PIFiA: Self-supervised Approach for Protein Functional Annotation from Single-Cell Imaging Data

Abstract

Fluorescence microscopy data describe protein localization patterns at single-cell resolution and have the potential to reveal whole-proteome functional information with remarkable precision. Yet, extracting biologically meaningful representations from cell micrographs remains a major challenge. Existing approaches often fail to learn robust and noise-invariant features or rely on supervised labels for accurate annotations. We developed PIFiA, (Protein Image-based Functional Annotation), a self-supervised approach for protein functional annotation from single-cell imaging data. We imaged the global yeast ORF-GFP collection and applied PIFiA to generate protein feature profiles from single-cell images of fluorescently tagged proteins. We show that PIFiA outperforms existing approaches for molecular representation learning and describe a range of downstream analysis tasks to explore the information content of the feature profiles. Specifically, we cluster extracted features into a hierarchy of functional organization, study cell population heterogeneity, and develop techniques to distinguish multi-localizing proteins and identify functional modules. Finally, we confirm new PIFiA predictions using a colocalization assay, suggesting previously unappreciated biological roles for several proteins. Paired with a fully interactive website (https://thecellvision.org/pifia/), PIFiA is a resource for the quantitative analysis of protein organization within the cell.

Anomaly detection for high-content image-based phenotypic cell profiling

“…Weakly supervised representation-learning uses experimental labels, such as the treatment label, to "guide" their representations such that replicates of the same label are encoded close to one another in the latent space and different labels are encoded far apart [26][27][28][29][30][31]. Beyond optimizing representations that are hard to biologically-interpret, experiment treatment-based weak supervision may lead to undesired consequences where the representations of treatments that lead to a similar phenotype are pushed away from one another in the latent space because they do not share the same label, possibly even pushing one representation closer to the control (Supplementary Figure 4A).…”

“…Thus, the biological function investigated may be incorrectly interpreted due to a simplified representation of the underlying data complexity. Recent representation-learning methods train machine learning models to encode lower-dimensional cell-level or well-level embeddings (called "latent representations") through weakly supervised learning [26][27][28][29][30][31] or self supervised learning [32]. While representation-learning methods show promising results in terms of capturing the differences between treatments, current representation methods are not optimized to model the change posed by the treatment with respect to the control, but to distinguish between the different treatments (see Discussion).…”

From pixels to insights: Machine learning and deep learning for bioimage analysis

“…[ 26 ] Similarly, a self‐supervised CNN was used on the yeast ORF‐GFP collection to group proteins based on their localization and function, without the need for manual compartment labeling. [ 25 ] Additionally, a self‐supervised method was used to learn feature representations of single cells without labeled training data in both yeast and human microscopy images. Given one cell from a microscopy image, the trained CNN was able to predict the fluorescence pattern in another cell from the same image.…”

“…[24] Predicted outputs of supervised ML/DL models can be overfitted protein interaction modules based on subcellular localization. [25] Unsupervised ML/DL methods are used for pattern detection. An example of unsupervised DL algorithms are autoencoders that capture the most important features of the input data and can be used, for example, to reduce the dimensionality of data, denoise data, or extract DL features.…”

“…Hierarchical clustering was applied to remove redundant assays from datasets when predicting drug candidates based on chemical structures and phenotypic profiles of previous high-throughput screens, and to determine the subcellular localization of the yeast proteome by clustering DL features of cell crops of the ORF-GFP collection images. [25,27] Other types of learning have also been used for bioimage analysis.…”