DEEP MOTIF DASHBOARD: VISUALIZING AND UNDERSTANDING GENOMIC SEQUENCES USING DEEP NEURAL NETWORKS.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Deep neural networks (DNNs) achieve high accuracy in transcription factor binding site (TFBS) classification.
• Understanding the biological basis of DNN predictions for TFBS remains a challenge.

Purpose of the Study:

• To develop a toolkit, the Deep Motif Dashboard (DeMo Dashboard), for visualizing and interpreting DNN models in TFBS classification.
• To elucidate how DNNs identify sequence motifs and dependencies for TFBS prediction.

Main Methods:

• Implemented saliency maps to identify nucleotide importance for predictions.
• Introduced temporal output scores for recurrent models to track prediction evolution.
• Utilized class-specific visualization via stochastic gradient optimization to find optimal binding sequences.

Main Results:

• Evaluated convolutional, recurrent, and convolutional-recurrent network architectures.
• The convolutional-recurrent network (CNN-RNN) architecture demonstrated the best performance.
• Visualization techniques revealed that CNN-RNN models capture both sequence motifs and their interdependencies.

Conclusions:

• The DeMo Dashboard provides valuable insights into DNN decision-making for TFBS classification.
• CNN-RNN models effectively learn complex sequence patterns relevant to transcription factor binding.