Cosine Similarity

Cosine similarity measures the cosine of the angle between two embedding vectors, ranging from -1 (opposite) to 1 (identical). In genomics, this is critical because the magnitude of a gene expression vector often reflects sequencing depth or transcript abundance, while the direction captures the relative expression pattern. By normalizing out magnitude, cosine similarity focuses purely on the regulatory or functional signature, making it the default metric for comparing gene expression profiles and regulatory element embeddings.

In genomic language models like DNABERT or Enformer, cosine similarity between the CLS token embeddings of two promoter sequences correlates with shared regulatory function. Key applications include:

• Regulatory element clustering: Grouping enhancers with similar tissue-specific activity
• Variant effect prediction: Measuring the cosine distance between wild-type and mutant sequence embeddings to estimate functional impact
• Cross-species alignment: Comparing orthologous regulatory regions by their embedding similarity rather than raw sequence identity

Cosine similarity serves as the scoring function in contrastive loss and SimCLR frameworks adapted for genomics. The model is trained to maximize cosine similarity between positive pairs (e.g., two augmented views of the same promoter, or orthologous sequences from different species) while minimizing it for negative pairs. This directly shapes the latent space geometry, organizing embeddings on a hypersphere where angular proximity encodes biological relatedness.

Because cosine similarity is scale-invariant, it provides inherent resistance to systematic shifts in vector magnitude caused by batch effects. If a sequencing run introduces a global scaling factor to all expression values, Euclidean distances would be distorted, but cosine similarities remain stable. This property makes it the preferred metric when integrating embeddings across multiple laboratories, sequencing platforms, or experimental conditions without explicit batch correction.

Cosine similarity enables fast approximate nearest neighbor (ANN) search in vector databases storing genomic embeddings. Since cosine similarity is equivalent to inner product on L2-normalized vectors, libraries like FAISS can index billions of sequence embeddings and retrieve functionally similar regulatory elements in milliseconds. This powers applications like:

• Finding all known enhancers similar to a novel candidate
• Querying gene expression atlases for analogous profiles
• Identifying off-target binding sites for engineered transcription factors

For vectors A and B, cosine similarity = (A · B) / (||A|| × ||B||). In genomic embeddings:

• A value of 1.0 indicates identical orientation (e.g., duplicate sequences or perfect functional orthologs)
• A value of 0.0 indicates orthogonal, unrelated functions
• A value near -1.0 is rare in standard genomic models but can emerge in contrastively trained spaces where opposing biological processes (e.g., activation vs. repression) are pushed apart
• Zero vectors (all-zero embeddings for padding tokens) produce undefined similarity and must be filtered before computation

The straight-line distance between two vectors in embedding space, calculated as the square root of the sum of squared differences. Unlike cosine similarity, Euclidean distance is magnitude-sensitive, meaning it penalizes vectors that differ in length even if they point in the same direction. In genomics, this can be problematic when comparing regulatory elements with varying expression levels but similar functional profiles. For normalized embeddings, Euclidean distance and cosine similarity are monotonically related.

The scalar result of multiplying corresponding elements of two vectors and summing the products. For unit-normalized vectors, the dot product is exactly equal to the cosine similarity. In genomic embedding models, the dot product is often used directly as a logit in the final classification layer, representing the unnormalized alignment between a sequence embedding and a learned weight vector. It is computationally cheaper than cosine similarity but conflates magnitude with directional alignment.

Also known as L1 distance, this metric sums the absolute differences between vector components. It is less sensitive to outliers than Euclidean distance and is often preferred in high-dimensional sparse spaces, such as k-mer frequency vectors. In single-cell RNA-seq analysis, Manhattan distance is sometimes used to compare gene expression profiles because it better captures the modular nature of biological pathways where a few genes dominate the signal.

A measure of linear correlation between two vectors, calculated as the covariance divided by the product of standard deviations. When applied to mean-centered vectors, Pearson correlation is mathematically equivalent to cosine similarity. In genomics, it is the standard metric for constructing gene co-expression networks, where the focus is on relative expression changes across conditions rather than absolute abundance levels.

A set-based metric that measures the overlap between two sets divided by their union. For binary vectors representing the presence or absence of k-mers or genomic features, Jaccard similarity provides an interpretable alternative to cosine similarity. It is commonly used in metagenomic binning to compare the k-mer profiles of assembled contigs and in MinHash algorithms for rapid genome-scale similarity estimation.

The geodesic distance on the unit hypersphere, defined as the arccosine of the cosine similarity divided by pi. This metric satisfies all formal triangle inequality requirements, making it a proper distance metric unlike cosine similarity itself. In genomic embedding spaces, angular distance is used when downstream algorithms require a true metric, such as hierarchical clustering of transcription factor binding motifs or phylogenetic tree construction from sequence embeddings.