MINpred

MINpred is an alignment-free, four-phase deep-learning tool for identifying and classifying nitrogen mineralization-related enzymes. It first distinguishes enzymes from non-enzymes, identifies nitrogen mineralization enzymes, assigns them to ten functional classes, and finally predicts class-specific EC numbers. The standalone command-line interface accepts protein FASTA files directly and can use TransDecoder to predict proteins from nucleotide or transcript FASTA files. Lightweight TFLite inference and separately archived, checksum-verified models keep the source repository small and make model retrieval reproducible.

Key Features

Hierarchical four-phase pipeline

• Phase I — Enzyme vs. non-enzyme: binary CNN using the 640-dimensional DPC+NMBroto representation (average 10-fold CV accuracy: 95.70%, MCC: 0.914; independent-test accuracy: 93.43%, MCC: 0.868).
• Phase II — Nitrogen mineralization vs. non-mineralization enzyme: binary CNN using DPC+NMBroto (average 10-fold CV accuracy: 95.23%, MCC: 0.9042; independent-test accuracy: 92.43%, MCC: 0.8487).
• Phase III — Nitrogen mineralization enzyme class: multiclass CNN covering ten enzyme classes using DPC+NMBroto (average 10-fold CV overall accuracy: 96.23%, MCC: 0.8706; independent-test overall accuracy: 93.83%, MCC: 0.7640).
• Phase IV — EC-number assignment: ten class-specific CNNs using the 2,400-dimensional CKSAAP representation, covering 69 specific EC numbers plus wildcard and Others outputs. Independent-test overall accuracy across the reported classifiers ranges from 88.43% to 98.24%.

Modern packaging and portable TFLite inference

• Managed through PEP 621 metadata in pyproject.toml, with a reproducible Python dependency environment in uv.lock.
• Uses TensorFlow Lite (.tflite) models for portable CPU inference; model files are downloaded from Zenodo only when required and verified using file size and SHA-256 checksums.
• Supports protein FASTA directly and nucleotide/transcript FASTA through the external TransDecoder dependency.

Installation

MINpred requires Python 3.9 or newer. Install the package and TensorFlow Lite runtime support with:

pip install '.[tensorflow]'

For a reproducible development environment using the committed uv.lock:

uv sync --extra tensorflow
uv run minpred models-status

The trained models are distributed separately through Zenodo. A required model is downloaded automatically on first use and validated against its published file size and SHA-256 checksum. To download models in advance:

minpred download-models all
minpred models-status

By default, models are stored in the operating system's user cache directory. Set MINPRED_MODEL_DIR to use a specific local model directory.

USU cluster deployment

The maintained cluster launchers mirror the DeepNEC deployment. minpred.sl loads dl-gpu, exposes the TransDecoder 6 installation bundled inside that module, validates the Python runtime, and runs the standalone CLI. The web wrapper run_minpred_web.slurm delegates to the same launcher so web and standalone predictions use one code and model checkout.

On biocluster.usu.edu, prepare the checkout-local dependency layer without replacing the module's TensorFlow installation:

module load dl-gpu
python -m venv --system-site-packages .venv
.venv/bin/python -m pip install --no-deps -e .
.venv/bin/python -m pip install platformdirs
mkdir -p models "$HOME/prediction-jobs/minpred"

The web deployment should use $HOME/naveen_tools/minpred/run_minpred_web.slurm as its remote script and $HOME/prediction-jobs/minpred as its remote job directory.

Prediction

minpred -i example/test.fasta -l Phase1 -od MINpred_results
minpred -i example/test.fasta -l Phase4 -n amidohydrolases -od MINpred_results

MINpred accepts protein FASTA input by default. For transcript or other nucleotide FASTA input, install TransDecoder and select nucleotide mode:

conda install -c conda-forge -c bioconda transdecoder
minpred -i transcripts.fasta --sequence-type nucleotide -l Phase4 \
  -od MINpred_results

Phase IV uses the Phase III enzyme-class prediction to select the matching EC model for each sequence. Use -n CLASS only to restrict Phase IV to one class.

In nucleotide mode, MINpred runs TransDecoder.LongOrfs, preserves its log and the predicted peptide FASTA inside the output directory, removes terminal stop symbols, writes translated_proteins.fasta, and then applies the selected MINpred phases to those peptides. TransDecoder is an external Conda dependency and is therefore not included in uv.lock, which covers the Python environment only. The supplied minpred_environment.yml installs both TransDecoder and the Python runtime.

Phases I–III use the 640-dimensional DPC+NMBroto hybrid representation. Phase IV uses the 2,400-dimensional CKSAAP representation.

Model files

The trained MINpred models were converted to TensorFlow Lite (.tflite) for efficient and portable CPU inference. They are downloaded automatically from Zenodo when required and verified using the file sizes and SHA-256 checksums recorded in minpred/model_manifest.json.

The model files are archived in Zenodo record 21897103 under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

License

The MINpred source code is distributed under the MIT License. The separately archived TFLite model files are distributed under CC BY 4.0. Please cite the Zenodo model record when using the trained models.

Citation

Until the associated journal article is published, cite the model release as:

Duhan, N., Norton, J. M., & Kaundal, R. (2026). MINpred: TFLite models for nitrogen mineralization-related enzyme classification (Version 1.0.0). Zenodo. https://doi.org/10.5281/zenodo.21897103

---

Developed by Naveen Duhan Kaundal Artificial Intelligence & Advanced Bioinformatics Lab (KAABiL) Utah State University, Logan, UT

Technical Queries / Bugs: Naveen Duhan (naveen.duhan@usu.edu)

Scientific Queries: Dr. Rakesh Kaundal (rkaundal@usu.edu)

Lab Website: https://kaabil.net

Released under the terms of the MIT License.