The problem. Most disease-associated variants from GWAS fall in noncoding regions, where their effect is regulatory and invisible to protein-centric annotation. Predicting whether a base change disrupts a regulatory element requires understanding the sequence determinants of chromatin — TF occupancy, accessibility, histone modification — across many cell types. That’s a lot of context to learn from sequence alone.

The idea. DeepSEA trains a deep convolutional network to predict hundreds of chromatin features (from ENCODE-style profiling) directly from a window of DNA sequence, learning a regulatory-sequence code with single-nucleotide sensitivity. To score a variant, it predicts chromatin state for the reference and alternate alleles and takes the difference — a quantitative prediction of regulatory disruption that improves prioritization of functional eQTLs and disease variants.

Why it matters. DeepSEA extends DeepBind (today’s first read) from single-protein specificity to genome-wide chromatin state, and it leans on the ENCODE catalogue I read earlier as training signal. It’s the direct predecessor of Enformer (next), and it addresses the exact gap my variant pipelines hit: SnpEff tells me a variant is “intergenic,” DeepSEA estimates whether it actually breaks regulation. That’s the noncoding-interpretation frontier.

Verdict. A foundational sequence-to-chromatin model; its context window is short by today’s standards and predictions are correlative, needing experimental follow-up. Read it as the step that made noncoding variants computationally interpretable.