Beluga

A chromatin model can only score a variant against features it was trained to predict: ask the 2015 DeepSEA network about a histone mark in a cell type outside its 919-profile panel and there is no output to difference. Beluga rebuilds that network against the full Roadmap Epigenomics release — 2,002 transcription factor binding, DNase I accessibility and histone mark profiles spanning more than 200 human cell types and tissues — and doubles the convolutional depth to absorb them. It reads 2,000 bp of DNA and returns one sigmoid probability per feature for the central 200 bp, so 1,800 bp of flanking sequence acts as context rather than as the region being labelled. Variant effects come from the same allele-pair procedure the line has always used: run the frozen network twice, on reference and alternate sequence, and difference the two 2,002-dimensional outputs. No variant data enters training, so rare and never-observed alleles are scored exactly as common ones are.

Jian Zhou, Chandra Theesfeld, Kevin Yao, Kathleen Chen, Aaron Wong and Olga Troyanskaya built it at the Lewis-Sigler Institute for Integrative Genomics at Princeton University and the Flatiron Institute, and published it in Nature Genetics in 2018. The name is not an acronym — the group code-names its sequence models after whales, and the checkpoint ships as deepsea.beluga.pth. The HumanBase server labels it the 2019 version of DeepSEA, after the year its PyTorch conversion was posted.

It was released as the sequence half of ExPecto, which scans ±20 kb around a transcription start site with Beluga in a 200 bp-step moving window and feeds the spatially transformed chromatin predictions into per-tissue expression models. The chromatin network is a separate artifact with its own checkpoint, and is used on its own wherever chromatin-level rather than expression-level predictions are wanted.

#Key Features

• 2,002 chromatin profiles in one head: Transcription factor binding, DNase I hypersensitivity and histone marks from ENCODE and the full Roadmap Epigenomics release are predicted jointly, so a sequence feature learned for one assay is available to every other.
• Allele-pair variant scoring: Chromatin effect is the difference between reference and alternate predictions. The shipped script also evaluates each variant at shifted window positions out to 800 bp, catching features whose peak is offset from the variant.
• Strand-symmetric inference: Each input is one-hot encoded and scored alongside its reverse complement, so predictions do not depend on which strand a region was submitted on.
• Frozen, redistributable checkpoint: The weights are a single 598 MB PyTorch state dict on Zenodo, also packaged as a Kipoi model and hosted behind a server taking VCF, FASTA or BED.

#Technical Details

The network is six convolutional layers in three blocks of two, at widths 320/320, 480/480 and 640/640, all with kernels of width 8 and ReLU activations, with max-pooling of width and stride 4 after the first two blocks and dropout of 0.2 inside them. A dropout of 0.5 precedes a fully connected layer that flattens the 67,840-unit representation into 2,003 units, followed by a second linear layer to the 2,002 outputs and a sigmoid. That is 149,536,251 parameters, twice the convolutional depth of the 919-feature predecessor and a 2,000 bp input window in place of 1,000 bp. Training used hg19 with chromosomes 8 and 9 withheld entirely and 4,000 samples from chr7:30,508,751–35,296,850 as validation, leaving chr8 free to serve as a clean whole-chromosome holdout for everything built on top.

#Applications

The direct use is prioritising noncoding variants: given a VCF, the model reports how each allele shifts every one of 2,002 chromatin features, which is the raw material for asking whether a GWAS or eQTL variant plausibly disrupts regulation and in which cell types. The HumanBase server takes VCF, FASTA or BED input for up to 10,000 variants or sequences per submission, while the standalone code and the Kipoi package handle genome-scale jobs. Because the outputs are a dense, cell-type-resolved representation of a sequence rather than a single score, they also serve as features for downstream models — expression prediction, disease-impact scoring, in silico mutagenesis scans.

#Impact

Beluga's scalability is what made its parent framework's survey of more than 140 million promoter-proximal mutations feasible, and the checkpoint became a standard starting point for chromatin-level variant work: the pretrained BelugaMultiplexer predicts the effects of all possible substitutions across a 2,000 bp window in one pass over the same 2,002 profiles. Three constraints bound the artifact. Coordinates are hg19 only, and the standalone code expects a local reference FASTA. The 2,000 bp window cannot see distal enhancer–promoter contacts, which is why the expression framework wraps it in a ±20 kb scan and why later models such as Enformer use far longer receptive fields. And the code and weights carry a Princeton academic-use agreement limited to teaching and not-for-profit research.