PEDLA: predicting enhancers with a deep learning-based algorithmic framework

[cgreene]

https://dx.doi.org/10.1101/036129

[gwaybio]

published doi: http://dx.doi.org/10.1038/srep28517

[gwaybio]

Goal

Predict enhancers using heterogeneous genomic data across different cell types

Biology

Use histone mod, TFs, chromatin accessibility, DNA methylation, conservation, etc. to train a neural network to extract the probability of an enhancer at given genomic locations. These probabilities are then used in a Hidden Markov Model to predict enhancer states.

Computational Aspects

• Significant amount of preprocessing of heterogeneous features
  • RPKM-like measurement for each feature at 200 base pair bins
  • 0-1 normalized
• Gold standard enhancers defined based on H3K27ac peaks
• Unsupervised pre-training with deep belief network
  • Weights of DBN used to initialize deep neural network
• DNN outputs probability of enhancer given the location
  • Output probability is used in subsequent HMM to predict enhancer states
  • More info in supplementary figure S1
• Validation defined as enhancer overlap with p300, NANOG, OCT4, SOX2 binding sites

Pros vs. Cons

Pros

• After extensive preprocessing it seems to be a reasonable way of integrating heterogeneous features
• Extensive performance metric reporting
• Very nice analytical comparison of method to alternative approaches (all machine learning based) and performs better than all
• Nice discussion of previous limitations in predicting enhancers and how PEDLA overcomes them
  • Low training samples
  • difficult to integrate heterogenous data
  • there is high class imbalance (PEDLA corrects for by normalizing by class imbalance probability in HMM)

Cons

• Questionable gold standard and validation decisions
  • Gold standard labeling is one of the most difficult steps for enhancer classifiers but difficult to compare methods
  • 200 base pair bins - if any prediction overlaps an H3K27ac peak anywhere, is it considered to have successfully identify an enhancer?
• PEDLA doesn't really work across different tissues
  • They claim it does but it really trains on one and then uses weights in a new training model on a different cell type/tissue
• In discussion: "...these studies inspired us to further improve PEDLA through the incoporation of 3D structure information."
  • No 3D structure features used

[cgreene]

Tagged as study. Closing for now. @gwaygenomics : if you want to highlight this example in the 'study' section, feel free to bring it in!