Skip to content

embpy: Universal Biological Embeddings

embpy architecture

embpy is a unified Python package for creating, managing, comparing, and benchmarking multi-modal biological embeddings. It provides a standardized interface to over 100 state-of-the-art embedding models across proteins, genes, cells, molecules, DNA, and natural language.

Core Capabilities

  • Universal Interface (BioEmbedder): A single class replaces dozens of disconnected model pipelines with unified .embed() calls.
  • Multi-Modal Foundation Models: Embed proteins (ESM-2, ProtT5, ESM-3), DNA (Evo, Borzoi), molecules (ChemBERTa, MolFormer), and cells (scGPT, Geneformer).
  • Standardized Output Contracts: Seamless export to anndata (for scverse ecosystems), pandas DataFrames (table), and raw dictionary payloads (payload).
  • Hardware Scalability: Runs smoothly on laptops (CPU / Apple Silicon MPS) and automatically scales to multi-GPU NVIDIA clusters.

Modern Packaging & Cross-Platform Features

embpy brings key structural and usability improvements:

1. Pure pip & uv Compatibility

  • Lightweight Core: A bare pip install embpy installs only essential libraries (numpy, pandas, anndata, scikit-learn, rdkit) without requiring a C++ compiler or CUDA drivers.
  • Modular Dependency Extras: Heavy deep learning stacks are isolated into clean optional extras ([models], [cpu], [gpu], [scib], [all]).
  • Elimination of Git Dependencies: Replaced non-standard git repository dependencies (e.g., scooby) with pure PyPI distributions to ensure reproducibility.

2. Cross-Platform POSIX Gating

  • Native Windows Support: POSIX-only C extensions (such as pysam and pyensembl) are automatically gated on Windows. Core biological embedding workflows function natively on Windows without build errors or WSL.
  • Multiprocessing Safety: Windows multiprocessing defaults to spawn rather than fork. embpy ensures transparent process isolation when scaling batch operations across CPU cores.

3. Decoupled Architecture & Lazy Imports

  • Heavy neural network weights and dependencies (torch, transformers, torch_geometric) are lazily loaded only when a specific model wrapper is invoked. Importing embpy completes in milliseconds.

Next Steps