iona-rerank-400m

iona-rerank-400m is a pretrained rescorer for peptide-spectrum matches (PSMs). It scores each database-search candidate from standard PSM features plus features of the Iona spectrum–peptide embedding similarity, so a run can be rescored without training on that run.

A rescorer trained on each run individually gives substantially more identifications (see Evaluation). Use this global model when per-run training isn't possible, for example when a run has very few spectra.

Model details

• Architecture: an MLP with 259 inputs, two hidden layers of 64 units and one output score.
• Inputs: a PSM feature table plus five embedding features: the cosine similarity between the spectrum and peptide embeddings, its rank within the spectrum, its difference from the best other candidate, its z-score, and the gap between the top two candidates.
• Files: model.safetensors holds the MLP weights. rescorer.json lists the 259 input feature names, the per-feature medians used to fill missing columns (0 for modification counts), and the mean and standard deviation used to standardize the inputs.

Usage

Use the scripts in the Iona code repository:

# 1. Embed every candidate of a run (GPU recommended).
python scripts/rerank_psm_embed.py --run my_run.parquet --out rows/my_run.parquet \
    --encoder path/to/iona-contrastive-400m \
    --student path/to/iona-peptide-embedder-400m

# 2. Score the candidates with this model.
python scripts/psm_rerank.py score --mode global --model path/to/iona-rerank-400m \
    --rows rows/ --labfeat features/ --out psms.parquet

The embedding features must come from iona-contrastive-400m (spectra) and iona-peptide-embedder-400m (peptides), the pair this model was trained with. To train and apply a per-run rescorer instead, run python scripts/psm_rerank.py score --mode perrun ... (the default mode).

Training

• Data: 8 runs (6 HEK293, 2 HCT116) with the top 10 MSFragger candidates per spectrum.
• Labels: target–decoy labels with iterative relabeling, seed 0.

Evaluation

Accepted PSMs at 1% FDR on 16 held-out HCT116 runs (1,867,364 spectra):

Method	PSMs at 1% FDR
MSFragger e-value	329,807
iona-rerank-400m (global)	411,607
Per-run rescorer, PSM features	515,674
Per-run rescorer, PSM + embedding features	523,720

Limitations

• Trained on HEK293 and HCT116 runs searched with MSFragger. Other search engines, cell types and instruments haven't been evaluated.
• The input features must be computed the same way as in training, including the embedding models above.