ALPHAGenome Embeddings: Human Genome hg38 in 3,072-Dimensional Latent Space

~22,000 non-overlapping 131 KB bins of the human genome (hg38/GRCh38) are embedded into a 3,072-dimensional latent space using the ALPHAGenome foundation model. The embeddings were created by the lagosproject and were last updated on the platform in May 2026. These pre-computed embeddings power an interactive browser visualization for exploring latent relationships between genomic regions.

Use Cases

• Visualizing latent relationships between genomic regions based on the 3,072-dimensional embeddings.
• Developing downstream machine learning models for genomic prediction tasks using pre-computed embeddings.
• Exploring the structure of the human genome in a compressed, continuous latent space.
• Benchmarking new genomic foundation models against the ALPHAGenome embedding representations.

Strengths

• Covers the entire human genome (hg38/GRCh38) divided into ~22,000 bins.
• Provides high-dimensional (3,072-D) embeddings for each genomic bin.
• Has a companion interactive visualization tool (UMAP Explorer) for exploration.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

lagosproject

Collection Method

Pre-computed using the ALPHAGenome DNA foundation model.

Freshness

Last updated 2026-05-26 13:48:32; freshness should be verified.