PepPr

Resolving spectral ambiguity in de novo peptide sequencing using peptide priors.

Code for the paper's benchmarks and figures lives in gagneurlab/PepPr_paper.

Install

conda create --name peppr -y python=3.11
conda activate peppr
pip install -r requirements.txt
pip install torch==2.7.0 --index-url https://download.pytorch.org/whl/cu128
pip install -e .

# Patch Casanovo with the PepPr hooks.
git submodule update --init
cd casanovo
git apply ../casanovo_integration.patch
pip install -e . --config-settings editable_mode=compat
cd ..

editable_mode=compat is required — the default strict editable install resolves Casanovo's nested casanovo/ package as a namespace and breaks import casanovo.

A ContraNovo integration is included too: cd ContraNovo && git apply ../contranovo_integration.patch.

Inference

Download and extract the model archive from Zenodo (DOI filled in on publication), then point at the checkpoints inside it:

tar xzf peppr_models.tar.gz
export PEPPR_MODELS="$PWD/peppr_models"

PEPPR_PRIOR_PATH="$PEPPR_MODELS/models/human/prior_model.pt" \
PEPPR_FUSION_PATH="$PEPPR_MODELS/models/human/fusion_model.pt" \
casanovo sequence \
    -m https://github.com/Noble-Lab/casanovo/releases/download/v5.0.0/casanovo_v5_0_0.ckpt \
    -c casanovo/casanovo/config.yaml \
    --teacher_forcing false --use_peppr true \
    -d output_dir -o results \
    -f spectra.mgf

The archive holds one prior per species under models/<species>/prior_model.pt; swap PEPPR_PRIOR_PATH to use a different one, and the same fusion head still applies. --use_peppr false gives the plain Casanovo baseline.

Train

Priors and fusion heads are trained separately and don't need to be retrained together. A fusion head fuses a prior's scores with one specific backbone's score distribution (PEPPR_BACKBONE, e.g. Casanovo or ContraNovo), so it's reusable across any prior trained for that backbone: swap species by pointing PEPPR_PRIOR_PATH at a different prior and keep the same fusion head. Only changing the backbone itself requires training a new fusion head, since that changes the score distribution the fusion head was fit to.

Training writes its intermediates into the current directory. Set PEPPR_WORK_DIR to write elsewhere.

Train a new prior

Reads the proteome from PEPPR_FASTA. No data archive is involved.

export PEPPR_FASTA=/path/to/proteome.fasta

# Digest the proteome, then train.
python peppr/prepare_data.py
PEPPR_PRIOR_PATH=/path/to/prior.pt python peppr/train_peptide_prior_model.py

Train a new fusion head

Needed only when switching PEPPR_BACKBONE; a fusion head trained for one backbone works with every prior trained for that same backbone.

# Backbone teacher scores + prior teacher scores over one corpus.
# PEPPR_BACKBONE selects which backbone's teacher scores to compute
# ('casanovo', the default, or 'contranovo').
PEPPR_FUSION_PATH=/path/to/fusion.pth python peppr/train_fusion_head.py

The part worth relocating deliberately is fusion/, holding the backbone teacher scores and fusion targets: those depend only on the training corpus, not on the prior, so point PEPPR_FUSION_WORK_DIR at a shared location to avoid recomputing them for every prior.

Environment variables

Variable	Meaning	Default
PEPPR_PRIOR_PATH	filepath of prior model	required
PEPPR_FUSION_PATH	filepath of fusion model	required

Inference needs only these two. Anything that genuinely requires a training directory fails with a message naming the variable to set.