Sequence-to-function model predicting cell-type- and disease-specific gene expression from DNA, trained on pseudobulk profiles from 22 million cells.
No providers recorded yet. Browse all providers
Sequence-to-function models read a stretch of genomic DNA and predict what will be transcribed from it. The models that defined the field — Enformer and its successor Borzoi — were trained almost entirely on bulk assays measured over whole tissues or cell lines. A bulk track averages over every cell in a sample, so a model fitted to it cannot tell you that a promoter variant matters in CD4+ T cells and nowhere else, or that an enhancer behaves differently in fibroblasts from an inflamed ileum than a healthy one.
Decima closes that gap by training on single-cell data instead. Developed at Genentech's gRED Computational Sciences group and first posted in October 2024, it predicts one gene's expression across thousands of cell type, tissue, and disease contexts at once, from the DNA sequence surrounding that gene. Rather than modeling individual cells — intractable at atlas scale — Decima aggregates single-cell and single-nucleus RNA-seq counts into pseudobulk profiles, one per unique combination of cell type, tissue, disease state, and source study, and learns to predict that whole vector from sequence.
The model has become the reference point for a small lineage of cell-type-resolved sequence models: DanioDecima extends the architecture to zebrafish embryogenesis, initializing from Decima and Borzoi weights to test how far mammalian regulatory grammar carries across 450 million years of evolution. The work appeared in Nature Methods in 2026.
Decima is initialized from Borzoi, a convolutional–transformer hybrid trained on bulk RNA-seq, CAGE, DNase-seq, ATAC-seq, and ChIP-seq, so it inherits epigenomic structure hard to learn from scRNA-seq alone. Borzoi's output head is replaced with global mean pooling along the sequence axis and a linear layer emitting 8,856 values. The input is a 524,288 bp window containing the gene's transcription start site with at least 163,840 bp upstream, encoded as five channels: one-hot DNA plus a binary mask marking the gene body — the extra channel Borzoi lacks. A two-component loss combines a Poisson term on total gene expression with a multinomial term across pseudobulks, pushing the model toward between-condition rather than between-position differences.
The corpus draws on atlases from SCimilarity plus human brain, skin, and retina atlases, totaling over 22 million cells and yielding an 8,856 × 18,457 pseudobulk matrix. Four replicates were trained, one per Borzoi replicate, and averaged. On 1,811 held-out genes, the mean Pearson correlation between measured and predicted expression was 0.80 per pseudobulk and 0.58 per gene. On 984 fine-mapped single-cell eQTLs from OneK1K, Decima beat Borzoi at separating causal variants from matched controls in 19 of 21 cell types, and predicted effect direction correlated with eQTL beta at rho = 0.42, rising to 0.58 and 87% directional accuracy for variants it predicted any effect for.
The immediate uses are variant interpretation and regulatory dissection. A noncoding GWAS or eQTL hit can be scored across cell types to nominate both the target gene and the cell population where the effect is realized — information bulk models cannot supply. Attribution over disease/healthy pseudobulk pairs surfaces the motifs behind a disease expression program. The generative direction matters for gene therapy: the paper's proof of concept evolves a 200 bp element predicted to fire in Crohn's disease fibroblasts while staying silent in neighboring gut cell types, the specificity AAV cargo design requires.
Decima established pseudobulk aggregation as a workable bridge between atlas-scale single-cell data and long-context sequence models, and supplied a cell-type-resolved sc-eQTL benchmark. The release is unusually complete: code, documentation, all four replicates as checkpoints and safetensors on Hugging Face, and the pseudobulk matrix as a documented dataset repo. Two caveats matter. Disease-versus-healthy prediction is the weakest axis, averaging 0.24 Pearson correlation across 565 matched pairs — real signal, but far below cross-cell-type performance. And the designed elements are validated in silico only. Licensing needs care: the code carries the Genentech Non-Commercial Software License, whose definition of commercial use reaches pre-clinical and clinical development, and the same four checkpoints are declared MIT on Hugging Face but CC BY-NC on Zenodo.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.