Loading…
Noncoding variant effect prediction from DNA sequence, scoring how an allele shifts 919 chromatin features across ENCODE and Roadmap cell types.
No providers recorded yet. Browse all providers
A single base change can destroy a transcription factor's binding site, while the same change a few hundred base pairs away leaves binding untouched — and which happens depends on flanking sequence a motif scanner reading only the binding window never sees. DeepSEA widens the window and learns the readout end to end, taking 1 kb of genomic sequence and predicting, for 919 chromatin features at once, the probability that the central 200 bp lies inside a peak. A variant is then scored by running the frozen network twice, on reference and alternate alleles, and differencing the two 919-dimensional outputs. No variant data enters training, so rare and previously unobserved alleles are scored exactly as common ones are.
Jian Zhou and Olga Troyanskaya built DeepSEA at the Lewis-Sigler Institute for Integrative Genomics at Princeton University, with Troyanskaya also at the Simons Center for Data Analysis of the Simons Foundation; it appeared in Nature Methods in 2015. Three choices carried the result: 1 kb of context instead of the roughly 300 bp windows earlier binding predictors used, a hierarchy of convolution and pooling layers reading sequence at several spatial scales, and multitask training in which all 919 predictors share one feature representation, so a motif useful for one transcription factor is immediately available to a physically interacting one. Together with Basset, it established the convolutional sequence-to-chromatin paradigm that shaped regulatory genomics for the rest of the decade. The name now covers a line of successor models and the server hosting them; this entry describes the original 919-feature model and its own reported numbers.
Three one-dimensional convolution layers with 320, 480 and 960 kernels of width 8 and ReLU activations, max-pooling of width and stride 4 after the first two, and dropout, feed a fully connected layer that integrates the full 1,000 bp before a sigmoid layer over the 919 outputs. Training ran in Torch7 on a Tesla K20m GPU under weight-decay and max-norm regularisation. Labels came from 200 bp bins carrying at least one TF binding event — 521,636,200 bp, or 17% of GRCh37 — with chromosomes 8 and 9 held out for testing and 4,000 chromosome 7 samples for validation. On held-out sequence the median AUC was 0.958 for TF binding, 0.923 for DNase I hypersensitive sites and 0.856 for histone marks, against a median 0.896 for gkm-SVM on the same TF binding data. Against 57,407 allelically imbalanced SNPs from DNase-seq footprinting across 35 cell types, it chose the more accessible allele with over 95% accuracy on the 6,726 variants whose predicted probability difference exceeded 0.1.
The clearest use is triaging noncoding variants. Boosted logistic regression classifiers over the 1,842-feature vector prioritised HGMD regulatory mutations, GRASP noncoding eQTLs and GWAS Catalog trait-associated SNPs against distance-matched 1000 Genomes controls, beating annotation-based methods that saw chromatin annotations the sequence model never did, and separating trait-associated SNPs even from nearby non-associated ones. Individual predictions line up with experiment, such as increased FOXA1 affinity at the breast cancer risk locus rs4784227. Users reach the model through a hosted server taking VCF, BED or FASTA input, a standalone bundle for larger jobs, the lab's Selene library for retraining it, or PyTorch conversions of the weights on Zenodo and Kipoi.
The 919-task dataset became a default benchmark for genomic sequence models, still reported by TransBind, Big Bird, GENA-LM and HyenaDNA long after the architecture was superseded, and the design is reused directly — CAMBUS is a DeepSEA-architecture network widened to a 5 kb input. Three limitations bound the original artifact: its 1 kb window cannot represent distal enhancer–promoter regulation, which later dilated-convolution and attention models were built to reach; coordinates are GRCh37/hg19 only; and the released checkpoints are Torch7-era, so the converted PyTorch weights are the practical route to running them, while the hosted server now defaults to newer models in the same line.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.