biotrainer-core

A lightweight Python package providing shared data classes, I/O utilities, and helper functions for protein-related projects.

Purpose

biotrainer-core is designed as a minimal, dependency-light foundation that can be reused across multiple protein ML projects (e.g. biotrainer) without pulling in heavy frameworks like PyTorch or Transformers.

Features

Data Classes

Pydantic-based data models shared across protein ML workflows:

• Sequence data — representations for protein sequences and their metadata
• Protocols — supported task types (residue-level, sequence-level, contact prediction, …)
• Metrics — structured containers for evaluation metrics
• Model results & predictions — standardised output formats for biotrainer models
• Embedding statistics — summary statistics over embedding spaces
• AutoEval — data classes for automated evaluation tasks, reports, and benchmark datasets (FLIP, PBC)

H5 File Handling

Utilities for reading and writing HDF5 files used to store protein embeddings:

• H5Database — high-level interface for embedding databases
• Low-level helpers for chunked I/O and dataset management

Functions

Pure utility functions with no side effects:

• Hashing — deterministic sequence hashing
• Seeding — reproducible random seed management
• Ranking — ranking helpers for evaluation
• Bootstrapping — confidence interval estimation via bootstrapping

Input Files

• FASTA parsing — read and validate FASTA files into typed sequence objects

Installation

pip install biotrainer-core

License

MIT — see LICENSE for details.