Allele-Specific Embedding

The foundational step requires phased sequencing data (e.g., from trio-binning, Hi-C phasing, or Strand-seq) to assign heterozygous variants to distinct parental haplotypes. The embedding model receives two separate input sequences—one for the maternal allele and one for the paternal allele—rather than a single ambiguous consensus. This allows the model to learn representations where a heterozygous regulatory variant in a promoter produces divergent embedding vectors for each allele, directly encoding the predicted difference in transcription factor binding affinity.

These embeddings explicitly capture allele-specific expression (ASE) and allele-specific chromatin accessibility within the latent space. By training on matched haplotype-resolved epigenomic data, the model learns to position alleles with active regulatory elements in a distinct region of the embedding manifold from silenced alleles. Key features include:

• Allelic bias vectors: The difference between maternal and paternal embeddings for a given locus
• Imprinting detection: Systematic separation of parent-of-origin-specific silencing patterns
• Variant effect prediction: The embedding distance between reference and alternate alleles correlates with functional impact

Specialized neural architectures process both alleles simultaneously with weight-sharing constraints to ensure consistent feature extraction. Common approaches include:

• Siamese networks: Two identical encoders process maternal and paternal sequences independently, with a contrastive loss applied to their output embeddings
• Cross-allele attention: A transformer variant where maternal tokens attend to paternal tokens at homologous positions, explicitly modeling trans-homolog interactions
• Haplotype-specific positional encoding: Modified positional embeddings that encode not just genomic coordinate but also parental origin as an additional dimension

Training objectives are designed to preserve and amplify allele-specific signals while maintaining robustness to sequencing noise. The primary loss components include:

• Triplet margin loss: Ensures that the embedding of an active maternal allele is closer to other active alleles than to its silenced paternal counterpart
• Allele-specific reconstruction loss: Applied in masked autoencoder frameworks, requiring the model to predict the correct nucleotide for each haplotype independently from a masked input
• Adversarial domain confusion: A gradient reversal layer removes batch effects while preserving allelic differences, preventing the model from collapsing distinct haplotypes into identical representations

Allele-specific embeddings enable clinically actionable insights that collapsed representations miss. Key use cases:

• Compound heterozygote resolution: Determining whether two recessive variants occur on the same allele (benign carrier) or opposite alleles (affected)
• Somatic mutation phasing: In cancer genomics, distinguishing whether driver mutations co-occur on the same DNA molecule, which affects drug sensitivity predictions
• Pharmacogenomic embedding: Encoding which parental copy of a drug-metabolizing enzyme carries a loss-of-function variant, directly informing dosing models
• Non-coding variant interpretation: Prioritizing regulatory variants that create allelic imbalance in enhancer or promoter embeddings

Haplotype-resolved embeddings double the representational footprint, requiring careful engineering:

• Memory: Storing separate maternal and paternal embeddings for every locus increases vector database size by 2x compared to collapsed representations
• Training data requirements: Phased reference panels (e.g., 1000 Genomes Project phased haplotypes) are essential for pre-training; trio-based ground truth data is needed for fine-tuning
• Inference latency: Siamese architectures process two forward passes per locus, though weight sharing allows batching of both alleles through the same encoder
• Long-range phasing: Embeddings for loci separated by centromeres or assembly gaps require specialized handling when phase blocks are disconnected

The computational process of assigning genetic variants to their maternal and paternal chromosomes of origin. Allele-specific embeddings are entirely dependent on accurate phasing to distinguish which variants are in cis (on the same molecule). Statistical phasing uses population reference panels, while trio-phasing leverages parental genotypes for near-perfect resolution. Physical phasing technologies like Hi-C and Strand-seq provide direct experimental evidence of haplotype structure.

Regulatory interactions where a genetic variant on one chromosome copy directly influences the expression of a gene on the same molecule. Allele-specific embeddings are designed to capture these effects by maintaining separate representations for each haplotype. Key mechanisms include:

• Promoter variants altering transcription factor binding
• Enhancer mutations disrupting long-range looping
• Allele-specific open chromatin measured by ATAC-seq

An epigenetic phenomenon where gene expression is determined solely by the parent of origin, resulting in monoallelic expression. Allele-specific embeddings must encode this parent-of-origin bias to accurately model imprinted loci. Classic examples include the IGF2/H19 locus and the SNRPN gene cluster. Imprinting is mediated by differential DNA methylation established during gametogenesis and maintained through somatic cell divisions.

Neural network architectures explicitly designed to process both haplotypes of a diploid genome rather than collapsing them into a single consensus sequence. These models maintain separate embedding channels for maternal and paternal alleles, enabling the detection of allele-specific binding and expression quantitative trait loci (eQTLs). Architectures include dual-tower encoders and haplotype-separated transformer inputs.

The differential expression of the two alleles of a gene, often quantified by counting RNA-seq reads that overlap heterozygous single nucleotide polymorphisms. ASE analysis provides a direct functional readout of cis-regulatory variation. Allele-specific embeddings can be trained to predict ASE from DNA sequence alone, learning the latent regulatory grammar that distinguishes active from inactive haplotypes.

A reference assembly where each chromosome is represented as two distinct haplotypes rather than a single mosaic sequence. The human pangenome reference and T2T-CHM13 assembly provide phased, gapless representations that enable allele-specific read mapping. These resources are critical for training embedding models that generalize across diverse human populations and structural variant landscapes.