Ambiguity Codes

Extends standard 4-channel one-hot encoding to handle IUPAC ambiguity codes by representing each degenerate base as a probability distribution over the four canonical nucleotides.

• R (A or G): [0.5, 0, 0.5, 0]
• N (any base): [0.25, 0.25, 0.25, 0.25]
• Preserves uncertainty information rather than making arbitrary base calls
• Compatible with convolutional neural networks expecting fixed-dimensional input channels

Strategies for incorporating ambiguity codes into genomic language model vocabularies without exploding token count or losing semantic meaning.

• Direct inclusion: Add IUPAC codes (R, Y, S, W, K, M, B, D, H, V, N) as distinct tokens in the vocabulary
• Canonicalization: Replace degenerate bases with the lexicographically smallest possible nucleotide before tokenization
• Multi-token expansion: Split ambiguous positions into multiple valid k-mers, each weighted by prior probability
• DNABERT and Nucleotide Transformer handle ambiguity codes differently, affecting downstream variant calling performance

Adapting MLM pre-training objectives to account for positions of natural uncertainty in reference genomes.

• Masking an N and predicting the distribution over {A, C, G, T} teaches the model to learn base composition biases
• Loss functions can be modified to use cross-entropy over a soft target rather than a hard nucleotide label
• Pre-training on genomes with realistic ambiguity patterns improves robustness to sequencing errors during fine-tuning
• Models learn contextual cues that resolve degenerate positions based on flanking sequence motifs

Ambiguity codes introduce complexity into strand-invariance enforcement because degenerate characters have symmetric complements.

• R (A|G) complements to Y (C|T), not a single base
• Canonical k-mer strategies must account for degenerate pairs: R↔Y, S↔S, W↔W, K↔M
• N is self-complementary, simplifying strand collapse for fully ambiguous positions
• Reverse complement augmentation must correctly transform IUPAC codes to maintain biological validity of training data

Deep learning variant callers must distinguish between true genetic variation and reference genome ambiguity codes that represent known polymorphic positions.

• An R in the reference at a known SNP site should not trigger a variant call when either A or G is observed
• Models like DeepVariant encode reference ambiguity as prior probabilities in the pileup tensor
• Confusion between ambiguity codes and true variants is a common source of false positives in clinical pipelines
• Proper handling requires integrating population frequency data with the IUPAC-encoded reference

Ambiguity codes in linear references are often flattened representations of structural variation best captured by pangenome graphs.

• Graph-based references replace a single N or R with explicit alternative paths through the genome
• Embedding methods like Graph2Vec or node2vec learn representations that preserve allelic diversity
• Transitioning from IUPAC ambiguity to graph-based representations eliminates information loss at polymorphic sites
• Pangenome embeddings outperform linear reference embeddings for variant-aware tasks in diverse populations