MycoNet

Small toolkit for classifying DNA/ITS sequences using k-mer/recoding, an embedding layer and a BiLSTM.

Quickstart

• Create environment & install dependencies (recommended):

pip install uv
uv install

• Train a tiny model (CLI):

python train.py path/to/input.fasta experiment_name 10 --epochs 50 --batch-size 256

• Predict using a trained experiment:

python predict.py queries.fasta experiment_name results.csv

Input format

• Files expect alternating header and sequence lines (header on one line, sequence on the next). Sequences should be long enough to pass configured filtering (--minseq, default 100).

Model formats

• Training saves a single-file Keras archive (.keras) and also attempts to export a TensorFlow SavedModel directory. The code prefers the .keras archive for reload with Keras 3. predict.py will load a .keras file if present, or wrap a SavedModel directory with TFSMLayer for inference.

Tests

• Run the full test suite with:

pytest -q

• Tests live under tests/. tests/conftest.py ensures the repo root is on sys.path and includes a cleanup fixture that removes temporary small_test* directories after the session.

Repository layout

• MycoNet/ — package modules (recode.py, make_model.py, kmer_embedding.py, data.py, utils.py).
• train.py — promoted training CLI (canonical training entrypoint).
• predict.py — prediction script (loads .keras or SavedModel dirs).
• tests/ — pytest tests and fixtures.

Notes

• Tested on Python 3.12 with TensorFlow/Keras; GPU usage is optional and TensorFlow will try to enable memory growth when GPUs are available.

Performance on simulated data: