Basset Framework

Basset simultaneously predicts chromatin accessibility across 164 cell types using a shared convolutional trunk. This multitask approach forces the network to learn universal regulatory grammars rather than cell-type-specific noise. The architecture uses three convolutional layers with 300, 200, and 200 filters respectively, followed by two fully connected layers of 1000 units each. Each cell type receives its own output neuron with a sigmoid activation, enabling the model to capture both shared and cell-type-specific regulatory logic.

Basset accepts raw 600-base-pair DNA sequences centered on DNase I hypersensitivity sites. The input is one-hot encoded into a 4 x 600 binary matrix representing A, C, G, and T nucleotides. Critically, the model requires no pre-computed genomic features, evolutionary conservation scores, or chromatin state annotations. This end-to-end learning from raw sequence enables the discovery of novel regulatory motifs without human bias, making it a true de novo motif discovery engine.

The first convolutional layer employs 19-width filters that scan across the input sequence, functioning as trainable position weight matrices. Each filter learns to detect a specific sequence motif, such as transcription factor binding sites. The rectified linear unit activation introduces non-linearity, while max-pooling across the spatial dimension reduces dimensionality and provides translational invariance. This design mirrors the biological reality that regulatory proteins bind short, degenerate sequence patterns.

A defining capability of Basset is its use for computational perturbation analysis. By systematically introducing every possible single-nucleotide variant into a query sequence and measuring the predicted change in accessibility, researchers can quantify the functional impact of non-coding mutations. This technique generates a mutation impact map that identifies critical regulatory nucleotides. The approach has been validated against experimental Massively Parallel Reporter Assays, confirming that Basset's predictions correlate strongly with measured regulatory activity.

Basset's interpretability stems from visualizing learned convolutional filters as sequence logos. By scanning the genome with each first-layer filter and extracting the maximally activating sequences, researchers can generate position frequency matrices that reveal the binding preferences of each filter. Many filters converge on known transcription factor motifs such as AP-1, CTCF, and NF-kB, while others discover novel regulatory patterns. This transparency was groundbreaking for deep learning in genomics, addressing the black-box criticism.

The shared convolutional layers learn a universal regulatory vocabulary that generalizes across cell types. Filters detecting core promoter elements or enhancer signatures are reused across all prediction tasks, while the fully connected layers specialize for cell-type-specific outputs. This architectural inductive bias enables Basset to make accurate predictions for cell types with limited training data by leveraging patterns learned from data-rich cell types, demonstrating early principles of epigenomic transfer learning.