Microscopes have been making cells famous since the 1600s, but nobody told the cells they'd eventually be measured in over 1,500 different ways simultaneously - and judged by artificial intelligence.
Image-based profiling is exactly what it sounds like: taking pictures of cells and extracting an absurd number of measurements from them. Size, shape, texture, how their organelles are arranged, whether their mitochondria look stressed out - the works. A new review in Molecular Systems Biology by Serrano, Peters, Wagner and colleagues surveys how this field has evolved from "look at this neat cell" to "let's profile billions of cells and let deep learning sort them out."
The Cell Painting Glow-Up
The workhorse of this field is something called Cell Painting, an assay that stains cells with six fluorescent dyes highlighting eight different cellular components. Picture it as giving cells a full makeover with glow-in-the-dark makeup for their nucleus, mitochondria, endoplasmic reticulum, and more. Software then extracts roughly 1,500 features per cell, creating what researchers call a "morphological profile."
These profiles turn out to be surprisingly useful. Drug companies can screen compounds and predict mechanisms of action based on how cells look after treatment - no expensive biochemical assays required. The JUMP-Cell Painting Consortium, a collaboration between pharmaceutical giants, has released profiles from over 136,000 chemical and genetic perturbations. That's 116,750 compounds and thousands of gene knockouts, all captured in around 700 terabytes of data. Your phone's storage is crying.
Deep Learning Enters the Chat
For years, the gold standard was CellProfiler, open-source software that extracts hand-crafted features from images. It's been used by thousands of biologists and works well, but there's a catch: it's computationally intensive and requires careful parameter tuning.
Enter deep learning. Neural networks can now learn features directly from images without anyone defining what "interesting" means beforehand. Self-supervised methods like DINO have begun outperforming CellProfiler on tasks like drug target classification while running 50 times faster. Attention-based architectures like X-Profiler combine convolutional networks with Transformers to filter out noise and capture subtle phenotypes that traditional pipelines miss.
The trade-off? Interpretability. When CellProfiler tells you a cell's nucleus is 15% larger, you know what that means. When a neural network says "embedding dimension 47 increased by 0.3," good luck explaining that to a biologist.
The Batch Effect Beast
Here's the dirty secret of high-throughput microscopy: cells photographed on Monday look different from cells photographed on Friday, even if they're genetically identical. Different microscopes, lamp intensities, staining concentrations, and room temperatures all introduce "batch effects" - systematic variations that have nothing to do with biology.
A 2024 Nature Communications study benchmarked ten batch correction methods borrowed from single-cell RNA sequencing. The winners? Harmony and Seurat RPCA performed best across scenarios ranging from same-lab-different-days to different-microscopes-different-continents. Still, no method works perfectly, and overcorrection can erase genuine biological signals along with the noise. It's less "cleaning up data" and more "performing surgery while blindfolded."
New Frontiers: Pooled Screens, Time, and the Third Dimension
The review highlights several emerging techniques pushing the field forward. Optical pooled screening combines CRISPR perturbations with in situ sequencing, letting researchers link phenotypes to genetic knockouts at single-cell resolution. Instead of one gene per well, you can screen entire genome-wide libraries in a single dish.
Temporal imaging adds movies to the mix - tracking how cells change over hours or days rather than capturing single snapshots. This is where tools like mapb2.io for visualizing complex relationships could help researchers make sense of dynamic cellular behaviors.
And then there's 3D. Organoids - miniature lab-grown organs - don't sit flat on a dish. They're spheres, and imaging them means capturing dozens of z-stacks and somehow segmenting individual cells from a blob of tissue. Current tools struggle with the computational load and variable image quality. Deep learning solutions exist but often require high-resolution confocal microscopy that isn't practical for high-throughput screens.
Quality Control: Still a Work in Progress
The authors note that robust quality control standards remain elusive. How do you know if your profiles are "good enough"? There's no universal benchmark, and what works for drug screening might fail for genetic perturbation studies. The field needs standardized metrics and workflows - preferably ones that don't require a PhD to implement.
Open-source tools and public datasets are helping. The Cell Painting Gallery hosts hundreds of terabytes of freely available data, and benchmarks like CPJUMP1 provide ground truth for method development. Reproducibility is improving, but there's a long road ahead.
The Bottom Line
Image-based profiling has matured from a niche technique into a pillar of drug discovery and functional genomics. Deep learning is accelerating everything, but challenges in 3D imaging, temporal data, batch effects, and quality control remain unsolved. The authors - many from the Carpenter-Singh Lab at the Broad Institute, where CellProfiler and Cell Painting were born - aren't just cataloging progress. They're mapping the terrain for whoever wants to tackle what's next.
Cells have never been more scrutinized. Now we just need to figure out what all those measurements actually mean.
References:
-
Serrano E, Peters J, Wagner J, et al. Progress and new challenges in image-based profiling. Mol Syst Biol. 2026. DOI: 10.1038/s44320-026-00197-7
-
Bray MA, Singh S, et al. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nat Protoc. 2016;11(9):1757-1774. DOI: 10.1038/nprot.2016.105
-
Chandrasekaran SN, et al. Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations. Nat Methods. 2024;21:1114-1121. DOI: 10.1038/s41592-024-02241-6
-
Arevalo J, et al. Evaluating batch correction methods for image-based cell profiling. Nat Commun. 2024;15:6516. DOI: 10.1038/s41467-024-50613-5
-
Funk L, et al. Optical pooled screens in human cells. Cell. 2019;178(6):1424-1435. PMID: 31626775
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.