This October 7 issue brings together four catch-up selections from an overlapping 14-day look-back: three journal publications and one preprint. The listed versions first appeared online between September 29 and October 5, 2026; they were not all first published today. The journal versions of 2-step and PiProteline follow earlier preprints.
The selection moves from cross-system retention-time prediction and fragment-aware self-supervised learning to targeted quantification and downstream proteomics analysis. Across these topics, realistic transfer conditions, suitable data splits, simple baselines and transparent QC are central to interpreting the reported results.
Reading depth varies. Three papers received targeted checks of their main text or publisher method text; MRMhub was assessed from publisher preview/indexed text, metadata and official documentation. No independent experimental or computational reproduction was performed. The paper-specific reading scopes below state these limits.
Coverage is incomplete: some ACS ASAP, Molecular & Cellular Proteomics in-press, Talanta and date-filtered search pages were inaccessible. This issue does not claim to cover all newly published work. Unverified leads were excluded, and an indexing date alone was not treated as a first-online date.
Times are changing but order matters: transferable prediction of small-molecule liquid chromatography retention times
Nature Methods · 2026
First online: 2026-10-01
Publication status: Peer-reviewed journal article. An earlier ChemRxiv preprint exists; the first-online date shown is for the journal version, not the first disclosure of the study.
Reading scope: Main text and selected Methods checked. Supplementary materials, raw data and code were not fully audited; no independent reproduction.
DOI: 10.1038/s41592-026-03243-2
2-step separates retention-order prediction from conversion to absolute retention time in reversed-phase LC. A molecular graph model predicts order, then high-confidence anchors from the target system support a robust time mapping. In chemically heterogeneous splits, the authors report better performance on five of six datasets and a near tie on the sixth. The cross-system evaluation is informative, but the method still needs target-system anchors, column descriptors and pH. A network trained without target-system data therefore still relies on target-system information. HILIC remains outside the demonstrated scope; some biological examples lack confirmation with authentic standards, and even the more realistic splits may be optimistic.
Read paperdIon: Fragmentation-Based Invariance for Self-Supervised Learning of Tandem Mass Spectra
arXiv, cs.LG · 2026
First online: 2026-10-05
Publication status: arXiv v1 preprint, not peer reviewed. Submitted 2026-10-05 at 13:09:47 UTC. arXiv lists the DOI with DataCite registration pending; use the arXiv paper link.
Reading scope: Preprint main text, Tables 1–3 and selected appendices checked. Remaining appendices and code were not fully audited; no independent reproduction.
DOI: 10.48550/arXiv.2610.06282
dIon builds self-supervised views from precursor-conditioned spectrum mixtures and partial spectra with the precursor withheld, encouraging representations that use fragment evidence rather than a precursor-only shortcut. Under matched downstream training conditions, the authors report gains of 2.3–8.4 percentage points in peptide precision over training from scratch. The frozen representation does not outperform a simple binned spectral-angle baseline, and the strongest retrieval results require supervised adaptation. Most evaluation labels come from search engines and may inherit selection and identification errors; computational cost is also substantial. This is a preprint, and its peptide/proteomics results do not establish transfer to metabolomics.
Read paperMRMhub: a scalable data-processing framework for large-scale targeted metabolomics
Nature Metabolism · 2026
First online: 2026-09-29
Publication status: Peer-reviewed journal publication, Correspondence.
Reading scope: Publisher preview/indexed text, author metadata and official documentation checked. The complete paper and supplementary benchmarks were not obtained; no independent reproduction.
DOI: 10.1038/s42255-026-01629-2
MRMhub connects cross-sample peak integration, quantification and quality control for targeted MRM. INTEGRATOR uses chromatographic patterns across the analysis sequence to address retention-time drift and interference, while QUANT supports validation, quantification, artefact correction, filtering and QC reports. Official examples cover 937 samples with 503 features and 4,591 samples with 829 features. These examples demonstrate processing scale, not independent accuracy validation. The complete paper and supplementary benchmarks were not obtained, so claims of superior accuracy or a specific speedup cannot be assessed here. The scope is targeted MRM and does not establish untargeted metabolite identification performance.
Read paperPiProteline: An R Package for Integrated Proteomics Data Analysis, from Label-Free Quantitation to Systems Biology
Biology · 2026
First online: 2026-10-01
Publication status: Peer-reviewed journal article. An earlier preprint was posted on 2026-07-14 (10.20944/preprints202607.0993.v1); the first-online date shown is for the journal version.
Reading scope: Publisher text returned by search, including Methods, Results and Discussion, and the official repository checked. The PDF and supplements were not reviewed in full, and supplementary Figure S1 was not independently rerun.
DOI: 10.3390/biology15191723
PiProteline combines preprocessing of label-free abundance matrices, differential analysis, functional enrichment and protein-interaction network analysis in R and Shiny. On a known spike-in benchmark, the authors report performance similar to DEP and protti and below limma. Its main contribution is an integrated analysis route. Merging same-gene isoforms by the per-sample maximum loses isoform information. For weighted networks, the authors recommend 20–30 samples per group; relaxed completeness filtering uses pairwise-complete correlations, which do not resolve the missing-data mechanism. The empirical composite network measure is not a calibrated discovery probability. Input must be on the abundance scale, with no automatic conversion of log2 intensities back to linear abundance.
Read paper