DDA-BERT

A data-dependent acquisition search engine hands its peptide-spectrum matches (PSMs) to a rescoring step, and that step almost always fits a shallow classifier — Percolator's support vector machine, mokapot inside MS²Rescore — on the targets and decoys of the one file or project in front of it. Deep learning enters only upstream, as predicted retention times and fragment intensities converted into features for that classifier. Whatever the decision stage learns is discarded when the next dataset loads, and its ceiling is set by how many PSMs that single dataset happens to contain.

DDA-BERT moves the decision itself into one transformer trained once across projects. Two encoders meet through cross-attention: a spectrum encoder over the m/z and intensity of MS2 peaks, and a peptide precursor encoder over the candidate sequence together with precursor m/z, charge, predicted retention time and ΔRT. There are no handcrafted correlation features and no separate representation-learning stage. The result ships as a fixed checkpoint that scores new files in inference mode, so nothing is refit when the instrument, species or sample type changes.

DDA-BERT comes from Tiannan Guo's group at Westlake University, with Westlake Omics, and was published in Nature Communications after a bioRxiv preprint. It is the data-dependent counterpart to DIA-BERT from the same lab, which scores peak groups in data-independent runs, and sits alongside DIA-CLIP among pretrained proteomics models.

#Key Features

• Cross-project training corpus: 12,285 LC–MS/MS runs from 110 independent projects, spanning 11 species and 10 mass spectrometer platforms, for roughly 271 million PSMs.
• Dual-modality corruption under a joint objective: 10% of peaks are dropped from each spectrum while one to three contiguous spans covering about 20% of peptide tokens are masked, and masked amino acid reconstruction is optimized jointly with PSM quality estimation — so the model must rely on peptide-spectrum correspondence, not memorized sequence.
• Iterative hard negative mining: high-scoring decoys from earlier training rounds, per-file shuffled sequences, and "false targets" — real peptide sequences that were not the top match for their scan — are fed back as negatives.
• Search-engine flexibility: rescores Sage output alone, or Sage combined with FragPipe (MSBooster) and AlphaPept, and runs either as a one-stop pipeline through protein inference or as a rescoring-only module.

#Technical Details

The architecture is nine transformer layers with an embedding dimension of 768 and sixteen attention heads, totalling approximately 95 million parameters. Peak m/z and intensity are encoded with sinusoidal position embeddings and learned linear layers; learned embeddings carry precursor m/z, charge, amino acid identity and position, with no causal mask since the peptide encoder is not autoregressive. Training ran for 30 epochs on 16 NVIDIA A100 GPUs over roughly seven days, at a batch size of 128 spectra and a peak learning rate of 5 × 10⁻⁵ with a 150,000-step linear warm-up followed by cosine decay. 190 raw files from an HLA immunopeptidomics dataset were held out for validation; a further 450 files from the same deposit were added to training, with peptide sequences present in the validation set stripped from the training data to prevent overlap.

At a 1% FDR threshold DDA-BERT identified 40,105 unique peptides on human FFPE colorectal tissue — 2.24% above FragPipe with MSBooster and 269.35% above AlphaPept — plus 19,797 in S. cerevisiae, 15,141 in A. thaliana and 12,834 in D. melanogaster, gaining 3.73–141.46%, 3.68–62.77% and 5.53–45.64% over the compared tools. On HLA class I immunopeptidomics it added 4.14–87.47% more peptides, and at 0.1 ng of HeLa digest, roughly 0.4 cells, it reported 432 PSMs. Entrapment analysis with FDRBench placed the peptide-level false discovery proportion at 0.88–0.99% against a nominal 1% threshold.

#Applications

DDA-BERT targets bottom-up proteomics groups whose limiting factor is identification depth: tissue cohorts, trace-level and single-cell samples, immunopeptidomics, and non-model organisms with too few matches to train a project-specific rescorer. It reads mzML and Bruker .d directly, with Sciex .wiff requiring prior conversion, and ships as a portable Linux executable, a pip-installable Python package and a Docker image, needing an NVIDIA GPU with bfloat16 support and about 20 GB of GPU memory.

#Impact

DDA-BERT's argument is that rescoring has been treated as a per-dataset learning problem when it can be a cross-project one, and that the shallow final classifier — not the feature predictors ahead of it — is what caps the framework. The authors are explicit that it complements per-file and per-project workflows rather than replacing them, since those remain cheap, CPU-friendly and well calibrated on large homogeneous studies. The constraints are real: a complete workflow on a file of about 23,000 spectra takes roughly 29 minutes on a single A100; scoring confidence falls for peptides longer than about 35 amino acids; predicted fragment m/z and intensity are not used as inputs the way MSBooster-style tools use them; and results remain bounded by the candidate PSMs the upstream search engine produced. The code, the shipped weights and the executable carry a custom license permitting non-commercial use only.