henhaobio
Sign in

bioRxivBioinformaticsPreprint

QuantUMS: uncertainty minimisation enables confident quantification in proteomics

Franziska KistneriD, Justus L. GroßmanniD, Ludwig Roman SinniD, Vadim DemicheviD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Mass spectrometry-based proteomics has been rapidly gaining traction as a powerful analytical method both in basic research and translation. While the problem of error control in peptide and protein identification has been addressed extensively, the quality of the resulting quantities remains challenging to evaluate. Here we introduce QuantUMS (Quantification using an Uncertainty Minimising Solution), a machine learning-based method which minimises errors and eliminates bias in peptide and protein quantification by integrating multiple sources of quantitative information. In combination with data-independent acquisition proteomics, QuantUMS boosts accuracy and precision of quantities, as well as reports an uncertainty metric, enabling effective filtering of data for downstream analysis. The algorithm has linear complexity with respect to the number of mass spectrometry acquisitions in the experiment and is thus scalable to infinitely large proteomic experiments. For an easy implementation in a proteomics laboratory, we integrate QuantUMS in our automated DIA-NN software suite.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (24 Jun 2023) · CC BY 4.0

QuantUMS, a machine-learning module in DIA-NN, weights MS1 and fragment signals by quality-derived variance estimates, improving DIA precision, removing ratio compression in high-accuracy mode, and scoring each quantity's reliability.

Target-decoy competition controls peptide identification errors, but quantification errors could not be estimated, so users relied on coefficient-of-variation filters or manual chromatogram checks. DIA-NN previously summed the top three fragments and set the MS1 signal aside; QuantUMS uses the MS1 and fragment signals, weights each by its quality score and reports quantity quality scores for precursors and proteins.

Key results

  1. ShownThe high-precision mode of QuantUMS reduced the spread of inferred precursor and protein ratios compared with legacy DIA-NN quantification. The largest gains were for low-intensity precursors and proteins.

    LFQbench-type human K562 / E. coli mix (A:B:C 1:1:1 human, 50:33:20 E. coli; 5-minute analytical flow gradient, timsTOF Pro dia-PASEF): compared with legacy DIA-NN, log2-standard deviations of ratios improved over 1.4-times for low-intensity precursors and 1.3-times (human) and 1.2-times (E. coli) for low-intensity proteins. · QuantUMS improves precision and boosts accuracy…; Figure legends

  2. ShownIn high-accuracy mode, median inferred ratios matched the expected mixing ratios, removing the ratio compression seen in the legacy and high-precision modes, though precision was not significantly better than legacy. Filtering on the per-quantity quality metric kept mainly accurately quantified features.

    Same mixed-species benchmark, A:C ratios plotted against the expected human 1:1:1 and E. coli 50:33:20 design. Run-specific protein-level filtering at the 0.75 quality quantile, in high-accuracy mode, gave almost perfect protein ratios. Results are shown as boxplots; no numeric error values are given in the text. · QuantUMS improves precision and boosts accuracy…; Figure legends

  3. ShownAt a matched effective FDR, both QuantUMS modes found more true differential E. coli features than legacy DIA-NN. Gains were larger for precursors than proteins, for the tighter A:B comparison, at stricter confidence and for low-abundance features. High-precision mode did slightly better than high-accuracy mode.

    Welch t-test on each precursor or protein; effective FDR is the share of human (false) hits. High-precision mode reported 871 differentially expressed proteins at 5% FDR for A:C. The E. coli total for comparison is 1078 proteins, counting those seen in two or more replicates of each of A and C. Detection rate by intensity rank places the advantage in low-abundance features. · QuantUMS improves precision and boosts accuracy…; Statistical analysis; Figure legends

  4. SuggestedIn 50 reanalysed chronic lymphocytic leukaemia (CLL) samples, QuantUMS found more differentially expressed proteins than legacy DIA-NN, which beat the original Spectronaut analysis. In the leukocyte-count and treatment-status tests, filtering by averaged protein quality raised significant hits at a given FDR.

    100-minute nanoLC gradient, timsTOF Pro dia-PASEF (PXD022216); limma tests against four phenotypic characteristics. The largest gains were for leukocyte count and treatment status. In those two tests, quality filtering raised the number of significant proteins at a given FDR by reducing the number of tests to which Benjamini-Hochberg correction is applied. Counts are not stated in the text. · QuantUMS improves precision and boosts accuracy…; Statistical analysis; Figure legends

  5. SuggestedHyperparameters trained on a subset of the benchmark (A+C, A alone or C alone) and applied to all three mixtures performed about as well as those trained on A+B+C. The authors read this as hyperparameters reflecting the LC-MS setup and sample matrix rather than the experimental design.

    High-accuracy mode, Supplementary Figure S2: A:C ratio boxplots and log2-standard deviations in three intensity bins were similar to training on the full data (Figure 1b). Tested on one dataset only; no numeric comparison is stated. · QuantUMS improves precision and boosts accuracy…; Figure legends

Methods, data and reproducibility

QuantUMS estimates a log-variance and a log-bias for each MS1 and fragment signal from its intensity and a DIA-NN quality score. It merges signals by minimum-variance weighted averaging into precursor quantities and then, through a variance-weighted MaxLFQ, into proteins. Hyperparameters are fit by backtracking gradient descent to maximise agreement between features and to remove abundance-linked bias.

  • LFQbench-type mixed-species benchmark: human K562 and E. coli tryptic digests in mixtures A, B, C (human 1:1:1, E. coli 50:33:20), 5-minute analytical flow gradient, timsTOF Pro dia-PASEF · human (K562 cells); E. coli · n = 3 mixtures (A, B, C) · PXD040205
  • Chronic lymphocytic leukaemia (CLL) samples, 100-minute nanoLC gradient, timsTOF Pro dia-PASEF, originally analysed with Spectronaut · human (CLL patient samples) · n = 50 samples · PXD022216
Methods & tools
QuantUMS, DIA-NN 1.8.1, DIA-NN 1.8.2 beta 22, MaxLFQ (variance-weighted variant), timsTOF Pro (dia-PASEF), Spectronaut, limma 3.50.3, R 4.1
Software & environment
Spectral libraries generated and refined with DIA-NN 1.8.1; QuantUMS implemented in DIA-NN 1.8.2 beta 22; statistics in R 4.1 with limma v 3.50.3.
Reproducibility gaps
There is no standalone QuantUMS code yet; an open-source release independent of DIA-NN is described as in preparation. The DIA-NN build and figure scripts are shared through an OSF view-only link.
Limitations (6)
  • Bottom-up protein quantities cannot reliably separate a change in a protein's level from a change in its post-translational modification, and the protein quality score does not address this. · stated by the authors
  • Removing ratio compression costs some precision: the high-accuracy mode gave no significant precision gain over legacy DIA-NN, so users must trade precision against accuracy. · stated by the authors
  • The variance model assumes known variances and independent measurement errors, which the authors note do not strictly hold; hyperparameter tuning and variance calibration are used to compensate. · stated by the authors
  • Tested only on DIA data with DIA-NN quality scores; use with selected or parallel reaction monitoring and multiplexed MS/MS is left to future work, and the quality metric reduces manual chromatogram checks rather than eliminating them. · stated by the authors
  • Both test datasets come from timsTOF Pro dia-PASEF runs, and the CLL comparison counts significant proteins without ground truth, so more hits there do not by themselves show better accuracy. · noted in this summary
  • Linear scaling to very large experiments is argued from algorithmic complexity alone; no runtime, memory or large-cohort measurements are reported. · noted in this summary

Read this if you quantify DIA proteomics data with DIA-NN and want variance-aware precursor and protein quantities, less ratio compression, or a per-quantity quality score to filter on before differential analysis.

Citations
50

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPino '19Pham '26Chang '18Maia '19Szyrwiel '25Frey '25
75% semantic matchBiochemistry
L. K. Pino, B. C. Searle, H.-Y. Yang+3 more·bioRxiv·Jul 30, 2019