henhaobio
Sign in

bioRxivBioinformaticsPreprint

Prioritizing peptides for targeted mass spectrometry experiments using deep learning

Shreyash SonthaliaiD, Bo WeniD, Pooja Dasgupta, Chris HsuiD, Michael J. MacCossiD, William Stafford NobleiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

One critical step in any targeted mass spectrometry experiment is selecting, from each protein of interest, a small number of peptides that respond well in the mass spectrometer and can serve as reliable proxies for protein quantification. Existing methods select target peptides either by relying on prior empirical measurements, limiting their applicability to previously observed peptides, or using machine learning to predict peptide behavior from sequence alone. However, current machine learning tools suffer from various limitations, including using detectability as an indirect proxy for intensity, relying on small training sets, or ignoring the precursor charge state. In this study, we introduce Bromo, a transformer-based deep learning model that ranks peptide precursors from a given protein by their relative response, taking charge state into account. Trained on millions of annotated peptide pairs derived from large-scale, publicly available data-independent acquisition mass spectrometry data, Bromo consistently outperforms existing sequence-based methods across diverse, independent datasets. Furthermore, we show that fine-tuning Bromo on experiment-specific data can account for differences in sample preparation, sample matrix, and instrument platform, all of which influence which peptides serve as optimal targets. This adaptability makes Bromo a practical tool for selecting target peptides for selected reaction monitoring and parallel reaction monitoring assay development across a wide range of experimental conditions.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (31 Aug 2026) · CC BY 4.0

Bromo, a pairwise transformer, ranks a protein's peptide precursors by expected mass spectrometry response, beating XGBoost, PREGO and PeptideRanger on held-out yeast data, and can be fine-tuned per instrument.

Targeted proteomics assays need a few high-responding peptides per protein; empirical selection cannot cover unobserved proteins, and existing sequence-based predictors use detectability as a proxy for intensity, train on small sets, or ignore charge state. Bromo instead learns which of two precursors (peptide plus charge) from the same protein gives more signal, trained on millions of pairs labelled from public data-independent acquisition (DIA) runs.

Key results

  1. ShownOn held-out yeast-astral data, Bromo's peptide-level rankings pick more of the truly top-responding peptides than PREGO and PeptideRanger. PeptideRanger was trained to predict general detectability, not response.

    Bromo was pretrained on human-astral, and the comparison was limited to peptides that every tool scored. Its top-k accuracy (TKA, the share of predicted top-k peptides that are in the observed top-k) at k = 3 was significantly higher than PREGO's (paired t-test, p = 6×10^−200) and PeptideRanger's (p = 1.95 × 10^−289). Yeast-exploris and yeast-lumos are plotted, but the text gives no statistics for them. · 3.2 Bromo outperforms existing methods for…; 2.7 Performance measures; 2.8 Baseline methods

  2. ShownBromo beats an XGBoost classifier trained on the same pair task and data splits. The gap is widest for protein groups with many candidate peptides in the Exploris and Lumos sets.

    TKA was significantly higher on yeast-astral, yeast-lumos and yeast-exploris. XGBoost used amino-acid composition plus N- and C-terminal residue indicators (a 128-dimensional pair vector) and was tuned over 10 Optuna trials. The text gives no TKA values or p-values here, and the authors say the comparison mixes architecture with input representation. · 3.2 Bromo outperforms existing methods for…; 2.8 Baseline methods

  3. ShownPeptide rankings agree closely between replicate runs on one instrument but much less across instrument platforms. Response therefore depends on the instrument as well as on sequence.

    In four human datasets (Orbitrap Astral, Orbitrap Fusion Lumos, Exploris 480, SCIEX 6600 TripleTOF), TKA at k = 3 was 0.850 to 0.930 between technical replicates on the same platform and 0.528 to 0.745 between different platforms. · 3.3 Instrument-specific fine-tuning mitigates…; 2.9 Peptide ranking comparison across…

  4. ShownFine-tuning the human-astral model on data from the target instrument was never worse than the pretrained model or a model trained from scratch. The largest gain reported in the text was on SCIEX TripleTOF data (human-pan).

    On human-pan, TKA at k = 3 rose from 0.501 to 0.585 (+0.0835; paired t-test, p = 3.4 ×10^−21). Significant but small gains were 0.0247 (human-exploris) and 0.0255 (human-lumos). Against training from scratch, fine-tuning was better or comparable on most datasets; human-lumos showed no such benefit. · 3.3 Instrument-specific fine-tuning mitigates…; 2.6 Model fine-tuning

  5. ShownBromo's ranking accuracy keeps rising as training data grow, while the XGBoost baseline plateaus early.

    Mean TKA at k = 3 was plotted against training set size on human-astral and human-pan, with standard deviation over 5 random subsamples of the training set. The text gives no numbers for this analysis. · 3.4 Bromo’s performance scales with increasing…; Figure legends

Methods, data and reproducibility

Twin weight-sharing transformer encoders embed two tryptic peptides (up to 30 residues) from one protein, join them with charge-state embeddings, and classify which precursor has the higher DIA-NN normalized MS2 intensity; pairwise outputs are tallied into a within-protein ranking. Pair labels come from per-run intensity comparisons, combined across runs by majority vote or a binomial test.

  • human-astral: human DIA dataset (Orbitrap Astral) used for pretraining; protein groups split 0.6/0.2/0.2 into train/validation/test · human · PXD056793
  • yeast-astral, yeast-lumos, yeast-exploris: yeast DIA from the same study (Orbitrap Astral, Fusion Lumos, Exploris 480), used for testing and fine-tuning · yeast · PXD056793
  • human-lumos and human-exploris: human DIA from the same study on other instruments, used for fine-tuning · human · PXD056793
  • human-pan: cancer cell lines across 28 tissue types on SCIEX 6600 TripleTOF; DIA-NN results downloaded from PRIDE; used for fine-tuning and label-scheme analysis · human (cancer cell lines) · n = 949 cell lines; 6864 DIA runs · PXD030304
Methods & tools
Bromo, PyTorch, Optuna 4.8.0, DIA-NN 1.8.1, XGBoost, PREGO, PeptideRanger, Orbitrap Astral / Fusion Lumos / Exploris 480, SCIEX 6600 TripleTOF, PRIDE / ProteomeXchange, PanoramaWeb, Zenodo
Software & environment
Optuna 4.8.0 and DIA-NN 1.8.1; PyTorch with a fixed seed (default 42, --seed) and deterministic cuDNN settings; Bromo released under Apache 2.0. No container or requirements file is mentioned.
Assay conditions & compute
Final model trained on a single NVIDIA L40 GPU for 15 epochs at batch size 4096, taking about 45 hours; inference runtime is reported in Supplementary Figure 3.
Reproducibility gaps
Search results, processed data and model checkpoints are on Zenodo, but the link is given only inside the bromo-manuscript repository, not in the text. Versions of PyTorch and other dependencies are not stated.
Limitations (6)
  • Bromo-selected peptides were never used to build and test a selected or parallel reaction monitoring (SRM/PRM) assay, so the evaluation is against DIA-measured rankings only. Transitions, retention times and scheduling still need other tools. · stated by the authors
  • Training used tryptic peptides with DIA MS2 signals at fixed normalized collision energy. Rankings may transfer less well to other proteases, unusual modifications or charge states, enrichment or depletion workflows, poorly represented matrices, or collision-energy-optimized targeted methods. · stated by the authors
  • Labels inherit DIA-NN artefacts (missed detections, interference, protein inference, normalization). They can also be wrong when a peptide's abundance differs from that of its assigned protein, for example through other isoforms, unlisted homologs or high-stoichiometry modifications. · stated by the authors
  • Same-instrument replicates reach only 0.850 to 0.930 TKA at k = 3, so the gold standard is noisy and achievable model accuracy is capped below 1.0. Datasets with few replicates give noisier labels. · stated by the authors
  • The baseline comparisons use yeast test sets from the same study and LC-MS/MS conditions as the human pretraining data. No comparison with PREGO or PeptideRanger is reported on data from an independent study. · noted in this summary
  • The text gives no TKA values for any baseline comparison, and p-values only for PREGO and PeptideRanger on yeast-astral (the XGBoost comparison is reported only as 'significantly higher'), so effect sizes can only be read from Figure 2. · noted in this summary

Read this if you design SRM or PRM assays and want a charge-aware, sequence-based peptide ranker that you can fine-tune on DIA data from your own instrument.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISXue '25Lapin '25Zumer '23Burq '24Wichmann '19Dincer '22
68% semantic matchBioinformatics
Liang Xue, Shivani Tiwary, Mykola Bordyuh+2 more·PROTEOMICS·Jun 30, 2025· 1 citations