DeepSEA Architecture

DeepSEA simultaneously predicts 919 chromatin feature profiles across 125 cell types and tissues, including transcription factor binding, DNase I hypersensitivity, and histone modification marks. This shared representation learning forces the convolutional layers to extract universal regulatory grammars rather than overfitting to a single assay. The multitask objective function combines binary cross-entropy losses for each profile, enabling the model to leverage correlations between related epigenomic marks—for example, H3K4me3 and promoter accessibility—to improve prediction accuracy on data-scarce cell types.

The model employs a three-layer deep convolutional network without pooling layers to preserve single-nucleotide resolution. Key architectural details:

• Layer 1: 320 convolutional kernels of size 8, capturing short motifs like core transcription factor binding sites (e.g., GC-box, TATA-box)
• Layer 2: 480 kernels of size 8, detecting motif interactions and spacing constraints
• Layer 3: 960 kernels of size 8, recognizing higher-order regulatory syntax and distal element signatures
• Receptive field: 1,000 base pairs of input sequence context
• Activation: Rectified linear units (ReLU) throughout
• Regularization: Dropout with 50% rate on the penultimate fully-connected layer
• Output: Sigmoid activation producing independent probability scores for each of the 919 target profiles

DeepSEA computes functional significance scores for non-coding variants by performing computational mutagenesis—predicting chromatin profiles for both reference and alternate alleles, then quantifying the absolute difference. The method:

• Input: 1,000-bp sequence centered on the variant position
• Process: Forward pass with reference sequence, then forward pass with alternate allele
• Output: A 919-dimensional vector of predicted effect scores, each representing the absolute change in probability for a specific chromatin feature
• Aggregation: The maximum effect score across all profiles serves as the variant's overall functional impact prediction This approach transforms an unannotated SNP into a quantitative, mechanistically interpretable functional prediction without requiring population-level association data.

DeepSEA was trained on genome-wide chromatin profile data from the ENCODE and Roadmap Epigenomics consortia, processing the human reference genome (GRCh37/hg19) in 200-bp non-overlapping windows. Training specifics:

• Dataset size: Approximately 15 million genomic windows covering chromosomes 1-21
• Validation: Chromosome 22 held out entirely for hyperparameter tuning
• Test: Chromosome Y excluded due to its unique evolutionary properties
• Class imbalance handling: Weighted loss function compensating for the sparsity of positive chromatin feature annotations, which typically occupy less than 1% of the genome
• Optimization: Stochastic gradient descent with momentum, batch size of 100, trained for 60 epochs This exhaustive training regime ensures the model captures the full diversity of regulatory element architectures across the human genome.

DeepSEA supports gradient-based interpretability to identify which nucleotides drive specific chromatin predictions. The method computes the partial derivative of a target output with respect to each input nucleotide, producing a saliency map that highlights functionally critical bases. Applications include:

• Motif discovery: Identifying known and novel transcription factor binding motifs without prior position weight matrix libraries
• Variant mechanism elucidation: Pinpointing whether a disease-associated SNP disrupts a specific transcription factor binding event
• Regulatory syntax analysis: Revealing spacing and orientation constraints between cooperative binding factors
• Disease variant prioritization: Ranking non-coding GWAS hits by their predicted disruption of tissue-relevant chromatin features This interpretability layer transforms DeepSEA from a black-box predictor into a hypothesis-generating discovery tool for regulatory genomics.

The architecture's sequence-only input requirement—predicting chromatin states directly from DNA without requiring experimental data—enables cross-species applications. By applying the human-trained model to orthologous genomic regions in model organisms, researchers can:

• Identify conserved regulatory elements that maintain similar chromatin profiles across evolutionary distances
• Detect lineage-specific regulatory innovations where sequence changes alter predicted chromatin states
• Prioritize model organism experiments by focusing on regions with high human-predicted regulatory activity
• Validate enhancer annotations through transgenic reporter assays guided by DeepSEA predictions This capability bridges the gap between human genomics and functional validation in tractable experimental systems.

A hybrid deep learning model that combines a convolutional layer with a bidirectional LSTM to predict DNA function. The convolutional layer detects regulatory motifs, while the recurrent layer captures the syntax and spacing rules between them.

• Optimizes the quantification of regulatory motif syntax
• Outperforms pure CNN architectures on DeepSEA benchmarks
• The LSTM component learns long-range dependencies between bound transcription factors

A computational perturbation technique that systematically introduces virtual mutations into a DNA sequence to quantify their predicted impact on a model's output. This is the primary method for interpreting DeepSEA predictions.

• Computes a variant effect score by comparing wild-type and mutant predictions
• Identifies which nucleotides are critical for transcription factor binding
• Enables high-throughput screening of non-coding variants without wet-lab experiments

A model interpretability method that attributes the prediction of a genomic neural network to its input sequence by accumulating gradients along a path from a baseline to the actual input. It satisfies the completeness axiom, ensuring attribution sums to the prediction difference.

• Provides nucleotide-resolution importance scores
• More robust than simple saliency maps for DeepSEA-style models
• Used to identify putative transcription factor binding sites driving predictions

A neural network training strategy where a single model simultaneously predicts multiple epigenomic assays across different cell types. DeepSEA pioneered this approach by sharing hidden representations across 919 chromatin profiles.

• Shared convolutional layers learn universal regulatory grammar
• Task-specific output heads specialize for individual assays
• Improves generalization on data-scarce cell types through representation sharing