Learning orientation-invariant representations enables accurate and robust morphologic profiling of cells and organelles
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (Arcadia Science)
Abstract
Cell and organelle morphology are driven by diverse genetic and environmental factors and thus accurate quantification of cellular phenotypes is essential to experimental cell biology. Representation learning methods for phenotypic profiling map images to feature vectors that form an embedding space of morphological variation useful for clustering, dimensionality reduction, outlier detection, and supervised learning problems. Morphology properties do not change with orientation, and thus we argue that representation learning methods should encode this orientation invariance. We show that prior methods are sensitive to orientation, which can lead to suboptimal clustering. To address this issue, we develop O2-VAE, an unsupervised learning method that learns robust, orientation-invariant representations. We use O2-VAE to discover novel morphology subgroups in segmented cells and mitochondria, detect outlier cells, and rapidly characterise cellular shape and texture in large datasets, including in a newly generated synthetic benchmark.
Article activity feed
-
l of 15,000 images were generated for each condition. Each image was assigned a unique random seed,
It seems likely this approach will require more data than a similar approach using pre-registered images. Do you have a sense of how much more data would be required for similar accuracy?
-
15,000 images
Have you determined how much data is required for reasonable accuracy? If we were designing an experiment, how many images would be necessary for reasonable accuracy?
-
trainsfer
typo: transferred
-
clases
Typo: Classes
-
Further, prior work shows that orientation-invariant encoders have better performance for supervised models on biological and medical data45, 46, 47, despite supervised learning having more techniques to enforce invariance (e.g. data augmentation). There are also theoretical arguments supporting this idea: a central idea in geometric deep learning is that enforcing known data invariances in neural networks should improve representations
I'm curious about the authors' opinion about whether a pattern needs to be apparent to the human eye in order to be detected by a computer. These statements suggest that the supervised models perform worse than autoencoders like the one described here.
-
mitohconria
typo, mitochondria
-
An object is ‘in contact’ with another object if its separation is smaller than 2 pixels.
I'm curious whether the authors considered using the model to identify specific organelle contacts of interest to further investigate using deconvolved images and confirming the contact.
-
images
typo, should be imaged
-
15,000 images were generated for each condition
This is useful to know. Do you have a sense that 15,000 images was near the minimum number needed or whether it was overkill for the training?
-
Moving beyond segmentation-only data, we take grayscale nuclei images undergoing mitosis24, and show that modelling texture improves the morphologic profiles
Texture is something that can be harder to quantify than cell shape so it is great that it can be captured by models like this one since it improves the morphologic profiles.
-
