Cell Painting image encoder that turns whole-slide multi-channel microscopy into morphological profiles in one pass, with no cell segmentation step.
No providers recorded yet. Browse all providers
PhenoProfiler is an end-to-end image encoder for Cell Painting microscopy: a whole-slide multi-channel image goes in, and a low-dimensional morphological profile comes out in a single forward pass. It was developed by researchers at the University of Macau and the University of Florida with collaborators across China and the United States, posted as a preprint in February 2025 and published in Nature Communications that December.
Image-based drug screening depends on turning millions of fluorescent cell images into vectors that can be compared across compounds and genetic perturbations. Established pipelines get there in stages: segment the multi-channel field into individual cells, crop each cell, push the crops through a feature extractor such as CellProfiler or DeepProfiler, then aggregate crop-level features back into a well-level profile. Every stage costs compute, and a poor segmentation quietly corrupts everything downstream. PhenoProfiler removes the staging entirely, and with it segmentation error as a failure mode.
The second departure is what the model is trained to predict. Rather than only classifying which treatment a well received — a label that is coarse and plate-specific — PhenoProfiler regresses onto continuous reference morphology profiles, with classification and contrastive objectives as auxiliary signals. That makes it a different animal from self-supervised Cell Painting encoders such as OpenPhenom-S/16, which learn purely from masked image reconstruction and use no perturbation information at all.
The architecture has three parts. A gradient encoder built on difference convolution enhances cell edges and boundary contrast in the raw multi-channel input; a transformer encoder with multi-head self-attention captures long-range structure across the field; and a pair of MLP heads drive the multi-objective learning module. Regression targets are the median morphology profiles of the cpg0019 collection; the three loss terms are weighted 0.1 (classification), 100 (regression), and 1 (contrastive), chosen by sensitivity analysis. Training used four NVIDIA A100 40GB GPUs with a batch size of 300 over up to 200 epochs, beginning with regression-only optimization for roughly 100 epochs before joint training.
Benchmarking spans over 230,000 whole-slide images from BBBC022, CDRP-BIO-BBBC036, and TAORF-BBBC037 — 231 plates and 4,285 compound and gene-overexpression treatments — evaluated by Folds of Enrichment (FoE) and Mean Average Precision (MAP) on biological matching. Against DeepProfiler, PhenoProfiler improves FoE by 64.9%, 3.4%, and 50.0% and MAP by 24.3%, 21.3%, and 15.3% across the three datasets. In the non-end-to-end setting on 8.42 million single-cell images spanning 450 treatments, it improves FoE by 11.9%, 9.7%, and 30.7% over DeepProfiler. On batch structure, its representations reach an Inverse Median Absolute Deviation of 0.603 on BBBC022, against 0.326 for DeepProfiler and 0.458 after DeepProfiler receives an explicit correction step. Leave-plates-out and leave-dataset-out splits show the advantage holds when training and test plates or datasets are disjoint.
The model targets high-content screening groups in pharma and academia that run Cell Painting assays at plate scale. Typical uses are mechanism-of-action clustering, hit triage by phenotypic similarity to annotated reference compounds, and comparison of gene-overexpression perturbations against compound treatments in a shared feature space. Because a whole plate is profiled without per-cell segmentation, the encoder suits screens where cells are confluent or otherwise hard to segment reliably. The web server lets bench scientists profile a new plate without configuring a deep learning environment.
PhenoProfiler's code is released under the MIT license, with trained checkpoints distributed via Google Drive and Synapse and an inference notebook for applying the fixed checkpoint to new plates. Its main contribution is architectural rather than scale-driven: segmentation-free whole-slide encoding beats crop-based pipelines on standard biological matching benchmarks while absorbing batch correction into the representation itself. The limitations follow from the training design. Regression targets derive from existing profile definitions, so the model inherits the notion of "morphology" encoded in those reference features rather than discovering one, and training is specific to Cell Painting-style assays. The MIT grant covers the code alone — both paper versions are CC BY-NC-ND 4.0 — and the weights sit on general file-sharing services rather than a model hub, which makes versioning harder to track.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.