Gaolaboratory/iona-base-100m

Iona

Iona is a transformer encoder foundation model for tandem mass spectra (MS/MS). It treats each centroided peak as a token and learns how peaks relate to each other through a per-head attention bias over the signed m/z difference (Δm/z) between every pair of peaks.

Model details

• Input: a centroided spectrum given as parallel mz and intensity arrays, with at most 512 peaks.
• Peak tokens: built from the normalized log intensity only. Tokens carry no absolute m/z; the model sees mass only through pairwise Δm/z.
• Δm/z attention bias: each attention head gets a learned bias curve over Δm/z. The curve is computed by a small MLP on Fourier features of Δm/z (256 frequencies, 0.001–190).
• Output: one contextual embedding per peak (last_hidden_state).

Size	Hidden size	Layers	Heads	FFN size
iona-base-25m	512	8	8	2048
iona-base-50m	640	10	10	2560
iona-base-100m (this model)	800	13	10	3200
iona-base-200m	1024	16	16	4096
iona-base-400m	1280	20	20	5120

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 AutoModelForPreTraining, AutoProcessor

repo = "path/to/iona-base-100m"  # this repository's Hub ID, or a local copy
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForPreTraining.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():
    peak_embeddings = model.iona(**inputs).last_hidden_state  # (batch, peaks, hidden_size)

The processor normalizes intensities, so pass raw intensity values. For a single embedding per spectrum, pool peak_embeddings over the peaks where inputs["attention_mask"] is 1.

Training

• Data: about 100 million consensus tandem mass spectra.
• Objective: masked-intensity modeling. 50% of peaks have their intensity token replaced by a learned mask token. The model predicts how intensity is distributed across the masked peaks, trained with a KL-divergence loss against the true normalized intensities.
• Optimization: AdamW (β₁ = 0.9, β₂ = 0.95), learning rate 1.3e-4 with 2,000 warmup steps and cosine decay, weight decay 0.01, bf16, 3 epochs.

Evaluation

Limitations

• Inputs must be centroided spectra with between 1 and 512 peaks. Profile-mode data needs to be centroided first.
• The model was pretrained on a single collection of consensus spectra. Performance on other instruments, fragmentation methods or sample types hasn't been characterized.