DIA-BERT

Conventional data-independent acquisition (DIA) search engines train their scoring model from scratch on the file in front of them. A single run yields only enough target and decoy peak groups to fit a shallow classifier before overfitting sets in, and whatever that classifier learns about how a genuine peak group looks is thrown away when the next file loads. DIA-BERT breaks that loop by pretraining one encoder-only transformer across 952 human DIA-MS files and reusing it everywhere.

The model is end to end. For each peptide precursor it takes the raw extracted ion traces together with the spectral library's fragment intensities and m/z values, and learns the representation and the score under a single objective — no hand-engineered correlation features, and no separate representation-learning stage feeding a gradient-boosted re-scorer, the arrangement DreamDIA uses. Because the pretrained scorer already knows what a real peak group looks like across hundreds of runs, adapting it to a new file is a short fine-tuning pass that need only absorb that run's retention-time drift and fragment-intensity idiosyncrasies.

DIA-BERT comes from Tiannan Guo's group at Westlake University, with Westlake Omics, and was published in Nature Communications after a bioRxiv preprint. It sits alongside DIA-CLIP among pretrained models built for data-independent acquisition proteomics.

#Key Features

• Pretraining at corpus scale: The pre-filtering model was trained on 276 million peptide-precursor instances and the scoring model on 318 million, drawn from 952 DIA-MS files covering varied human specimens.
• End-to-end peak-group scoring: Ion traces and library metadata go straight into the network, which jointly optimizes representation and score instead of handing crafted features to a downstream classifier.
• Per-run adaptation from a shipped checkpoint: The released scorer is fine-tuned on each file's own targets and decoys, split five ways to monitor validation loss, then re-scores peak groups at a 1% FDR threshold.
• Learned peak-area estimation: A second transformer regresses the summed peak area directly, pretrained on 34 million precursors from simulated DIA runs with known true areas.
• Cross-run normalization built in: Retention-time-binned precursor normalization feeds MaxLFQ protein quantification, whose output then adjusts the precursor values back.

#Technical Details

Each precursor is represented as a 330 × 16 matrix: MS1 traces for the precursor m/z and five isotopic peaks, plus MS2 traces for library fragments with their heavy and light isotopes, each extracted over a ±20 ppm window across 16 consecutive spectra centred on the expected retention time. Library attributes become numerical embeddings concatenated onto that matrix, which passes through two convolutional blocks (convolution, average pooling, layer normalization) and eight self-attention blocks before collapsing to a 1-D tensor. Run-level attributes such as gradient and instrument type and precursor-level attributes such as iRT, delta-RT, m/z, charge and sequence length are embedded separately, concatenated with that tensor, and mapped through two fully connected layers under a binary cross-entropy loss. The quantification model mirrors the architecture with a regression loss; its synthetic pretraining corpus was built with EncyclopeDIA and Prosit from the reviewed human proteome, perturbed with retention-time shifts, precursor interference and intensity rescaling, then fine-tuned on 36 human cell-line DIA files.

Against DIA-NN 1.9.2 on five human cancer tissue sets of three samples each, DIA-BERT identified 22% more peptide precursors and 51% more proteins in library-based mode, and 56% more precursors and 73% more proteins in library-free mode, with conservative two-species FDR estimates below 0.01 throughout. On a three-species human, yeast and C. elegans benchmark the margins narrow to 4% and 6%. Pretraining on 1, 35, 106 and 317 of the 952 files showed identifications rising approximately linearly with the log₁₀ of corpus size.

#Applications

DIA-BERT is aimed at bottom-up proteomics groups running library-based or library-free DIA on Orbitrap instruments — tissue cohorts, cell-line panels, and any experiment whose limiting factor is depth at the low-abundance end. The precursors and proteins it recovers beyond DIA-NN are significantly less abundant than the ones both tools find, so the practical gain lands near the detection limit. It ships as an install-free Windows and Linux application with the three checkpoints bundled, alongside the Python source, and consumes a standard spectral library that must be built with external tools.

#Impact

DIA-BERT is a working argument that DIA identification is a pretrain-then-adapt problem rather than a per-file learning problem, and the scaling curve across 1 to 952 pretraining files is the evidence: capability tracks corpus size. The constraints are real. Version 1.0 was trained on Orbitrap data and the authors do not recommend it for timsTOF or tripleTOF acquisitions; training used fully tryptic peptides only, leaving semi-tryptic peptides and broader modifications unevaluated; a CUDA GPU with tens of gigabytes of memory is needed in practice, and per-file fine-tuning means analysis cost scales with the number of runs. The code and the bundled weights carry a custom license permitting academic use and prohibiting commercial use.