Compare MLPs on antimicrobial peptides

The existing ABP_GN notebook compares Fourier, circular Z, finite Mellin, and Haar wavelet magnitudes of corrected BLOSUM100 on Gram-negative antibacterial peptide classification. It preserves 3,562 supplied training peptides and 1,973 test peptides, with the previous seed-7 stratified training validation split. The raw CSVs are local inputs and are not redistributed.

See also

Download the editable AMP notebook and read the repository’s research/README.md for execution instructions.

Run the comparison

Install .[ml,wavelets,validation] and notebook dependencies. Set TRAIN_CSV and TEST_CSV in the controls cell, then run all cells. Defaults look first in research/data/, then ~/Downloads/input_csv/. Required columns are Sequence and Class; classes are Positive and Negative.

python -m research.execute_amp_notebook \
  --output validation/amp_transform_mlp/my_comparison.executed.ipynb

Choose a fresh output path. The notebook saves fresh timestamped results under ignored validation/amp_transform_mlp/: metrics, every validation trial, row-indexed test probabilities, split indices, feature names, source/input hashes, and environment versions. It plots ROC, precision-recall, and training-loss curves.

What is controlled

Defaults use 20 channels, a 256-position grid, 5,120 magnitudes per peptide, one 64-unit ReLU MLP candidate, and one model seed. All four models therefore have 327,809 parameters. StandardScaler fits only the fitting partition. Validation selects candidates; the winner is refit on the full supplied training set. The MLP’s internal early stopping uses training-only data. Threshold stays at 0.5.

This full-coefficient policy differs from the historical dyadic Fourier audit. Adding candidates or transform settings changes the search budget. Adding model seeds measures fitting variability, not dataset or family-level uncertainty. Wavelet component='real' is a separate experiment because magnitudes lose signs.

Recorded exploratory baseline

One seed, existing previously studied test set

Transform

Test balanced accuracy

MCC

ROC-AUC

Average precision

Fourier

76.44%

0.529

0.834

0.749

Z

76.38%

0.525

0.845

0.782

Mellin

74.66%

0.488

0.835

0.787

Wavelet

74.72%

0.493

0.818

0.719

These are recorded runs, not recomputed by building the documentation. The API unification replay reproduced all test probabilities and metrics. The supplied test set has already been used in earlier experiments; small score differences do not establish superiority. A confirmatory study needs repeated seeds and a family-disjoint or new external test set, with policy fixed before evaluation.