied to biology itself: mRNA molecules are actively transported to specific cellular neighborhoods, where local translation supports functions like cell migration, polarity, and synaptic signaling. A framework that quantifies localization at scale turns a qualitative observation into a measurable, learnable feature.

Part of STARIT’s significance lies in its standardization. Imaging-based spatial transcriptomics has suffered from a proliferation of bespoke analysis approaches, each lab rasterizing, encoding, and modeling its data differently, which makes results hard to compare and reproduce. By defining a common tensor format for subcellular molecular information, STARIT gives the field a shared substrate on which models can be built, benchmarked, and transferred. A model trained on one tissue or platform could, in principle, be fine-tuned on another, much as pre-trained vision models are adapted across image domains today. That kind of interoperability could accelerate a shift from one-off analyses to cumulative, community-wide model development.

The framework also highlights a broader conceptual shift underway in computational biology: the migration of biological data into representations suited to deep learning. Just as protein structures became representable as geometric graphs amenable to neural networks, and histology slides became pixel grids for diagnostic AI, the interior of the cell is now becoming an image. Each of these transitions unlocked capabilities that the previous data formats could not support. STARIT suggests that the transcriptome’s spatial texture is the next frontier, and that the tools of computer vision—already battle-tested on natural images—may be repurposed with relatively little modification to mine it.

Challenges remain, as they do for any young method. Rasterizing transcripts into images requires choices about resolution, channel organization, and normalization, and the authors’ demonstration that the approach works on simulated and real data is a first step rather than a final verdict. Scaling to tissues with millions of cells and tens of thousands of genes will demand careful engineering, and interpreting what a vision network has learned about subcellular biology will require new explainability tools. Still, the direction is clear. By refusing to average away the molecular geography of the cell, STARIT gives biologists a way to see cell states that were always there but never before systematically measurable.

For a field that has spent two decades perfecting the art of counting molecules, the message of this study is that counting was only ever half the picture. The other half is written in the positions of those molecules, in patterns of nuclear clustering and cytoplasmic dispersal that encode how a cell is organized and, quite possibly, what it is becoming. With STARIT, that half of the picture is finally rendered in a form machines can learn from—and the cells hiding in plain sight within our tissues may soon give up secrets that gene counts alone could never reveal.

Subject of Research: A computational framework that represents subcellular transcript localization in imaging-based spatial transcriptomics as image tensors for deep learning-based cell state characterization

Article Title: Spatial Transcriptomics As Rasterized Image Tensors (STARIT) characterizes cell states with subcellular molecular heterogeneity

Article References: Velazquez, D., Hallinan, C., An, R., Clifton, K., & Fan, J. (2026). Spatial Transcriptomics As Rasterized Image Tensors (STARIT) characterizes cell states with subcellular molecular heterogeneity. PLOS Computational Biology, 22(10), e1014845. https://doi.org/10.1371/journal.pcbi.1014845

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014845

Keywords: spatial transcriptomics, STARIT, subcellular localization, deep learning, computer vision, cell states, single-cell analysis, computational biology, gene expression, image tensors, PLOS Computational Biology, mRNA