Peax Interactive Concept Learning for Visual Exploration of Epigenetic Patterns

Epigenomic data expresses a rich body of diverse patterns that help to identify regulatory elements like promoter, enhancers, etc. But finding these patterns reliably genome-wide is challenging. Peax is a tool for interactive visual pattern search and exploration of epigenomic patterns based on unsupervised representation learning with convolutional autoencoders. The visual search is driven by manually labeled genomic regions for actively learning a classifier to reflect your notion of interestingness.

Transcript

1. 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
   

   View Slide

2. View Slide

3. Search
   

   View Slide

4. Search
   ? ? ? ? ? ?
   

   View Slide

5. View Slide

6. The Epigenome
   

   View Slide

7. Davis et al. (2018) The Encyclopedia of DNA elements (ENCODE): data portal update.
   

   View Slide

8. Davis et al. (2018) The Encyclopedia of DNA elements (ENCODE): data portal update.
   

   View Slide

9. 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
   

   View Slide

10. Why not just search
    computationally?
    

    View Slide

11. • Little to no ground truth
    • Peak calling is not solved
    • Feature calling is very hard
    • Formally defining patterns is hard
    

    View Slide

12. • Little to no ground truth
    • Peak calling is not solved
    • Feature calling is very hard
    • Formally defining patterns is hard
    Visual quality control
    

    View Slide

13. • Little to no ground truth
    • Peak calling is not solved
    • Feature calling is very hard
    • Formally defining patterns is hard
    Visual quality control
    

    View Slide

14. • 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
    

    View Slide

15. How does Peax
    find similar patterns?
    

    View Slide

16. 1. Encoding 2. Active Learning & User Interface
    

    View Slide

17. 1. Data Processing 2. Convolutional Autoencoder
    

    View Slide

18. 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
    

    View Slide

19. 3 kb 12 kb 120 kb
    DNase-seq
    R2 .98 .90 .78
    

    View Slide

20. 3 kb 12 kb 120 kb
    Average histone mark ChIP-seq
    R2 .84 .69 .73
    

    View Slide

21. Is the learned encoding useful for similarity search?
    Yes!
    

    View Slide

22. Is the learned encoding useful for similarity search?
    Yes!
    Yes!
    

    View Slide

23. Ours: CAE
    ED
    SAX
    DTW
    XCORR
    UMAP
    TSFRESH
    

    View Slide

24. Ours: CAE
    ED
    UMAP TSFRESH XCORR DTW
    SAX
    

    View Slide

25. Ours: CAE
    ED
    UMAP TSFRESH XCORR DTW
    SAX
    

    View Slide

26. View Slide

27. Query view
    Reconstructions
    

    View Slide

28. List view
    Labeling
    Tabs
    

    View Slide

29. Embedding
    

    View Slide

30. Progress
    

    View Slide

31. View Slide

32. Find Asymmetrical Peaks
    Using a Visual Query as the Start
    

    View Slide

33. 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
    

    View Slide

34. 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
    

    View Slide

35. 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
    

    View Slide

36. 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
    

    View Slide

37. 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
    

    View Slide

38. 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
    

    View Slide

39. 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
    

    View Slide

40. 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
    

    View Slide

41. 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
    

    View Slide

42. 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
    

    View Slide

43. 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
    

    View Slide

44. Find Differentially Accessible Peaks
    Using Existing Peak Calls
    

    View Slide

45. ENCODE e11.5 DNase-seq from face and hindbrain
    Differential, central, strong peak calls
    Balance positives and negatives
    Initial Classifier
    

    View Slide

46. ENCODE e11.5 DNase-seq from face and hindbrain
    Differential, central, strong peak calls
    Balance positives and negatives
    Initial Classifier
    

    View Slide

47. ENCODE e11.5 DNase-seq from face and hindbrain
    Differential, central, strong peak calls
    Balance positives and negatives
    Initial Classifier
    

    View Slide

48. Resolve Conflicts
    Refine Labels & Assess Recall
    

    View Slide

49. Explore Local Neighborhoods
    

    View Slide

50. Final Results Excluding Labels
    

    View Slide

51. 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
    

    View Slide

52. PRERINT, SOURCE CODE, MODELS, ETC.
    peax.lekschas.de
    lekschas@seas.harvard.edu
    @flekschas
    lekschas.de
    CONTACT
    Thank You!
    

    View Slide