# eDNA_Stream_E01

## Scope
- Establish whether squiggle-based taxonomy is feasible on the small computational budget that is available

## Dataset:
- NO-MISS bacterial isolates, v10.4.1 chemistry on PromethION flow cells (https://epi2me.nanoporetech.com/nomiss_96bc_p2i_sup_2026/)
    - Downsampling at the pod5 level intended for compute/time reasons -> the full dataset is 293 pod5 / 1.6 TB of raw data 

## Dependencies:
- dorado >= v 2.0.0
- aws CLI
- conda/mamba (mamba recommended)
- python env with:
    - python 3.14
    - snakemake
    - python-dotenv
- custom-models package: live in a separate repo (`git@git.tk-ai.eu:Tom/eDNA_Stream_E01_custom_models.git`) included here as the `custom_models` git submodule; the workflow pip-installs it from that checkout (see `workflow/envs/torch_*.yaml`)
    - clone with: `git clone --recurse-submodules ...` (or `git submodule update --init` on an existing clone)
    - update to a newer version: `git submodule update --remote custom_models`, then commit the new pinned commit
- .env file with NCBI API key -> slightly speeds up reference fasta download

## Running:
```bash
cd workflow
snakemake --cores <cores> --resources gpu=<n_gpus> --use-conda
```

## Notes:
- Reads aligned to the FP traps by minimap are excluded for training -> esp. the E. coli reads are highly represented and should not be there, but are stable even when only using stringent alignment