Deep Motif Dashboard: Visualizing and Understanding Genomic Sequences Using Deep Neural Networks

1 Introduction

• DNNs are viewed as black boxes because of their complexity, making their predictions hard to understand in biomedical settings.Understanding predictions is important for doctors and researchers using models.Source: §1 · p. 1
• The DeMo Dashboard explores neural models and three visualization strategies, including saliency maps, temporal output scores, and class optimization for class-specific motif patterns.The paper introduces these methods to help understand how models make predictions.Source: §1 · pp. 1–2
• TFBS classification asks whether an input DNA sequence contains a site where a particular transcription factor binds.Transcription factors are regulatory proteins that bind DNA to regulate cell machinery.Source: §1 · p. 1
• CNN-RNN outperforms the other tested models and its visualizations show focus on traditional motifs and long-range dependencies among them.The evaluation uses a large-scale benchmark TFBS dataset.Source: §1 · p. 2

2 Deep Neural Models for TFBS Classification

• TFBS computational methods are needed because ChIP-seq is slow and expensive and does not reveal patterns common across positive sites; prior CNN work left the best neural architecture unclear.The models use one-hot nucleotide inputs and a softmax binary classifier within a shared end-to-end framework.Source: §2 · p. 2

2.1 Convolutional Neural Network (CNN)

• The CNN uses temporal convolutions to extract local motifs, then max-pooling to reduce dimensions, provide translation invariance, and capture higher-level features across wider input regions.Source: §2.1 · p. 3

2.2 Recurrent Neural Network (RNN)

• The RNN uses bidirectional LSTMs to model sequence dependencies, averages timestep outputs in each direction, concatenates them, and classifies.LSTM gates control when information is written, read, or forgotten.Source: §2.2 · p. 4

2.3 Convolutional-Recurrent Network (CNN-RNN)

• The CNN-RNN feeds convolutional feature vectors sequentially into an LSTM, averages its outputs across time, and uses the result for classification to model temporal patterns among motifs.Source: §2.3 · p. 4

3 Visualizing and Understanding Deep Models

• DeMo Dashboard visualizes saliency maps, temporal output scores, and class optimizations to inspect how deep models make TFBS predictions.The dashboard is intended to improve understanding of model predictions in a biomedical setting.Source: §3 · p. 5

3.1 Saliency Maps

• Saliency maps use first-order derivatives at a test sequence to rank nucleotide influence on the class score, with magnitudes showing importance regardless of derivative direction.The derivative is computed by one backpropagation step; the map highlights which nucleotides need to change least to affect the score most.Source: §3.1 · p. 5

3.2 Temporal Output Scores

• Temporal output scores track recurrent-model predictions across progressively longer sequence prefixes, showing where the decision changes; the method is applied to RNNs and CNN-RNNs, including reverse sequences.CNNs process the sequence at once, so this visualization is not used for them.Source: §3.2 · p. 5

3.3 Class Optimization

• Class optimization finds a locally optimal input sequence maximizing positive-TFBS probability for a particular TF, using stochastic gradient descent while holding model weights fixed.This class-specific visualization is not tied to a particular test sequence.Source: §3.3 · pp. 5–6

3.4 End-to-end Automatic Motif Extraction from the Dashboard

• Automatic motif extraction selects length-9 subsequences from saliency maps and temporal scores, while the class-optimized sequence itself serves as a motif.Saliency motifs maximize summed contiguous saliency; temporal motifs mark the strongest negative-to-positive score change.Source: §3.4 · p. 6

3.5 Connecting to Previous Studies

• Unlike image and language interpretability studies, this work seeks DNA signals using DNN understanding; against Alipanahi et al., saliency maps reportedly find motif patterns better than convolution activations.Prior work also applied convolution-activation visualization to a convolutional-recurrent model for noncoding variant prediction.Source: §3.5 · p. 6

4.1 Experimental Setup

• The benchmark uses 108 K562 TF datasets, averaging 30,819 training sequences per dataset, with 1,000 test sequences and 101 DNA bases per sequence.Positive and negative sequences are evenly split; negatives are dinucleotide-preserving shuffles, and a separate model is trained for each TF.Source: §4.1 · pp. 6–7

4.2 TFBS Prediction Performance of DNN Models

• CNN-RNN achieves the best mean AUC among the three DNN architectures; all deep models significantly beat MEME, CNN beats RNN, and CNN-RNN beats CNN.The pairwise comparisons use the best-AUC model of each architecture.Source: Table 2 · p. 7, Table 3 · p. 7, §4.2 · p. 7

4.1 Experimental Setup

• The authors compare the best-performing hyperparameter variants of CNN, RNN, and CNN-RNN against MEME-ChIP and a prior CNN, using motif-extraction baselines as well.They show smaller and larger variants around each architecture's best model to check for underfitting or overfitting.Source: §4.1 · p. 7

4.3 Understanding DNNs Using the DeMo Dashboard

• The dashboard places class optimization above saliency maps and temporal output scores for the same positive test sequence, alongside known JASPAR motifs.The class optimization represents the positive class generally, rather than a specific test sequence.Source: Figure 2 · p. 8
• Across visualization-derived motif matches, CNN ranks first, followed by CNN-RNN and RNN; the authors hypothesize CNN-RNN's AUC advantage reflects sequential interactions among motifs.Tomtom comparisons cover 57 of 108 TF datasets with JASPAR motifs; sequence-specific techniques average matches across 500 positive sequences.Source: Table 4 · p. 9, §4.3 · p. 10
• CNN saliency concentrates on short motif-like subsequences, RNN saliency spreads across the sequence, and CNN-RNN saliency highlights motifs plus influential distant nucleotides.For GATA1, CNN-RNN is 99% confident while nucleotides outside the motif also influence its prediction.Source: §4.3 · p. 9
• Temporal output switches from negative to positive usually occur near JASPAR motifs, with no clear difference between forward and backward patterns.In the NFYB example, CNN-RNN switches to positive while RNN does not; their saliency maps also focus on different regions.Source: §4.3 · p. 9
• CNN class optimizations often resemble known TF motifs, while recurrent-model optimizations are less clear and can show reverse-complement motifs.Class-optimized sequences can give a general indication of what triggers a positive TFBS prediction.Source: §4.3 · p. 9

• DNNs are highly accurate for TFBS classification but difficult to interpret, slowing their practical adaptation.Source: §5 · p. 10
• DeMo Dashboard explores three DNN architectures for TFBS prediction and introduces three visualization methods to clarify how models work.Source: §5 · p. 10
• Its visualizations still require human examination; the authors frame the dashboard as a starting point and hope it prompts further genomic-sequence visualization studies.Source: §5 · p. 10
• The authors plan to extend the dashboard to related applications, noting DNNs' excellent results in epigenomic analysis [22].Source: §5 · p. 10