WinnowNet

This is the proteomic SIP branch of WinnowNet, a deep-learning PSM filtering workflow for proteomic stable isotope probing data. This repository currently provides a production pure_cnn_pct prediction wrapper for Sipros5-style raw inputs and training utilities for the current 7-channel CNN feature schema.

Setup

WinnowNet is developed and tested on Linux. Create the micromamba environment used by the scripts:

micromamba create -n winnownet python=3.12 pytorch=2.5.1 numpy pandas scikit-learn einops matplotlib
micromamba activate winnownet

All examples below use:

micromamba run -n winnownet <command>

Input Files

Use Sipros5 to generate the raw WinnowNet input files:

• *_filtered_psms.tsv
• *.FT1
• *.FT2

For each run, keep the files in the same directory with matching stems. For example:

sample_01_filtered_psms.tsv
sample_01.FT1
sample_01.FT2

The TSV must include PSM identifiers (PSMId or SpecId) and peptide sequences. SIP abundance columns such as MS1IsotopicAbundances and MS2IsotopicAbundances are used when present.

Prediction

Run production prediction directly on the Sipros5 output directory or on a comma-separated list of *_filtered_psms.tsv files. script/winnownet.py extracts features, scores PSMs, applies the production threshold logic, and writes accepted target PSMs.

The production wrapper uses 256 matched peaks per PSM.

micromamba run -n winnownet python script/winnownet.py \
  --target data/sample_01 \
  --decoy data/sample_02 \
  -o results/accepted_target_psms.tsv \
  --decoy-output results/accepted_decoy_psms.tsv

In these examples, data/sample_01 is the target sample and data/sample_02 is the unlabeled control/decoy sample.

The default model is:

model/pure_cnn_all_pct.pt

Override it with --model:

micromamba run -n winnownet python script/winnownet.py \
  --target data/sample_01 \
  -o results/accepted_target_psms.tsv \
  --model model/pure_cnn_all_pct.pt

Useful prediction options:

• --target: Target *_filtered_psms.tsv file, directory, glob, or comma-separated list.
• --decoy: Optional unlabeled control/decoy input. When provided, the score threshold is estimated at 1% target-decoy FDR. If no unlabeled control samples are provided, WinnowNet automatically uses the pretrained model's score-threshold calibration.
• --decoy-output: Optional accepted decoy audit TSV; requires --decoy.
• --config: SIP configuration file. Default: script/SIP.cfg.
• --sip-atom-abundance: Optional natural isotope abundance override for no-decoy prediction, for example C13=1.07.
• --device: auto, cpu, cuda, or cuda:N. Default: auto.
• --jobs, --cores, --threads-per-job: Control parallel file scoring and CPU use.
• --batch-size: CNN scoring batch size. Default: 1024.

Training The Current CNN Model

The current production model architecture is pure_cnn_pct. It expects 7-channel CNN features with schema cnn_7ch_v1 and 256 matched peaks per PSM; old 3-channel CNN checkpoints and features are not compatible.

First generate CNN feature pickles from Sipros5 outputs. Directory/list mode automatically pairs each *_filtered_psms.tsv file with matching .FT1 and .FT2 files and writes one .pkl beside each TSV.

micromamba run -n winnownet python script/SpectraFeatures.py \
  -i data/sample_01,data/sample_02 \
  -f cnn \
  -c script/SIP.cfg \
  -t 8 \
  -j 4 \
  --max-peaks 256

For a single run:

micromamba run -n winnownet python script/SpectraFeatures.py \
  -i data/sample_01_filtered_psms.tsv \
  -1 data/sample_01.FT1 \
  -2 data/sample_01.FT2 \
  -o data/sample_01.pkl \
  -f cnn \
  -c script/SIP.cfg \
  -t 8 \
  --max-peaks 256

Then train pure_cnn_pct with target and decoy feature pickles. The --target-pct and --decoy-pct values must align with the target and decoy inputs.

micromamba run -n winnownet python script/WinnowNet_CNN.py \
  --target data/sample_01 \
  --decoy data/sample_02 \
  --target-pct 5 \
  --decoy-pct 1 \
  -m model/pure_cnn_all_pct.pt \
  --model-arch pure_cnn_pct \
  --epochs 50 \
  --learning-rate 2e-4 \
  --train-batch-size 1024 \
  --eval-batch-size 1024 \
  --class-weight none \
  --pct-loss-weight 2.0 \
  --exclude-protein-prefix Con_

Training writes the best model to the path passed with -m and per-epoch checkpoints to a sibling directory named after the model, for example pure_cnn_all_pct_checkpoints/.

Optional filters:

• --target-exclude: Per-target-input strict lower bound for MS2IsotopicAbundances; use 0 to disable for an input.
• --decoy-exclude: Per-decoy-input strict upper bound for MS2IsotopicAbundances; use 0 to disable for an input.
• -p: Load a compatible pretrained 7-channel checkpoint before training.

For a grid search over pure_cnn_pct hyperparameters:

micromamba run -n winnownet python test/grid_search_pure_cnn_pct.py

References

1. Feng, S., Zhang, B., Wang, H., Xiong, Y., Tian, A., Yuan, X., Pan, C., and Guo, X. Enhancing peptide identification in metaproteomics through curriculum learning in deep learning. Nature Communications 16, 8934 (2025). https://doi.org/10.1038/s41467-025-63977-z
2. Xiong, Y., Mueller, R. S., Feng, S., Guo, X., and Pan, C. Proteomic stable isotope probing with an upgraded Sipros algorithm for improved identification and quantification of isotopically labeled proteins. Microbiome 12, 148 (2024). https://doi.org/10.1186/s40168-024-01866-1
3. Sipros5 is used to generate the *_filtered_psms.tsv, .FT1, and .FT2 inputs consumed by WinnowNet.