iona-denoise-200m

iona-denoise-200m scores every peak of a tandem mass spectrum (MS/MS) as signal or noise. It is the Iona 200m encoder with a per-peak classification head, fine-tuned for noise-peak detection.

Model details

• Architecture: Iona encoder (16 layers, hidden size 1024, 16 heads) plus a two-layer classification head (hidden size 512). 203.4M parameters in total.
• Input: a centroided spectrum given as parallel mz and intensity arrays, with at most 512 peaks.
• Output: one logit per peak. Noise is the positive class, so sigmoid(logit) is the predicted probability that a peak is noise.

Usage

The checkpoint includes its modeling code, so it loads with trust_remote_code=True. You need torch, transformers, sentence-transformers and pytorch-metric-learning installed.

import torch
from transformers import AutoModelForTokenClassification, AutoProcessor

repo = "path/to/iona-denoise-200m"  # this repository's Hub ID, or a local copy
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForTokenClassification.from_pretrained(repo, trust_remote_code=True).eval()

mz = [175.119, 262.151, 363.198, 476.282]
intensity = [1200.0, 5300.0, 800.0, 2400.0]
inputs = processor(mz, intensity, return_tensors="pt")

with torch.no_grad():
    noise_prob = torch.sigmoid(model(**inputs).logits.float())  # (batch, peaks)
keep = (noise_prob < 0.5) & inputs["attention_mask"].bool()

Pass raw intensity values; the processor normalizes them. To denoise whole mzML files, use the iona-denoise command from the Iona package.

Training

• Pretraining: masked-intensity modeling on about 100 million consensus tandem mass spectra for 540,423 steps (see iona-base-200m).
• Fine-tuning: per-peak noise classification on a dataset of MS/MS spectra with per-peak noise labels, 4 epochs, learning rate 2e-4 (0.5× for the encoder) with cosine decay, effective batch size 12, bf16, seed 1.
• Checkpoint selection: the best of 3 seeds across the pretraining checkpoints available at this scale, by validation AUPRC. The test split wasn't used for selection.

Evaluation

On the validation and test splits of the noise-labeled dataset. The splits share no peptides with the training split or with each other.

Metric	Validation	Test
AUPRC	0.9526	0.9524
AUROC	0.9459	0.9459
F1 (logit ≥ 0)	0.8803	0.8793
Per-spectrum AUROC (mean)	–	0.9508

Limitations

• Inputs must be centroided spectra with between 1 and 512 peaks. Longer spectra were excluded from training and evaluation.
• The model was fine-tuned and evaluated on one dataset. Performance on other instruments, fragmentation methods or sample types hasn't been characterized.