The power of image-based profiling lies in its scalability and its agnosticism. Rather than measuring a predefined molecular target, profiles capture the phenotypic consequences of perturbations, allowing researchers to group compounds by mechanism of action, detect off-target effects, and identify subtle morphological changes that molecular assays might overlook. This has made the approach central to phenotypic drug discovery, functional genomics, and disease modeling. In functional genomics, image-based profiles quantify the phenotypic impact of CRISPR knockouts or RNAi perturbations, helping map gene function and pathway relationships at single-cell resolution. In disease modeling, morphological features derived from patient cells can distinguish disease subtypes, identify biomarkers, and screen for phenotype-correcting compounds. The landmark 2017 review by Caicedo and colleagues established the foundational standards for the field, outlining critical steps such as illumination correction, cell segmentation, feature quantification, feature processing, and profile generation. Since then, the field has advanced rapidly, driven by progress in experimental protocols, complementary profiling modalities, and rapidly advancing computational tools.

The experimental side of the workflow begins with careful design. Negative controls such as untreated or vehicle-treated samples establish baselines for normalization, while biological replicates capture natural variability and improve statistical power. Plate layout constrains replicate number; in a standard 96-well plate using the inner 60 wells, six replicates per treatment with flanking positive controls is common. The authors recommend including as many replicates as feasible, typically a minimum of four for treatments and at least nine for controls. Maintaining consistent seeding density and media conditions is equally critical, as inconsistent conditions can introduce artifacts like overcrowding or edge effects. Higher-density formats such as 1536-well plates minimize data loss from discarding edge wells affected by spatial biases. The Cell Painting assay, which stains multiple subcellular compartments across five to six channels, has become the predominant fluorescence-based approach, providing a rich and unbiased representation of cellular morphology. Label-free techniques including brightfield and phase-contrast imaging are increasingly explored for their non-invasive nature and compatibility with live-cell experiments.

Once images are acquired, the computational core of the pipeline takes over. Whole-image quality control removes out-of-focus images or those containing large smudges and debris that would decrease segmentation quality and introduce noise. Illumination correction follows, learning and correcting patterns of uneven background lighting and signal variations from optical artifacts so that downstream measurements reflect biological differences rather than technical noise. Cell segmentation, the delineation of individual cells from the background, has witnessed a decisive shift toward deep learning. Classical approaches implemented in tools like CellProfiler and Ilastik use intensity thresholding and watershed algorithms, but foundational architectures like U-Net catalyzed a transition to convolutional neural networks that achieve superior segmentation accuracy. This paved the way for tools such as StarDist and Cellpose that offer models capable of identifying cell borders with minimal parameter tuning. More recently, transformer-based foundation models like CellSAM and CellViT leverage these architectures to further improve performance and generalizability across diverse imaging conditions.

Feature extraction represents the most consequential transformation in the field. Classical workflows extract hundreds to thousands of hand-crafted features from segmented objects, quantifying aspects like size, shape, texture, intensity, and spatial relationships. These features provide a mathematically well-defined representation of cell morphology, and their transparency supports hypothesis-driven research. However, the feature set might miss particular phenotypic changes and can be prone to redundancy and noise. Deep learning methods learn relevant morphological representations directly from raw pixel data, producing embeddings that are typically less redundant and frequently outperform classical features in downstream tasks. The review highlights striking examples of both directions: the count of cells or mitochondria outperformed convolutional neural networks to predict assay outcomes, and mechanism of action prediction using K-nearest neighbors on CellProfiler features achieved an F1 score of 0.41, comparable to DeepProfiler’s 0.42. Yet self-supervised DINO frameworks increased mean average precision by 29 percent when classifying chemical perturbations compared to hand-crafted features, and zero-shot image-to-image mechanism of action classification using contrastive learning achieved 61.3 percent in top-10 accuracy compared to 24.5 percent with CellProfiler.

Two major neural network architectural families dominate deep learning-based feature extraction. Convolutional neural networks process images through stacked convolutional filters that perform localized transformations to extract features hierarchically, learning increasingly descriptive representations while maintaining relatively low computational demands. Vision transformers, adapted from natural language processing, partition an input image into non-overlapping patches that are flattened and linearly embedded, then processed in parallel through self-attention blocks that capture global relationships among all image patches at once. The scaling properties of transformers have made the architecture especially useful with large datasets. Training strategies have evolved in parallel. Supervised learning requires ground truth labels that are frequently unavailable in profiling experiments. Weakly supervised learning uses easily obtained annotations like perturbation labels, which are biologically meaningful but do not directly reveal phenotypic effects, as auxiliary tasks. Self-supervised learning relies entirely on unannotated image data, defining objective functions derived from the data itself, such as reconstructing masked patches or distinguishing between augmented views of the same image.

Post-processing deep learning features requires careful attention to order of operations. Applying a sphering transform to deep learning features followed by robust z-scoring against plate-level statistics yields the most effective post-processing strategy for self-supervised features, and performing z-scoring before sphering was found to substantially degrade performance in most cases. Feature selection, which is essential for classical features due to their high correlation, might be unnecessary for deep learning embeddings since only weak correlations are generally observed. Batch effect correction remains a persistent challenge regardless of feature type. A recent study evaluated ten batch correction techniques from single-cell RNA sequencing on image-based profiling data and concluded that none could adequately remove batch effects in the most complex cases involving multiple microscope types and a large number of compounds, underscoring the ongoing need for new methods.

The field is now moving toward foundation models, large models trained on massive and diverse datasets that can be adapted to solve a wide range of tasks. Early work repurposed natural-image convolutional networks by pseudo-RGB channel mapping, while more recent efforts have trained self-supervised vision transformers on millions of microscopy images. One masked autoencoder-based strategy scaled vision transformer models to approximately 1.9 billion parameters and datasets with roughly 16 million images, demonstrating that performance improves at that scale. Channel adaptation remains a significant hurdle, as most models have been trained using the five-channel Cell Painting assay while imaging panels vary with experimental goals. Several strategies have emerged, including models that process one channel at a time, channel-agnostic masked autoencoders that simultaneously process all channels adaptively, and novel transformer architectures incorporating channel-aware attention mechanisms.

Emerging frontiers are expanding the dimensions of profiling beyond static two-dimensional snapshots. Temporal imaging captures dynamic cell processes that static methods cannot, circumventing the survivorship bias of fixed time point profiles. Live-cell painting using minimally toxic dyes enables time-lapse imaging, and advances in spinning disk confocal and light sheet microscopy make acquisition of large temporal datasets possible. The third spatial dimension remains underexplored but holds promise for revealing more nuanced cellular signatures, particularly for complex cell shapes like neurons or macrophages with extended protrusions that may appear disconnected in a two-dimensional slice. Virtual staining, which predicts fluorescence signals from label-free modalities such as brightfield imaging using machine learning, offers lower cost and reduced phototoxicity. However, virtual staining has not yet been used in any published studies for image-based profiling applications, and its generalizability across unseen cell types and phenotypes remains highly variable and dependent on the biological context of training data.

Single-cell image-based profiling offers an effective, affordable, high-content alternative to single-cell RNA sequencing. While population-level aggregation creates robust representations that minimize technical noise, averaging leads to loss of information about cell heterogeneity and the diversity of cell states. Methods that capture heterogeneity, subpopulations, and cell state transitions have been applied to identify drug toxicity, cell types, cell cycle stages, and cell health alterations in rare subpopulations. Emerging evidence suggests that single-cell analysis may improve mechanism of action prediction compared to population-level analysis. Optical pooled screening has emerged as a transformative technology, enabling high-throughput single-cell resolution mapping of genotype-to-phenotype relationships at unprecedented scale by integrating high-content imaging with in situ genotyping methods. The review identifies six core challenges for the next decade: developing end-to-end workflows, fortifying methodological infrastructure for temporal, three-dimensional, and virtual staining modalities, bridging profiles to mechanisms through interpretable features, establishing rigorous quality control standards, building systematic benchmarking frameworks, and reaping the rewards of artificial intelligence and foundation models while accounting for their high computational costs. The authors emphasize that these challenges form a cohesive foundation for the next generation of image-based profiling, and that realizing the field’s promise depends on collective resolve to confront and overcome the obstacles that remain.

Subject of Research: Image-based profiling methods and computational challenges in high-content microscopy

Article Title: Progress and new challenges in image-based profiling

Article References: Serrano, E., Peters, J., Wagner, J., Graham, R. E., Chen, Z., Feng, B. Y., Miranda, G., Kalinin, A. A., Vulliard, L., Tomkinson, J., Mattson, C., Lippincott, M. J., Kang, Z., Sitani, D., Bunten, D., Seal, S., Carragher, N. O., Carpenter, A. E., Singh, S., … Way, G. P. (2026). Progress and new challenges in image-based profiling. Molecular Systems Biology, 22(5), 624-658. https://doi.org/10.1038/s44320-026-00197-7

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00197-7

Keywords: image-based profiling, Cell Painting, deep learning, high-content screening, cell segmentation, foundation models, Progress, challenges, image-based, profiling, scientific research

The two technologies at the heart of the study capture fundamentally different views of drug activity. The first, thermal proteome profiling, or TPP, grew out of the cellular thermal shift assay introduced in 2013. The principle is elegant: when a small molecule binds a protein, it often shifts the protein’s melting temperature, either stabilizing or destabilizing it. Researchers heat intact cells to a range of temperatures before lysis, then use mass spectrometry to measure how much of each protein remains soluble at each temperature. Coupling this to whole-proteome analysis, first demonstrated by Savitski and colleagues in 2014, allows an unbiased survey of thousands of drug-protein interactions in a single experiment. A typical whole-cell TPP experiment yields tens to hundreds of proteins with perturbed thermal stability, but these include not only direct drug targets but also proteins in complexes with, or downstream of, the true target, making it difficult to separate cause from consequence.

The second technology, Cell Painting, works from the opposite direction. Instead of measuring protein behavior directly, it captures the morphological fingerprint a drug imposes on a cell. Cells are treated with compounds and then stained with a fixed panel of fluorescent dyes labeling the nucleus, endoplasmic reticulum, nucleoli and cytoplasmic RNA, Golgi apparatus and actin cytoskeleton, and mitochondria. Software such as CellProfiler then extracts thousands of features per cell, producing a rich quantitative description of each treatment’s morphological perturbation. Because the assay is inexpensive and highly multiplexable, it can be applied to thousands of compounds at once. Compounds with similar mechanisms of action tend to induce similar morphological changes, so machine learning can infer the mechanism of an unknown compound by comparing it to well-annotated neighbors. The weakness is that small molecules often have many targets, and the accuracy of predictions depends heavily on the quality of available annotations.

The Uppsala team’s insight was that these two data types are complementary in a very specific way. TPP detects target engagement directly but misses proteins, sometimes because a peptide fails to ionize in the mass spectrometer, sometimes because a protein is more thermally stable than the tested temperature range, and sometimes because membrane-associated proteins sediment during centrifugation. Cell Painting, meanwhile, can suggest targets through compounds that produce similar shapes but cannot confirm physical binding. The researchers therefore built a pipeline that starts with the list of thermally stabilized or destabilized proteins from a TPP experiment and constructs a protein-protein interaction network using the STRING database, which aggregates experimentally scored interactions, curated pathway databases, and text mining. In parallel, the compound of interest is located within a Cell Painting dataset of 5259 compounds from the SPECS drug repurposing repository, profiled in U2OS bone cancer cells, and the known targets of its closest morphological neighbors are used to fill gaps in the network.

A key technical challenge was clustering the Cell Painting data reliably. Popular algorithms such as k-means and HDBSCAN are stochastic and can produce very different cluster structures on repeated runs, particularly on high-dimensional, poorly separated data like morphological profiles. The team turned to SC3s, a consensus clustering method originally developed for single-cell RNA sequencing, which combines principal component analysis with thousands of repeated k-means runs and summarizes the results into a consensus matrix. Running the algorithm 2000 times across five different cluster numbers, between 100 and 180, produced highly reproducible clusters for most test compounds. The cluster-derived target candidates were then merged with the TPP network, and betweenness centrality scores, a graph-theoretic measure of how often a node lies on shortest paths between others, were used to filter the merged network. Because drug targets have been shown to occupy central positions in interaction networks, the researchers reasoned that true targets would rank among the highest-scoring nodes.

To validate the approach, the team used publicly available TPP datasets for five well-characterized drugs: the BET bromodomain inhibitors (+)-JQ1 and I-BET151, the BRAF inhibitor vemurafenib, the ALK and MET inhibitor crizotinib, and the histone deacetylase inhibitor panobinostat. For (+)-JQ1, 67 proteins showed dose-dependent thermal shifts, and the resulting network correctly placed the known targets BRD4, BRD3, and BRD2, along with the documented off-target HADHA, among the top-ranked nodes. For I-BET151, 40 perturbed proteins yielded a network containing BRD4, BRD3, and BRD2. Intriguingly, although the two compounds share targets, only 11 perturbed proteins overlapped between them, suggesting they regulate different pathways downstream of target binding. Gene Ontology and Reactome enrichment analysis of the network communities recovered mechanisms consistent with the known biology of (+)-JQ, including lysine-acetylated histone binding and activation of cell death.

The validation extended further. For panobinostat, TPP alone detected only three of eleven known targets, but Cell Painting cluster analysis contributed eight more, and the combined network placed six targets among the top-ranked proteins. For vemurafenib, the primary target BRAF was entirely absent from the TPP data, while the off-target FECH was missing from the Cell Painting cluster, so only the integrated network captured both. The method correctly identified known targets and mechanisms for four of the five compounds. The exception, crizotinib, proved instructive rather than disappointing: its true targets ALK, ROS1, and MET are transmembrane proteins that classical TPP protocols rarely detect, and neither U2OS nor K-562 cells expressed ALK or ROS1. Instead, the analysis highlighted the Bcr-Abl pathway through the off-targets ABL1 and BCR, a biologically plausible finding given the high expression of these proteins in the K-562 leukemia cells used for TPP. The team then scaled the validation to a public Proteome Integral Solubility Alteration, or PISA, dataset covering 49 compounds, a high-throughput TPP variant in which all temperature treatments are pooled before mass spectrometry. The combined model achieved target prediction with a ROC AUC of at least 0.7 for 21 compounds, compared with 15 for PISA alone and 14 for Cell Painting alone, and outperformed the individual methods in the majority of cases.

As a proof of principle, the researchers applied the pipeline to sinomenine, a plant-derived alkaloid used in China to treat rheumatoid arthritis and pain, whose mechanism of action has remained largely obscure. They generated their own PISA data in U2OS cells treated with 10 and 30 micromolar sinomenine, finding 175 and 121 significant thermal shifts respectively, with 94 proteins affected at both doses. The integrated network revealed a strikingly multimodal profile. Among the highest centrality scores were the beta-2 adrenergic receptor ADRB2 and the NMDA receptor subunit GRIN2C, both endogenously expressed in U2OS cells, supporting an action on nervous system receptors consistent with the compound’s analgesic use and with previous mouse studies showing elevated GABA signaling. Enrichment analysis also flagged neurotransmitter receptor activity and chemical synaptic transmission. Equally intriguing was the appearance of cyclin-dependent kinase 2, CDK2, alongside pathways governing RNA binding, DNA replication, chromosome organization, and double-strand break repair, pointing to potential anti-tumor effects that corroborate earlier sinomenine studies. Additional communities implicated actin cytoskeleton organization, membrane trafficking, RHO GTPase signaling, and metabolic processes including the citric acid cycle, with the structural protein ACTB and the mitochondrial transporter SFXN1 scoring highly.

The authors are careful to note the method’s limits. It cannot detect a target unless that target shows thermal shifts or is annotated for morphologically similar compounds, and the workflow deliberately trusts TPP-derived nodes over Cell Painting-derived ones, which meant that in 7 of 15 cases a target seen only in imaging data failed to enter the final network. Cell line choice matters enormously, as the crizotinib case demonstrated, and the team recommends using the same cell line for both assays when investigating a compound with unknown targets. Clustering reproducibility also faltered for vemurafenib, the compound with the weakest morphological signal. Even so, the payoff is substantial: a typical mass spectrometry run covers only about half the proteome, roughly 10,000 of 20,000 canonical proteins, and TPP alone can leave researchers with 35 to 64 candidate targets to chase. The integrated workflow narrows that list to ten or fewer, promising to slash the time and cost of follow-up validation experiments. With the full pipeline written in Python and released on GitHub, the Uppsala team has handed chemical biologists a practical tool for turning two imperfect lenses on drug action into one considerably sharper picture, one that could accelerate both drug development and drug repurposing for compounds whose secrets have resisted decryption for decades.

Subject of Research: Integration of Cell Painting morphological profiling and thermal proteome profiling for drug target identification and mechanism of action inference

Article Title: Cell painting and thermal proteome profiling for inference of drug targets and mechanism of action

Article References: Johansson, C., Johansson, M., Kasi, P. B., Larsson, M., Jakobsson, P.-J., Göransson, U., Carreras Puigvert, J., Spjuth, O., & Jansson, E. T. (2026). Cell painting and thermal proteome profiling for inference of drug targets and mechanism of action. Molecular Systems Biology, 22(8), 1270-1291. https://doi.org/10.1038/s44320-026-00214-9

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00214-9

Keywords: Cell Painting, thermal proteome profiling, drug discovery, mechanism of action, target deconvolution, protein-protein interaction networks, proteomics, sinomenine, PISA, mass spectrometry, chemical biology, Uppsala University