One-Hot Encoding

Each nucleotide base (A, C, G, T) is mapped to a unique index in a binary vector of length 4. The presence of a base is indicated by a 1 at its designated position, while all other positions are 0.

• A → [1, 0, 0, 0]
• C → [0, 1, 0, 0]
• G → [0, 0, 1, 0]
• T → [0, 0, 0, 1]

This scheme prevents the model from inferring false ordinal relationships (e.g., that G is 'greater than' A) that would be implied by integer encoding.

Synthetic genomic data generators, such as GANs and VAEs, require continuous numerical inputs. One-hot encoding serves as the standard bridge between discrete nucleotide sequences and the differentiable operations of neural networks.

• The generator outputs a probability distribution over the 4 bases at each position.
• The discriminator receives either a real one-hot sequence or a softmax-relaxed synthetic sequence.
• This format is directly compatible with convolutional layers that scan for local patterns.

The encoding scheme is deterministic and requires no learned parameters, making it a transparent first step in any genomic deep learning pipeline.

• Zero ambiguity: The mapping is fixed and universally understood.
• Debugging ease: A malformed vector (e.g., multiple 1s or all 0s) is immediately detectable.
• Baseline performance: Serves as a benchmark against which more complex, learned nucleotide embeddings are compared.

The standard IUPAC nucleotide code includes ambiguous characters (e.g., N for any base, R for A or G). One-hot encoding can be extended to handle these.

• A simple approach maps 'N' to a uniform vector: [0.25, 0.25, 0.25, 0.25].
• This represents maximum uncertainty at a given position.
• For a strict binary representation, ambiguous bases are often filtered out or imputed to a specific base during preprocessing.

The resulting matrix for a genomic sequence is highly sparse: for a sequence of length L, the one-hot matrix has dimensions L x 4, with exactly L ones and 3L zeros.

• This sparsity can be exploited for memory-efficient storage using compressed sparse row (CSR) formats.
• However, for very long sequences (e.g., whole chromosomes), the memory footprint can become a bottleneck, motivating the use of more compact k-mer or learned embedding strategies.

One-hot encoding perfectly preserves the original sequence information without loss, which is critical for applications like motif preservation in synthetic data generation.

• A generated softmax vector can be argmax-converted back to a discrete base with perfect fidelity.
• This lossless property ensures that the encoding step itself does not introduce artifacts into the synthetic genomic data pipeline.

A learned, dense vector representation that maps discrete nucleotide bases (A, C, G, T) into a continuous space. Unlike one-hot encoding's sparse, orthogonal vectors, embeddings capture biochemical similarities—for instance, purines (A, G) may cluster closer than pyrimidines (C, T). These are the direct input layer for most deep learning models in genomics.

The occurrence count of short subsequences of length k in a genome. This is a critical statistical property that synthetic genomic data generators must preserve. Key points:

• A sequence's k-mer profile acts as a statistical fingerprint
• Mismatches between real and synthetic k-mer distributions indicate mode collapse
• Typical k values range from 3 to 7 for capturing local sequence context

The proportion of guanine-cytosine bases in a DNA sequence. This is a key genomic signature that synthetic data generators must control to avoid introducing artificial distributional shifts. Why it matters:

• Real genomes have species-specific GC content (e.g., ~41% in human)
• GC bias affects sequencing coverage depth and PCR amplification
• Uncontrolled bias leads to synthetic data that fails downstream variant calling

A metric for evaluating synthetic genomic data quality by comparing the distribution of generated sequences to real sequences in a feature space. How it works:

• Extract features from both real and synthetic datasets using a pre-trained model
• Model each distribution as a multivariate Gaussian
• Calculate the Frechet distance between the two Gaussians
• Lower scores indicate higher fidelity synthetic data

An evaluation paradigm where a predictive model is trained on synthetic genomic data and tested on real data. This measures the utility of synthetic data by its ability to substitute for real data in downstream tasks. Example: Train a variant caller on synthetic reads, then evaluate its precision/recall on real sequencing data. High TSTR performance validates the generator's practical value.

A mathematical framework that adds calibrated noise to generative model training, providing a provable guarantee that synthetic genomic data does not reveal the presence of any single individual. Core concept: The privacy budget (epsilon) quantifies the trade-off—lower epsilon means stronger privacy but potentially less useful synthetic data. Essential for sharing genomic datasets across institutions.