Deep Representation Learning FOR Morphological Profiling

Morphological Profiling

CellProfiler extracts ~ 100 - 1000 cell level features

We perform feature / dimensional reduction and calculate the mean for each perturbation

Example Use Case: Classifying Mechanism of ACTIONS

Deep learning can learn Meaningful Representations

Word2Vec

DCGANs

maybe for Profiles ?

+

=

Cells treated with A

Cells treated with B

Cells treated with A and B ?

or not?

Deep Profiles

Can we learn a robust representation from unlabeled data ? 

1000-dim CP feature vector
https://swarbrickjones.files.wordpress.com
cropped cell images
1000-dim CP feature vector
cropped cell images

latent space

LEarning A Meaningful Latent Space

subpopulation Profiles

unsupervised clustering in latent space

(semi) supervised training using hand-labeled data

NOISY LABELS: From Cell level to whole image level

Classifying and Segmenting Microscopy Images Using Convolutional Multiple Instance Learning, Arxiv 2015, O Kraus, LJ Ba, B Frey 

Unguided SUBPOPULATION DETECTION USING ATTENTION?

Spatial Transformer Networks NIPS 2015 Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu
Made with Slides.com