SBM Lab
Proteomics and bioinformatics, measured. We run MaxQuant, FragPipe, DIA-NN and R pipelines on public data with known answers, score the results, and publish the code and the limits.
Venn Diagram Generator
Compare 2 to 4 sets and download the diagram as PNG or SVG. Useful for gene lists, DEG overlaps and other data analysis.
Open the toolLatest
Same Gene List, Different GO Enrichment: How Much the GO Release Changes the Result (2016 to 2025, Measured)
We ran the same gene lists through the same GO over-representation analysis with four Gene Ontology releases from 2016 to 2025. Planted answers and random lists behaved, but a real dexamethasone list kept only 26 of its significant terms, and 'response to glucocorticoid' was missed with the 2016 annotations.
Correct Batch Effects, Then Test? With an Unbalanced Design, ComBat Turns a 5% False-Positive Target into 29%
Correcting batch effects with ComBat or removeBatchEffect and then testing the cleaned data looks harmless. We measured it on simulated data and on a real Bioconductor data set with fake labels: when groups are spread unevenly over batches, false positives rise from 5% to 28-29%, while putting batch in the limma model stays at 5%.
DIA-NN 2.7.0: Download, Upgrade From 2.3.2, and What Changed on a Known-Ratio Benchmark
I installed DIA-NN 2.7.0 and reran the same six SWATH files I had scored with 2.3.2. It identified 6% more proteins per run and was about as accurate on the proteins both versions quantified, but the extra low-abundance values pulled E. coli accuracy from 85% to 77%. Install steps, the .wiff trap, commands and an upgrade checklist.
More
Three Samples per Group: What We Learned Switching from the t-Test to limma
At 3 vs 3, Mann-Whitney cannot reach p < 0.05, a t-test misses a two-SD difference more often than not, and limma doubles what survives FDR without lowering the false share. A known-answer simulation and four papers.
Injection Order: The Quietest Trap in LC-MS Metabolomics
Signal drift is always there in an LC-MS run. It only ruins a study when the injection order runs parallel to the experimental groups, and no batch-correction algorithm can undo that. What the literature says, and what to fix in the sequence.
Does WGCNA Really Work? Running It End to End on Public Proteomics Data
We ran the standard WGCNA flow on 110 CPTAC kidney tumours and re-checked it in a second cohort. Modules are real, but only 2 of 7 grade-linked modules beat size-matched random sets, and the top one was blood-vessel content.
STRING Confidence Score and Hub Proteins, Measured: What 0.4 vs 0.7 Changes and Whether Hubs Are Biology or Attention (STRING 12.5)
Measured on STRING 12.5: 37% of links at the default 0.4 cut-off need text mining, the top hub at 0.4 has 17 papers, and random protein lists look significantly connected 20/20 times.
DIA-NN Quantification Modes Scored on Known Ratios: QuantUMS, Legacy MaxLFQ and Fragment-Level Re-Quantification, With and Without MBR
One set of DIA-NN 2.3.2 identifications, seven ways to turn it into protein quantities, scored against known ratios. Without MBR the MaxLFQ-type routes land within about 4 points of each other. With MBR, the strongly changed proteins MBR adds are compressed: QuantUMS handles them best, and re-quantifying them from raw fragments is much worse.
Proteomics CV: What Replicate CVs Look Like in MaxQuant, FragPipe and DIA-NN (Measured on a Known-Ratio Benchmark)
Median replicate CV on the same public benchmark: DIA-NN 4.5%, MaxQuant LFQ 8.1%, FragPipe 10.1%, and 20.9% for MaxQuant's raw Intensity column. Why the number depends on the column, on which proteins survive, and on abundance, and whether a low CV means a correct ratio.