Skip to content

NotesFEATUREencoder.md

Latest commit

 

History

History
42 lines (28 loc) · 1.63 KB

NotesFEATUREencoder.md

File metadata and controls

42 lines (28 loc) · 1.63 KB
Raw

DNA + Genomic Feature Encoder

Overview

The encode function transforms DNA sequences and associated genomic features into a numerical format suitable for convolutional neural networks (CNNs). It is designed to be flexible and biologically interpretable, supporting both per-base and global features.

The main purpose is to prepare input for models predicting transcription factor (TF) binding, epigenetic states, or other genomic properties.

Features

  • DNA One-Hot Encoding:
    Each nucleotide is encoded as a 4-dimensional one-hot vector (A, C, G, T). This allows convolutional kernels to learn motif patterns directly from sequence.

  • Optional Features:
    Additional genomic features can be included, such as:

    • ATAC: Chromatin accessibility (per-base or global)
    • METH: DNA methylation (per-base or global)
    • Future features: ChIP-seq signals, conservation scores, etc.

  • Flexible Feature Selection:
    Users can specify which features to include via a feature_list. The order of features is preserved, which is critical for reproducibility and consistency when initializing motif-based convolutional filters.

  • Global Feature Broadcasting:
    Scalar features (e.g., average accessibility over the region) are broadcast across the sequence to be compatible with the per-base CNN input.

Usage

# DNA only
enc = encode("ACGTACGT", {}, ["DNA"])

# DNA + global ATAC
enc = encode("ACGTACGT", {"ATAC":1}, ["DNA","ATAC"])

# DNA + per-base ATAC + per-base methylation
atac = [1,0,1,0,1,0,1,0]
meth = [0,1,0,1,0,1,0,1]
enc = encode("ACGTACGT", {"ATAC":atac,"METH":meth}, ["DNA","ATAC","METH"])