Bromo

A targeted proteomics assay lives or dies on a handful of peptides: for each protein, an analyst picks two or three tryptic peptides to monitor by selected reaction monitoring (SRM) or parallel reaction monitoring (PRM), and they must give reproducible signal tracking the protein's abundance. The classical way to choose them is to measure them first — fine for proteins somebody has already studied, useless for the rest. Sequence-based predictors were meant to close that gap, but most learn detectability — whether a peptide is seen at all. That says nothing about which of five observable peptides gives the largest signal, and it ignores charge state, even though one sequence at 2+ and 3+ can ionize and fragment quite differently.

Bromo, from the Noble and MacCoss labs at the University of Washington, reframes the problem as ranking rather than regression. It takes an ordered pair of precursors — two peptide sequences with their charge states, drawn from the same protein — and predicts which will yield the larger MS2 response in a data-independent acquisition (DIA) experiment. Because peptides from one protein share its starting abundance, differences in their observed intensity are attributable to the peptides themselves, so within-protein pairs supply supervision straight from complex biological samples — no equimolar synthetic standards needed. Counting positive labels over every pair induces the full ranking an assay designer needs. This entry describes version 2 of the bioRxiv preprint, posted 31 August 2026, which added baselines and reworked fine-tuning.

#Key Features

• Pairwise ranking objective: Binary classification over precursor pairs from one protein, matched to the assay-design decision rather than to absolute intensity regression.
• Charge state as a first-class input: A learned charge embedding accompanies each precursor, so 2+, 3+ and 4+ forms of a sequence are ranked separately, not collapsed.
• Trained on measured response: Labels come from DIA-NN normalized precursor intensities in real samples, not detectability annotations or synthetic peptide panels.
• Consensus labels across runs: A binomial test over every run detecting both precursors fixes each label; over 98% of pairs seen in five or more runs kept it across 20 resamplings.
• Instrument-specific fine-tuning: Continued training adapts the pretrained checkpoint to a new sample matrix or platform, and never performed worse than training from scratch.

#Technical Details

Bromo is a Siamese transformer: twin shared-weight encoders, each with a 256-dimensional embedding, positional encoding, and four transformer layers of eight attention heads and feed-forward dimension 512. Learned attention pooling collapses each precursor to one 256-dimensional vector; the pair is concatenated with two 16-dimensional charge embeddings into a 544-dimensional representation and classified by fully connected layers of width 128 and 2. Training uses cross-entropy with label smoothing and AdamW under an Optuna-tuned configuration; the final run took roughly 45 hours on one NVIDIA L40 GPU. Inputs are tryptic peptides of length 7–30 at charges 2–4.

Seven DIA datasets spanning human and yeast samples on four platforms — Orbitrap Astral, Orbitrap Fusion Lumos, Exploris 480 and SCIEX 6600 TripleTOF — underpin the work. Pretraining used human-astral (8,070 protein groups, 443,778 precursors, 18.4 million precursor pairs); the largest, human-pan, spans 6,864 runs from 949 cancer cell lines across 28 tissue types. Performance is top-k accuracy (TKA) against the observed intensity ranking. On yeast test sets from three instruments Bromo beat an XGBoost baseline trained on identical pairs, PREGO, and PeptideRanger at k = 3. Technical replicates on one instrument agree at TKA 0.850–0.930, an empirical ceiling; agreement across instruments falls to 0.528–0.745, and fine-tuning on SCIEX TripleTOF data lifted TKA at k = 3 from 0.501 to 0.585.

#Applications

The intended user is a proteomics group building a quantitative assay for proteins never measured on their instrument. Starting from a FASTA file, the command-line package performs in-silico digestion and ionization, scores every within-protein precursor pair, and returns a ranked shortlist to carry into SRM or PRM method development for biomarker verification and clinical assays. Groups with their own DIA data can fine-tune the released checkpoint to their sample preparation and platform.

#Impact

Bromo argues that public DIA repositories are a better foundation for peptide selection than the small synthetic panels earlier tools used, and its training-size curves agree: accuracy keeps climbing with more peptide pairs while the XGBoost baseline plateaus. Code is Apache 2.0 and the checkpoints sit on Zenodo, so results are reproducible end to end. The limits are stated plainly: Bromo covers tryptic peptides and DIA-derived labels, inherits DIA-NN's detection and inference errors, and was trained at fixed normalized collision energies, so it does not model the collision-energy optimization done during method development. No SRM or PRM assay was experimentally validated from Bromo-selected peptides, so the evidence establishes ranking quality, not end-to-end assay performance; a complete method still needs transition selection, scheduling, and retention-time prediction from tools such as AutoRT.