DeepSEA

DeepSEA employs a multi-task learning architecture to simultaneously predict 919 chromatin features from a single DNA sequence input. This includes 125 DNase I hypersensitivity profiles, 690 transcription factor binding ChIP-seq profiles, and 104 histone modification profiles. By sharing hidden representations across diverse experimental assays, the model leverages common regulatory logic to improve generalization and reduce overfitting compared to training separate models for each assay.

The model uses a three-layer convolutional neural network without pooling to scan sequence motifs at multiple resolutions. The first layer applies 320 convolutional filters of size 8 to detect short primary motifs. The second layer uses 480 filters of size 8 to capture motif interactions. The third layer applies 960 filters of size 8 to model higher-order regulatory grammar. This hierarchical design exponentially expands the receptive field while preserving spatial information critical for precise binding site localization.

DeepSEA computes a functional significance score for non-coding variants by performing in silico mutagenesis. For any single nucleotide variant, the model predicts chromatin profiles for both the reference and alternate allele sequences. The difference between these predictions, measured as the absolute log-fold change in binding probability, quantifies the regulatory impact of the variant. This enables prioritization of expression quantitative trait loci (eQTLs) and disease-associated variants from genome-wide association studies.

Input sequences are encoded using one-hot encoding, where each nucleotide (A, C, G, T) is represented as a binary vector of length four. This sparse representation preserves the exact positional identity of every base without imposing assumptions about nucleotide similarity. The model processes the full 1,000 bp context window centered on the region of interest, allowing it to capture both proximal promoter elements and distal enhancer motifs that influence chromatin state through long-range interactions.

To prevent information leakage caused by sequence homology between training and test sets, DeepSEA employs a rigorous cross-chromosome validation strategy. The model is trained on chromosomes 1–21 and evaluated on chromosome 22, with chromosome X reserved for additional testing. This ensures that performance metrics reflect true generalization to unseen genomic contexts rather than memorization of homologous sequences shared across chromosomes.

By analyzing the learned convolutional filters, DeepSEA reveals the cis-regulatory grammar encoded in genomic DNA. The first-layer filters converge on known transcription factor binding motifs, including AP-1, CTCF, and NF-κB consensus sequences. Higher-layer filters capture combinatorial interactions between motifs, such as cooperative binding and spacing constraints. This provides a data-driven approach to motif discovery without requiring prior knowledge of transcription factor sequence preferences.