DIA-NN: Neural networks and interference correction enable deep coverage in high-throughput proteomics

Key results

1. 01

   ShownOn HeLa DIA runs, DIA-NN reports more precursors than OpenSWATH, Skyline and Spectronaut across estimated false discovery rate (FDR) levels. The margin is largest at stringent FDR and on short gradients, and its 0.5h run beats the open-source tools' 2h runs at 1% FDR.

   HeLa digest on QExactive HF, gradients 0.5h to 4h (PXD005573). Effective FDR was estimated with a human-maize entrapment library (202310 human, 9781 maize precursors). At 1% estimated FDR, DIA-NN on 0.5h exceeds Skyline and OpenSWATH on 2h, which the authors equate to ~four-fold higher throughput; exact counts appear only in Figure 1B. · Results; Spectral library; FDR estimation using the compound spectral…

2. 02

   ShownOn the LFQbench human/yeast/E. coli mixture benchmark (TripleTOF 6600), DIA-NN recovered the yeast and E. coli spike-in ratios more precisely than Spectronaut and produced more valid A:B ratios. This shows the method transfers to a second instrument platform.

   HYE110 runs with 64 variable windows, three injection replicates per mixture, 1% q-value, scored with the LFQbench R package. Median coefficient of variation (CV) for human peptides 5.4% and human proteins 2.9% with DIA-NN vs 7.0% and 3.8% with Spectronaut; yeast and E. coli ratio precision shown as box plots without reported test statistics. · Results; 1. Performance of DIA-NN in the LFQbench test; Figure legends

3. 03

   ShownDIA-NN is written in C++ and processes each run separately into compact files, so it runs fast with low memory on a standard workstation. The authors present it as suited to series of hundreds to thousands of runs.

   On 2x 6-core Xeon E5645 @2.4Ghz: under 12 minutes for the four HeLa runs and under 17 minutes for the three yeast runs (library-free). Peak memory stayed under 3.3Gb (HeLa) and 1.8Gb (yeast). No runtime comparison with other tools is reported. · Results; 4. Hardware requirements, speed and GUI

4. 04

   SuggestedDIA-NN can search without a DDA spectral library by predicting fragment spectra and retention times from a sequence database, and it can build libraries from gas-phase fractionation DIA runs. Library-based search still reaches greater depth.

   Library-free HeLa search against UniProt UP000005640 (maximum peptide length 50, one missed cleavage), plus a PeptideAtlas-restricted variant. Yeast and plasma libraries were built from gas-phase fractionation runs, then used to analyse 23-minute microflow SWATH triplicates on TripleTOF 6600. Counts appear only in Supplementary Figures 2 and 3. · 2. Library-free processing; 3. Generating spectral libraries with DIA-NN; Figure legends

5. 05

   SuggestedThe workflow sets its own mass correction, extraction accuracy and retention-time window and needs no retention-time standards. By contrast, Skyline and OpenSWATH mass tolerances were tuned per run, and OpenSWATH could not process the shortest gradient.

   OpenSWATH could not recognise iRT standards in the 0.5h HeLa run and was excluded from it. Skyline (7ppm or 5ppm) and OpenSWATH (15ppm or 10ppm) mass tolerances were picked per run to maximise IDs at 1% q-value, with RT windows set by gradient length; DIA-NN v1.5.7 and Spectronaut used defaults apart from disabled protein inference and output FDR filtering. · Results; Configuring DIA-NN and Spectronaut; Configuring Skyline and OpenSWATH