Sei

A chromatin-level sequence model can tell you a variant lowers one H3K27ac signal in one cell line, but not what that means for the sequence as a whole: the answer is thousands of correlated, track-by-track numbers with no summary. Sei answers in two stages. A convolutional network reads 4 kb of genomic sequence and predicts, for the base at its centre, peak membership in each of 21,907 cis-regulatory profiles — transcription factor binding, histone marks and chromatin accessibility across more than 1,300 cell lines and tissues. Those predictions are projected onto 40 sequence classes, a vocabulary of regulatory programs obtained by clustering the model's own genome-wide output. A variant becomes a signed score on a named class — E7 monocyte/macrophage enhancer, P promoter, CTCF–cohesin — rather than a wall of track deltas.

Kathleen Chen and Jian Zhou built Sei with Olga Troyanskaya at Princeton University and the Flatiron Institute of the Simons Foundation, Zhou by then at UT Southwestern Medical Center; it appeared in Nature Genetics in 2022. It continues the line the lab began with DeepSEA, which established the sequence-to-chromatin paradigm alongside Basset: where the original read 1 kb of hg19 and 919 features, Sei reads 4 kb of GRCh38 and 21,907. Where Enformer widened the receptive field, Sei widened the readout and then made it legible.

#Key Features

• Forty sequence classes as a regulatory vocabulary: Louvain clustering over Sei predictions for 30 million 4-kb windows tiling the genome at 100 bp steps yields 61 clusters; the largest 40 become sequence classes covering over 97.4% of the genome, grouped into promoter, 12 enhancer classes, CTCF, TF, Polycomb, heterochromatin, transcription and low-signal.
• Directional variant effects: A variant's score is the alternate-minus-reference difference in sequence class score, so the sign separates mutations that raise a regulatory activity from those that abolish it. Histone predictions are first normalised by predicted nucleosome occupancy, so a mark rising only because nucleosome density rose is not read as a gain.
• Broad chromatin readout: 9,471 transcription factor binding, 10,064 histone mark and 2,372 accessibility profiles spanning roughly 1,000 DNA-binding proteins and 77 histone marks, mostly processed by the Cistrome Project and the rest by ENCODE and Roadmap Epigenomics.
• Runs on any input sequence: The framework takes FASTA, BED or VCF against hg19 or hg38 and returns chromatin predictions and sequence class scores, with no retraining. Each variant falls in exactly one class, so analyses partition signal without overlapping annotations.

#Technical Details

The network stacks residual blocks with paired linear and nonlinear paths at 480, 640 and 960 channels — the linear path trains quickly and efficiently, the nonlinear path carries interaction structure — followed by five residual dilated convolution layers that widen the receptive field without downsampling. A B-spline spatial basis transformation then reduces 256 spatial bins to 16 spline dimensions — far cheaper than a fully connected layer — before a hidden layer and a 21,907-way sigmoid output; the two 21,907-wide output layers dominate the roughly 890 million parameters. Training used PyTorch and the lab's Selene library with on-the-fly sampling, holding out chromosomes 8 and 9 for test and 10 for validation and excluding ENCODE blacklist regions. On held-out sequence the model averaged 0.972 AUROC and 0.409 AUPRC across all 21,907 profiles, and on the 2,002 profiles it shares with the lab's earlier DeepSEA Beluga model it improved 19% on average, measured as AUROC/(1 − AUROC).

#Applications

Sequence class scores hold up against independent genetics data: variants predicted to raise enhancer activity correlate with higher GTEx eQTL expression, Polycomb gains with lower, and variants perturbing enhancer, promoter or CTCF classes are markedly less likely to be common in 1000 Genomes. Across 47 UK Biobank GWAS traits the partition recovers tissue-specific architecture — monocyte enhancer E7 for monocyte count, brain enhancers for cognitive traits, liver/intestine E9 for high cholesterol — yielding 83 class–trait associations that survive conditioning on baseline annotations. Scoring 853 HGMD regulatory disease mutations put over 99% of the strongest effects in enhancer, promoter or CTCF classes, including the one in five predicted to be gains of function.

#Impact

Sei now serves as a frozen feature extractor for downstream regulatory models — Otari concatenates Sei chromatin features into its transcript graph node attributes — and the lab distributes precomputed sequence class annotations for hg38 and hg19. The same group's later Puffin pursues base-resolution interpretability at promoters instead. Three constraints bound Sei: the 4-kb window cannot represent long-range enhancer–promoter contacts; the sequence classes come from the model's own predictions, so they inherit the Cistrome-dominated training compendium and are defined for the human genome only; and while weights, training data and annotations sit on Zenodo under CC BY 4.0, the framework code carries a BSD-derived licence limited to academic and research use, with commercial use routed through Princeton's licensing office.