The problem. Regulation is long-range: enhancers can sit tens of kilobases from the gene they control. Convolutional models like DeepSEA see only a short window, so they miss distal regulatory interactions — precisely the ones that explain much of gene-expression variation and many noncoding variant effects. You need a model whose receptive field spans the regulatory neighbourhood.

The idea. Enformer replaces deep convolutions’ limited range with transformer self-attention, letting the model integrate information across ~100 kb of DNA. Trained to predict thousands of genomic tracks (expression, chromatin, TF binding) from sequence, its attention can link a gene to distant enhancers. The longer context substantially improves expression prediction and, importantly, sharpens variant-effect prediction and enhancer–gene assignment over shorter-range predecessors.

Why it matters. This is the Transformer — the architecture I read about in the attention paper — arriving in regulatory genomics, the third step in today’s DeepBind → DeepSEA → Enformer arc from short motifs to genome-scale context. It’s also the lineage that leads to DNA foundation models (Evo, on my earlier list). For interpreting the noncoding variants my pipelines produce, long-range context is what makes the predictions credible.

Verdict. A landmark sequence model, influential and widely benchmarked; it’s still correlative, computationally heavy, and imperfect on truly long-range or cell-type-specific effects. Read it as attention giving genomics the long context regulation actually uses.