Multi-Scale Genomic Feature

The foundational vectorization strategy that converts raw nucleotide sequences into dense, learnable feature vectors. Multi-scale features extend this by creating embeddings at multiple resolutions simultaneously—from single-base pair tokenization to kilobase-scale binning. Common approaches include:

• One-hot encoding at base-pair resolution for local motif detection
• k-mer frequency vectors capturing short sequence patterns
• Learned embeddings from DNA language models like DNABERT or Enformer
• Binned aggregate signals for distal regulatory context

The multi-scale approach ensures models capture both transcription factor binding sites (local) and domain-level chromatin states (global).

A genome-wide matrix quantifying interaction frequencies between all pairs of genomic loci, serving as the primary prediction target for 3D genome folding models. Multi-scale features are critical because Hi-C data exhibits distance-dependent signal decay—contacts between loci separated by 1 kb differ fundamentally from those at 1 Mb. Effective models must:

• Capture fine-scale interactions (1-10 kb) for enhancer-promoter loops
• Model intermediate-range contacts (10-500 kb) for TAD structure
• Represent compartment-level interactions (>1 Mb) for A/B compartment assignment

The multi-scale representation aligns the input feature hierarchy with the natural scale-dependent structure of chromatin interaction data.

A deep learning model from DeepMind that predicts gene expression and epigenetic tracks from 200 kb DNA sequences using a multi-scale architecture. Key design elements include:

• Convolutional stem capturing local sequence motifs at base-pair resolution
• Transformer layers with dilated attention patterns for long-range interactions
• Multi-scale feature hierarchy where lower layers detect binding sites and upper layers integrate distal enhancer effects
• Input representation combining one-hot encoded DNA with evolutionary conservation scores

Enformer demonstrates that multi-scale features are essential for accurate prediction of regulatory activity, directly informing 3D genome folding models that must operate across similar distance scales.

A self-interacting genomic region where DNA sequences physically contact each other more frequently than with sequences outside the domain. Multi-scale features are essential for TAD boundary prediction because:

• Local features (1-10 kb) capture CTCF binding site motifs and orientation
• Intermediate features (10-100 kb) model cohesin loading and extrusion dynamics
• Domain-level features (100 kb-1 Mb) represent insulation scores and directionality indices

Models using only single-scale representations often fail to detect boundaries because they miss the convergent CTCF motif orientation that requires analyzing sequence context at multiple distances simultaneously.

A deep learning architecture representing genomic loci as nodes and their interactions as edges in a graph. Multi-scale features are encoded as node attributes at different resolutions:

• Fine-scale nodes represent individual regulatory elements with sequence-level features
• Intermediate nodes aggregate signals across TAD-sized regions
• Coarse nodes capture compartment-level chromatin states
• Hierarchical message passing propagates information between scales

This graph-based multi-scale representation naturally mirrors the hierarchical organization of chromatin, from individual CTCF binding events to whole-chromosome territory positioning. GNNs trained on multi-scale features outperform flat convolutional architectures on loop prediction tasks.

A deep convolutional neural network that predicts 3D genome folding directly from DNA sequence alone, serving as a foundational architecture for sequence-to-structure prediction. Akita's multi-scale design includes:

• Base-pair resolution convolutions detecting sequence motifs and binding sites
• Dilated convolutions expanding the receptive field to capture long-range dependencies
• Multi-scale feature fusion combining local motif signals with distal context before predicting contact maps
• Input window of ~1 Mb to capture full TAD structures and flanking regions

Akita demonstrates that multi-scale genomic features enable accurate prediction of Hi-C contact maps without requiring experimental epigenomic data, validating the sufficiency of sequence-derived multi-scale representations.