This work was defended as an M.Sc. thesis at the Faculty of Informatics, Masaryk University, and published as a workshop paper at the GenBio workshop at ICML 2026: [OpenReview], [poster].
Interactive information-content map: https://genomeinfo.dyn.cloud.e-infra.cz/.
- Test existing genomic language models.
- Determine their compression ratio on the human genome (or other organisms) and compare it with standard compression models (e.g., gzip, 7zip).
- It could also be interesting to analyze how much of the compression is attributed to the tokenizer versus the language model itself.
- Since, unlike natural language, the compressibility of different genomic regions will vary significantly, it might be interesting to create an "information content map" across the human genome.
src/llm_and_compression/arithmetic_encoder/GPTProbabilityModel.py- LLM interface responsible for token-by-token language modelling designed to be plugged into Arithmetic encoder.src/llm_and_compression/arithmetic_encoder/coder.py- implementation of the encoding algorithm.
- Clone the repository.
- Install uv package:
pip install uv- Install the package in editable mode:
uv pip install -e .An example notebook demonstrating how to use the implemented framework for compressing genomic sequences using LLMs can be found at: src/notebooks/compression_demo.ipynb.