MODEL LEARNING PATH
DeepSEA
DeepSEA maps a DNA window to chromatin-feature probabilities using convolution and max pooling. Comparing reference and alternate sequence predictions produces variant-effect features.
Loading the interactive page…
An early landmark in multitask sequence-to-chromatin modeling and noncoding-variant prioritization. The author lab's subsequent Selene library preserves an explicit DeepSEA implementation, providing a concrete continuation of this architectural approach.
Chapters
Read local sequence
A one-hot DNA window enters the first motif-detecting convolution and max pool.
Follow motif detection through a convolution, activation and pooling.
Combine motif evidence
A second convolution/pool stage and a third convolution build higher-order sequence features.
Explain how repeated convolution and pooling expand sequence context.
Predict assay features
Flattened features feed a classifier with separate probabilities for chromatin assays.
Map sequence features to independent assay probabilities.
Compare alleles or train
Reference/alternate predictions support variant scoring; observed assay labels support binary supervision.
Compare allele-specific predictions without treating differences as causal proof.
All exercises in this guide
- ReLU
- Dense 1-D dilated convolution
- Affine Projection
- Activations and gated feed-forward transitions
- Windowed max pooling and coordinate routing
- Affine projection and embedding lookup
- DeepSEA Variant Effects
- Edit and crop an alternate-allele DNA window
- Reverse-complement sequences and align strand-aware outputs
- Crop, orient and average genomic predictions
- Align insertion/deletion predictions to reference coordinates
- Splice and chromatin variant-effect scores
- Masked Sum
- Masked Log-Sum-Exp
- Binary Cross-Entropy with Logits
- Masked Softmax
- Masked Mean
- Stable categorical and binary classification losses