Loading…
Peptide identification for diaPASEF proteomics, scoring fragment coelution across retention time and ion mobility with a pretrained CNN.
No providers recorded yet. Browse all providers
A peptide is declared present in a data-independent acquisition run when the elution profiles of its fragment ions rise and fall together. Judging how well they agree is the most consequential call the search software makes, and it has always been made by a hand-picked formula — a dot product, a cross-correlation, a cosine similarity, an entropy function. OpenSWATH defined 23 such scoring functions, MaxDIA 60, DIA-NN 73. Each encodes a human guess at what coelution looks like, and each degrades once signals shift, blur, or drop out.
Beta-DIA replaces that guess for the hardest feature with a small convolutional network, DeepProfile, pretrained once and shipped inside the package. diaPASEF data on timsTOF instruments carries ion mobility alongside retention time, so an elution group is not a set of curves but a 2D image: fragment intensity binned over acquisition cycles and over mobility. DeepProfile reads that image and returns a coelution probability. The learned scores do not displace the classical ones: they join them in a single target-decoy discriminator, and the paper's ablations find the two families complementary.
Beta-DIA comes from Jian Song and Xiaohui Wu at Soochow University, with Hebin Liu and Chengpin Shen of Shanghai Omicsolution, whose instrument runs supplied the pretraining corpus. It belongs to a small group of DIA search tools built around a pretrained scorer, alongside DIA-BERT and DIA-CLIP, aimed squarely at trapped-ion-mobility data.
DeepProfile is a compact CNN. The input map is normalized twice — once per ion, once globally — and each version passes through two rounds of convolution, ReLU, and max pooling into a 16-dimensional vector. Those two vectors are concatenated with a 16-dimensional embedding of the number of valid fragment ions, and the resulting 48-dimensional feature goes through two fully connected layers to a class probability. Inputs are 14 × 13 × 50 and 56 × 13 × 50 tensors — ions × acquisition cycles × mobility bins of 0.001 Vs cm⁻². The two checkpoints together hold well under 50,000 parameters. Pretraining used roughly 2.5 million positive 2D elution groups derived from sixteen diaPASEF runs, balanced 1:1 against suboptimal elution groups as negatives, with Adam and cross-entropy loss.
Two further learned components are fitted per run rather than pretrained: DeepMall, a two-layer bidirectional GRU with attention scoring spectrum similarity from 14 per-fragment features, and twelve small binary classifiers whose ensemble merges all scores into one discriminant for false discovery rate estimation. Across eight test sets spanning timsTOF Pro, HT, SCP, and Ultra, Beta-DIA reported 8–18% more peptides and 2–8% more protein groups than DIA-NN at 1% FDR on standard-cycle data, and roughly 50% more peptides and 30% more protein groups on single-cell input. In the authors' own control, a DeepProfile pretrained on 35 files from the target cohort was slightly outperformed by the heterogeneously pretrained one.
Beta-DIA is a drop-in analysis stage for laboratories running single-shot diaPASEF on
timsTOF hardware, with its largest gains where identifications are scarcest: plasma,
dilution series, and single-cell proteomics. It reads Bruker .d folders and emits a
report whose columns mirror DIA-NN's, so downstream statistics and visualization pipelines
need no rework. It requires a predicted spectral library generated by DIA-NN, supports
methionine oxidation but not phosphorylation or acetylation, and needs an NVIDIA GPU with
more than 10 GB of memory.
Deep learning entered DIA proteomics at the periphery: predicting spectral libraries before the search, rescoring peptide-spectrum matches after it. Beta-DIA puts a pretrained model inside the search itself, and its most useful contribution is evidence that this works without asking every laboratory to train on local data. The work remains a preprint, and the release is only partly open: the search core ships as obfuscated bytecode, though the checkpoints and most modules are readable Apache-2.0 source. It also depends on DIA-NN for spectral library prediction, a tie the authors intend to break with a predictor of their own.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.