Slide 1
Slide 1 text
Peax
Fritz Lekschas, Ph.D. candidate
Harvard University
Interactive Concept Learning for Visual
Exploration of Epigenetic Patterns
April 17, 2019
Brand Peterson, Daniel Haehn, Eric Ma, Nils Gehlenborg, and Hanspeter Pfister
Bio-IT World
Slide 2
Slide 2 text
No content
Slide 3
Slide 3 text
Search
Slide 4
Slide 4 text
Search
? ? ? ? ? ?
Slide 5
Slide 5 text
No content
Slide 6
Slide 6 text
The Epigenome
Slide 7
Slide 7 text
Davis et al. (2018) The Encyclopedia of DNA elements (ENCODE): data portal update.
Slide 8
Slide 8 text
Davis et al. (2018) The Encyclopedia of DNA elements (ENCODE): data portal update.
Slide 9
Slide 9 text
Davis et al. (2018) The Encyclopedia of DNA elements (ENCODE): data portal update. Maurano et al. (2012).
>90% of disease-
associated variants
found in GWAS are
located in non-
coding regions
Slide 10
Slide 10 text
Why not just search
computationally?
Slide 11
Slide 11 text
• Little to no ground truth
• Peak calling is not solved
• Feature calling is very hard
• Formally defining patterns is hard
Slide 12
Slide 12 text
• Little to no ground truth
• Peak calling is not solved
• Feature calling is very hard
• Formally defining patterns is hard
Visual quality control
Slide 13
Slide 13 text
• Little to no ground truth
• Peak calling is not solved
• Feature calling is very hard
• Formally defining patterns is hard
Visual quality control
Slide 14
Slide 14 text
• Little to no ground truth
• Peak calling is not solved
• Feature calling is very hard
• Formally defining patterns is hard
Visual quality control
Interactive visual query
Slide 15
Slide 15 text
How does Peax
find similar patterns?
Slide 16
Slide 16 text
1. Encoding 2. Active Learning & User Interface
Slide 17
Slide 17 text
1. Data Processing 2. Convolutional Autoencoder
Slide 18
Slide 18 text
Trained 6 autoencoders: 3 window sizes × 2 data types
Data types: DNase, histone mark ChIP
Window and bin sizes: 3 kb (25 bp), 12 kb (100 bp), 120 kb (1000 bp)
120 DNase-seq datasets from ENCODE
49 histone mark ChIP-seq experiments from Roadmap Epigenomics:
H3K4me1/me3, H3K27ac/me3, H3K9ac/me3, H3K36me3
Slide 19
Slide 19 text
3 kb 12 kb 120 kb
DNase-seq
R2 .98 .90 .78
Slide 20
Slide 20 text
3 kb 12 kb 120 kb
Average histone mark ChIP-seq
R2 .84 .69 .73
Slide 21
Slide 21 text
Is the learned encoding useful for similarity search?
Yes!
Slide 22
Slide 22 text
Is the learned encoding useful for similarity search?
Yes!
Yes!
Slide 23
Slide 23 text
Ours: CAE
ED
SAX
DTW
XCORR
UMAP
TSFRESH
Slide 24
Slide 24 text
Ours: CAE
ED
UMAP TSFRESH XCORR DTW
SAX
Slide 25
Slide 25 text
Ours: CAE
ED
UMAP TSFRESH XCORR DTW
SAX
Slide 26
Slide 26 text
No content
Slide 27
Slide 27 text
Query view
Reconstructions
Slide 28
Slide 28 text
List view
Labeling
Tabs
Slide 29
Slide 29 text
Embedding
Slide 30
Slide 30 text
Progress
Slide 31
Slide 31 text
No content
Slide 32
Slide 32 text
Find Asymmetrical Peaks
Using a Visual Query as the Start
Slide 33
Slide 33 text
Select Pattern for Querying
Initial Sampling
Binary Labeling
Train First Classifier
Active Learning Sampling
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially Freely browse and select a
region for quering
Size of the selected region is
fixed and based on the
autoencoder
We use HiGlass (Kerpedjiev et
al. 2018) as the genome
browser
Slide 34
Slide 34 text
Select Pattern for Querying
Initial Sampling
Binary Labeling
Train First Classifier
Active Learning Sampling
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially Freely browse and select a
region for quering
Size of the selected region is
fixed and based on the
autoencoder
We use HiGlass (Kerpedjiev et
al. 2018) as the genome
browser
Slide 35
Slide 35 text
Initial Sampling
Binary Labeling
Train First Classifier
Active Learning Sampling
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Increase distance of samples
to the query
Sample regions in dense areas
Maximize pairwise distance
between samples
All in the latent space
Slide 36
Slide 36 text
Binary Labeling
Train First Classifier
Active Learning Sampling
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Select regions that match and
do not match the query
Inconclusive regions can
simply be skipped
Slide 37
Slide 37 text
Train First Classifier
Active Learning Sampling
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
A random forrest classifier is
trained online with the labels
Each time a new set of
samples is requested a new
classfier is trained
A new classifier can also be
trained in between after labels
have changed
Slide 38
Slide 38 text
Active Learning Sampling
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Regions are sampled by their:
- prediction uncertain
- proximity to the target
- in dense neighborhoods
- with high pairwise distance
Slide 39
Slide 39 text
Training Progress
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Progress is tracked for every
trained classifier
Uncertainty is the overall
prediction probability
Change of the prediction
probaility
Convergence and divergence
Slide 40
Slide 40 text
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Convergence and divergence
2D UMAP embedding of all
encoded regions
Probability color encoder:
⬤ means matching
⬤ means non-matching
⬤ means unpredictable
View is interactive and dots are
selectable
Slide 41
Slide 41 text
Embedding View
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Convergence and divergence
2D UMAP embedding of all
encoded regions
Probability color encoder:
⬤ means matching
⬤ means non-matching
⬤ means unpredictable
View is interactive and dots are
selectable
Slide 42
Slide 42 text
Resolve Conflicts
Explore Final Results Spatially
al. 2018) as the genome
browser
Convergence and divergence
View is interactive and dots are
selectable
Peax warns about false
positives and negatives when
the labels and the classifier's
predictions disagree
Slide 43
Slide 43 text
Explore Final Results Spatially
al. 2018) as the genome
browser
Convergence and divergence
View is interactive and dots are
selectable
The query view is interactive
A bed-like track shows the
prediction probabilities:
⬤ means matching
⬤ means non-matching
⬤ means unpredictable
Slide 44
Slide 44 text
Find Differentially Accessible Peaks
Using Existing Peak Calls
Slide 45
Slide 45 text
ENCODE e11.5 DNase-seq from face and hindbrain
Differential, central, strong peak calls
Balance positives and negatives
Initial Classifier
Slide 46
Slide 46 text
ENCODE e11.5 DNase-seq from face and hindbrain
Differential, central, strong peak calls
Balance positives and negatives
Initial Classifier
Slide 47
Slide 47 text
ENCODE e11.5 DNase-seq from face and hindbrain
Differential, central, strong peak calls
Balance positives and negatives
Initial Classifier
Slide 48
Slide 48 text
Resolve Conflicts
Refine Labels & Assess Recall
Slide 49
Slide 49 text
Explore Local Neighborhoods
Slide 50
Slide 50 text
Final Results Excluding Labels
Slide 51
Slide 51 text
CONCLUSION
Leverage machine learning to improve visual exploratory pattern search
Generative models work well as they show the encoding quality
FUTURE WORK
Explore other types of encoders (VAE, GAN, LSTM, etc.)
Compare and evaluate different active learning strategies
Slide 52
Slide 52 text
PRERINT, SOURCE CODE, MODELS, ETC.
peax.lekschas.de
[email protected]
@flekschas
lekschas.de
CONTACT
Thank You!