Full-DIA

In single-cell proteomics, most of the protein matrix is empty. A search engine reports a protein in a cell only when that cell's run clears a run-specific false discovery threshold, so a matrix over a few hundred cells is riddled with gaps that are part biology and part detection statistics. Whatever fills them — zeros, a probabilistic minimum, a Bayesian reconstruction — bends the clustering and enrichment that follow.

Full-DIA removes the gaps instead of filling them. It keeps the best elution group and all fragment intensities for every library peptide in every run regardless of that run's q-value, controls error globally across the experiment rather than per run, and denoises the resulting complete fragment-ion matrix with a self-supervised autoencoder before rolling it up into peptides and proteins. Every number in the finished matrix is a measured intensity, corrected — none is imputed.

Identification rests on a pretrained scorer. diaPASEF on timsTOF instruments resolves ions in mobility as well as retention time, so an elution group is a two-dimensional image rather than a set of curves; DeepProfile, a small convolutional network trained once on sixteen heterogeneous runs and shipped inside the package, reads that image and returns a coelution probability. Full-DIA comes from Jian Song, working between Soochow University and the Meyer lab at Cedars-Sinai Medical Center, with Shanghai Omicsolution, whose runs supplied the pretraining corpus. It carries DeepProfile over unchanged from Beta-DIA, and joins DIA-BERT and DIA-CLIP among search tools built around a pretrained scorer.

#Key Features

• Complete protein matrix: fragment intensities are kept for every peptide in every run, so the sample-by-protein matrix is dense under one global 1% FDR rather than sparse under per-run thresholds.
• Coelution as an image: DeepProfile scores the retention time × ion mobility map of an elution group taken from profile data, one variant scoring the monoisotopic ions alone and the other all four isotope channels jointly.
• Frozen features, refined head: a second pass adapts DeepProfile to the file at hand by updating only its classification layers, leaving the feature extractor as pretrained.
• Autoencoder intensity correction: DeepQuant reconstructs fragment intensities from normalized signal and coelution scores, weighting its loss by fragment correlation and peptide confidence.
• GPU-resident search: peak picking, scoring, and quantification run on the GPU, making analysis faster than DIA-NN on three of four benchmark sets.

#Technical Details

DeepProfile normalizes its input map per ion and globally, passes each version through two rounds of convolution and max pooling into a 16-dimensional vector, and maps the concatenated pair, plus an embedding of the fragment ion count, through two dense 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 — and the two shipped checkpoints together hold just under 50,000 parameters. Pretraining used roughly 2.5 million positive 2D elution groups from sixteen runs mixing sample types, gradients, and window-slicing schemes. Three further components are fitted per experiment: DeepMall, a bidirectional GRU scoring spectral similarity; DeepQuant, whose layer widths follow the number of runs; and twelve binary classifiers fusing 791 functional scores with the learned ones into one target-decoy discriminant.

Against DIA-NN 2.2 at 1% global FDR, Full-DIA reported 14%, 32%, 55%, and 14% more precursors on the plasma, SC-293T, SC-LPS, and SC-Cycle datasets, with protein group counts changing by 1%, 7%, 19%, and −4%. On a plasma dilution series the median correlation between intensity and loading amount was 0.934 for peptides and 0.953 for proteins, against 0.882 and 0.946. On cell-cycle data the complete matrix separated the four phases with a silhouette score of 0.25 versus 0.20, and at 0.2 ng of HeLa digest peptide identifications rose roughly 16%.

#Applications

Full-DIA is an analysis stage for laboratories running single-cell or low-input diaPASEF on timsTOF hardware, and its gap-free matrix matters most where the statistics are sensitive to imputation: differential expression, cell-cycle phasing, and pathway enrichment. It reads Bruker .d folders and writes a report whose column names mirror DIA-NN's, so existing analysis pipelines need no rework. It requires a predicted spectral library from DIA-NN, supports methionine oxidation but not phosphorylation or acetylation, and needs an NVIDIA GPU with more than 10 GB of memory.

#Impact

The peer-reviewed evaluation is evidence that a scorer pretrained once on sixteen unrelated runs transfers without per-laboratory pretraining: the authors' own control, a DeepProfile pretrained on 35 files of a urine cohort, was slightly outperformed by the heterogeneous one. What the field is likelier to borrow is the matrix — every peptide quantified in every run under one global error rate, so completeness comes from measurement. The pretrained part is small: only DeepProfile ships as a checkpoint, and DeepQuant cannot in principle be one, since its layer widths depend on how many runs the user supplies. The pretraining runs are not publicly deposited, and the workflow still depends on DIA-NN for library prediction.