Context-Aware Significance of Cross-gene Attribution for Discovering Elements (CASCADE): a position-specific statistical framework for identifying model-derived candidate regulatory elements from in silico saturation mutagenesis of DNA sequence models.
A gene's sequence around its transcription start site is embedded by a fine-tuned DNA language model and decoded into predicted expression across many cell types (top). CASCADE recovers regulatory motifs at cell-type-, tissue-, and gene-specific resolution (bottom).
training/-- a training pipeline for a cell-type-resolved sequence-to-expression model: a convolutional encoder and multi-task decoder that predicts expression across many cell types at once from a precomputed per-gene sequence embedding.cascade/-- the CASCADE seqlet-calling method itself: a drop-in, per-position-null replacement for TF-MoDISco's seqlet detector, plus an example driver script.
Each subdirectory has its own README with the full detail, requirements, and usage examples.
For each gene, a frozen sequence-embedding backbone (not included here --
bring your own, e.g. a DNA language model) produces a per-base embedding matrix
X_g in R^(L x D) for a fixed window around the gene. A convolutional decoder
(two 1-D convolutions with batch normalization and ReLU, dropout, global max
pooling, layer normalization, a low-rank gene-latent bottleneck, and a
two-layer MLP head) maps X_g to one predicted expression value per cell
type, for all cell types simultaneously ("multi-task" training). See
training/README.md for the full architecture and training recipe.
TF-MoDISco calls a seqlet by rolling a window over an attribution track and
comparing the window score to one pooled null fit across all positions.
CASCADE fits that same null per position instead, so a position with
uniformly high background attribution gets a stricter significance bar than a
position where sequences differ a lot from each other. Everything else
(seqlet extraction, non-maximum suppression, downstream clustering into
motifs) is unmodified MoDISco. See cascade/README.md for the full method and
an example driver.
Released under the MIT License (see LICENSE).
