Bioinformatics

Best Bioinformatics Tools in 2026, Installed and Timed: What I Actually Ran This Month

Not a list of names: the proteomics, RNA-seq and single-cell tools I installed and ran on public data in October 2026, with versions, run times, the install traps I hit, and licences.

·7 min read
#bioinformatics tools#MaxQuant#FragPipe#DIA-NN#limma#DESeq2#scanpy#STRINGdb#biomaRt#proteomics software

Rewritten in October 2026. The earlier version of this page was a generic list of tool names. This version only includes tools I installed and ran myself in October 2026, on public data, with the versions, timings and problems from those runs. Each row links to the full write-up; the scripts and outputs are on GitHub in sbmlab-benchmarks.

"Best" depends on your data, so instead of ranking tools I'll tell you what each one did when I actually used it: how long it took, what broke during setup, and what the licence allows. Everything ran on one workstation (Intel i9-14900K, 128 GB RAM; Linux tools in WSL2 with 64 GB of it, DIA-NN natively on Windows), mostly with 16 threads, unless noted.

The Table

Tool (version)What I ranTimeThe trap I hitLicenceWrite-up
MaxQuant 2.8.16 Orbitrap DDA runs, LFQ + match-between-runs61.6 min (52.0 min without MBR); ~11 GB for the main processThe decoy column is now Decoy, not Reverse: the base-R filter pg[pg$Reverse != "+", ] silently returns zero rowsFreeware (free after registration)MaxQuant scored
FragPipe 24 (MSFragger 4.4.1, IonQuant 1.11.20)Same 6 runs8.9 min (+4.0 min raw → mzML), 20 GB peakSub-tools use the java on PATH: Java 8 fails, a modern Java (I used 21) must come first; headless mode won't start without diaTracer even for DDAMSFragger/IonQuant: free for academic, non-commercial use; commercial via FragmaticsFragPipe vs MaxQuant
DIA-NN 2.3.2 (Windows build, called from WSL, 24 threads)6 SWATH runs, library-free, MBR11 min library prediction + 6 h 17 min search, of which ~3.5 h was reading SCIEX .wiff files.wiff needs the Windows build; the output folder must exist; the first-pass report has no protein quantitiesAcademia edition (limited functionality, non-profit research) free; Enterprise paid. Current release is 2.7.0DIA-NN benchmark
ThermoRawFileParser 2.0.0.raw → mzML for FragPipe4.0 min for 6 files (3 at a time)-L is MS level, not log level: -L 1 writes MS1 only and the search finds 0 scansApache 2.0 (the bundled Thermo RawFileReader has Thermo's own licence)FragPipe post
limma 3.58.1 / DEqMS 1.20.0Differential expression at n = 3 on known ratiossecondsWithout a fold-change cut-off, DEqMS's extra power flagged an LFQ normalisation offset: 284 truly unchanged proteins called (limma: 104)GPL / LGPLlimma vs DEqMS
imputeLCMD 2.1 + impute (kNN)Six imputation approaches on known ratiossecondskNN erased a real 4-fold change (median +0.02 instead of +2 without MBR)GPLImputation scored
DESeq2 1.42.1 / edgeR 4.0.16 / limma-voomairway RNA-seq, 4 vs 40.4–3.1 s eachThe design (pairing by cell line) changed the results more than the choice of toolLGPL / GPLDESeq2 vs edgeR vs limma
scanpy 1.12.4PBMC3k, QC → clusters → markers~23 s, 1 GBScrublet needs scikit-image; seurat_v3 HVGs need scikit-misc (neither installed by default)BSD-3Single-cell on PBMC3k
STRINGdb 2.14.350 enrichment runs, two backgroundsnot timed; the first call downloads STRING's human network files (the links file alone is 79 MB)Default background is the whole genome: random lists of detected proteins came back with a median of 103 "significant" GO/KEGG termsGPL; STRING data CC BY 4.0STRING background test
Ensembl BioMart (martservice)Pig ↔ mouse 1:1 orthologues—biomaRt 2.58.2 failed inside useEnsembl(..., version = 111) while listing archives; querying the martservice endpoint with BioMart's XML workedEnsembl data: no restrictions; biomaRt Artistic-2.0BioMart orthologues
scikit-learn 1.9.1Biomarker classifier, leaky vs proper CVa few secondsFeature selection outside the Pipeline gave a mean AUC of 0.73 over 50 shuffles of the labels (correct: 0.5)BSD-3Data-leakage test
Public database APIs (UniProt, Ensembl, PDB, STRING, Reactome, KEGG…)One protein through 13 databasesmedian 0.09 s to 8 s per callEnsembl returned HTTP 500 on 4 of 5 calls in one round; KEGG is academic-onlyMostly CC0 / CC BY; KEGG notDatabases tested

Times are for the datasets in the linked posts; your data will differ. They are here to give a sense of scale (seconds vs minutes vs hours), not as benchmarks.

What I'd Pick, by Task

  • DDA label-free proteomics: FragPipe if you qualify for its academic licence or have a commercial one (about 5× faster than MaxQuant here including the mzML conversion, 7× without it, and it quantified more proteins, more accurately on the shared ones); MaxQuant if you need freeware with no academic restriction or a long track record in your field. Check either against known ratios before trusting it.
  • DIA proteomics: DIA-NN is the only DIA tool I ran. Its SWATH workflow beat the two Orbitrap DDA workflows on protein count and ratio spread, but that compares different instruments, not software. Budget time for vendor file reading, and record the version.
  • Proteomics statistics: limma on log intensities, with a sensible valid-value filter, and always pair the FDR cut-off with a fold-change threshold. DEqMS when you have peptide counts per protein.
  • RNA-seq differential expression: my default is DESeq2, but in my comparison none of the three was consistently more conservative; get the design matrix right first.
  • Single-cell: scanpy worked without surprises beyond the two missing dependencies. (Seurat is the common R alternative; I didn't run it for this post.)
  • Enrichment: in STRING, setting the background to the detected proteins turned about 100 false terms per random list into zero. I didn't compare tools, but set the background whichever tool you use.
  • Machine learning on omics: scikit-learn with every data-driven step inside a Pipeline, and a shuffled-label check.

Check Your Versions

Results change between versions, so the first thing to record in a methods section is what you actually ran. These are the two snippets I used for the versions above.

R:

pkgs <- c("limma", "DEqMS", "DESeq2", "edgeR", "STRINGdb", "biomaRt", "imputeLCMD", "clusterProfiler")
for (p in pkgs) cat(sprintf("%-16s %s\n", p, as.character(packageVersion(p))))
cat(R.version.string, "\n")
limma            3.58.1
DEqMS            1.20.0
DESeq2           1.42.1
edgeR            4.0.16
STRINGdb         2.14.3
biomaRt          2.58.2
imputeLCMD       2.1
clusterProfiler  4.10.1
R version 4.3.3 (2024-02-29)

Python:

from importlib.metadata import version
import sys
for p in ["scanpy", "anndata", "scikit-learn", "pandas", "polars", "numpy"]:
    print(f"{p:14s} {version(p)}")
print("Python", sys.version.split()[0])
scanpy         1.12.4
anndata        0.13.4
scikit-learn   1.9.1
pandas         3.0.6
polars         2.0.0
numpy          2.5.3
Python 3.12.3

Command-line tools print their version when started: diann-linux reports DIA-NN 2.3.2 Academia, and java -jar MSFragger-4.4.1.jar --version reports MSFragger version MSFragger-4.4.1. (scanpy.__version__ still works but now prints a deprecation warning; importlib.metadata is the quiet way.)

Licences Matter More Than Speed for Some Readers

Three of the most useful tools above are not free for commercial use: MSFragger and IonQuant (academic, non-commercial; commercial licences through Fragmatics), DIA-NN (Academia edition, with limited functionality, for non-profit research; Enterprise edition for industry), and KEGG data. If you work in a company, or might build a product, check these first. The R/Bioconductor and Python packages listed are open-source (GPL, LGPL, BSD or Artistic licences).

Limits

  • These are the tools I happened to need this month for proteomics, RNA-seq and single-cell work. Many good tools (Seurat, Spectronaut, Skyline, Perseus, nf-core pipelines, Galaxy) are not here because I didn't run them for these posts, not because they're worse.
  • One machine, public datasets, mostly small designs (3 vs 3, 4 vs 4). Times and memory scale with your data.
  • Licence summaries are from the tools' own pages in October 2026; check before relying on them.

Related