Benchmarks
Nexju against the original tools, side by side
Nexju, ADMIXTOOLS 2, EIGENSOFT and DATES run the same analyses on the same AADR files. The results match, and Nexju finishes first.
Watch both run
Pick an analysis. Both lanes play their measured wall times on one clock, sped up when the run is long.
DNA file to a qpAdm fit
A 23andMe-format file of 552,000 SNPs as the target, 3 sources and 11 right groups on the AADR 1240K. Each step ran once, the Nexju fit 3 times. The merge into the whole AADR alone took 6 minutes and 14 GB of memory.
Doing it yourself
EIGENSOFT + R
- plink --23file0.06 s
- convertf to EIGENSTRAT1.84 s
- mergeit into the whole AADR361.4 s
- qpadm() in R13.3 s
On Nexju
Nexju’s engines
- Convert your file0.76 s
- Copy the 14 populations0.68 s
- Add your genome1.80 s
- qpAdm fit0.11 s
Same result: The weights agree to 3 decimals, and both p-values (0.0092 and 0.0089) reject the model. mergeit drops 8,139 SNPs where the genome has no call, which most likely shifts the jackknife blocks.
Every measurement
Wall time for each tool on the same input, and what was compared between the two outputs.
Same job, same files
7 analyses| Analysis | Nexju | Original | Faster | Same result |
|---|---|---|---|---|
| DNA file to a qpAdm fitA 23andMe-format file of 552,000 SNPs as the target, 3 sources and 11 right groups on the AADR 1240K. Each step ran once, the Nexju fit 3 times. The merge into the whole AADR alone took 6 minutes and 14 GB of memory. | 3.35 s | 376.6 sEIGENSOFT + R | 112× | The weights agree to 3 decimals, and both p-values (0.0092 and 0.0089) reject the model. mergeit drops 8,139 SNPs where the genome has no call, which most likely shifts the jackknife blocks. |
| qpAdm fitEngland_Saxon from 3 sources and 10 right groups, 851 AADR genomes, 1.2M SNPs. One core each. | 0.39 s | 19.9 sADMIXTOOLS 2 (R) | 51× | The same p-value (1.585e-05) and the same three weights to 6 decimals. |
| qpAdm rotationEvery 1- to 3-source model from a pool of 6: 41 fits, 1,043 genomes, the same rights. Nexju on all 14 cores, R on one. | 0.93 s | 32.7 sADMIXTOOLS 2 (R) | 35× | The same 5 of 41 models fit (p ≥ 0.01) in both. |
| smartpca run1,486 genomes, 1.2M SNPs, 6 PCs from 234 reference genomes, the rest projected. One core each. | 5.55 s | 257.5 sEIGENSOFT (C) | 46× | All 234 eigenvalues and all 1,486 × 6 coordinates are identical at the printed 4 decimals. |
| DATES runASW dated as a mixture of YRI and CEU, 261 genomes, 1.2M SNPs, jackknife by chromosome. Nexju on all 14 cores, C on one. | 4.39 s | 17.3 sDATES (C) | 3.9× | The same date, 6.726 generations. The jackknife gives 6.703 ± 0.372 against 6.697 ± 0.371, because the two fits start from different random number generators. |
| qpGraph fit8 AADR populations, 1.2M SNPs, 4 admixture events, from genotypes. One core each. | 1.23 s | 8.97 sADMIXTOOLS 2 (R) | 7.3× | The same graph score, 1249.748217. |
| qpGraph searchfind_graphs with 2 admixture events on the same 8 populations, search only. Nexju on all 14 cores, R on one. | 0.44 s | 41.3 sADMIXTOOLS 2 (R) | 94× | Both searches are random and stop in different local optima, so the best graphs differ. |
Before the first run
Most of the time you save comes before any analysis starts.
Doing it yourself
- Install R and ADMIXTOOLS 2, or compile ADMIXTOOLS, EIGENSOFT and DATES from C.
- Download the AADR. The 1240K genotype file alone is 7.2 GB.
- Convert your 23andMe or AncestryDNA file with PLINK and convertf, then merge it into the whole AADR, which rewrites every genome in it to add yours. On our laptop that merge took 6 minutes and 14 GB of memory.
- Find population labels in a 13 MB annotation table and write a parameter file for each tool.
On Nexju
- Sign in with Google.
- Upload your raw DNA file once. Each run adds your genome to a copy of only the populations it uses. That took 1.8 s in the same test.
- Pick populations by name, date, place or nearest N25 neighbour, from 23,900 genomes kept current with new papers.
- Press Run. The analysis takes the Nexju times above.
How we measured
- One Apple M4 Pro laptop with 14 cores and 24 GB of memory.
- Each tool runs in a fresh process, and its wall time includes start-up. One untimed run first puts the files in the page cache. The times are medians of 3 runs (5 for qpGraph).
- Both tools read the same files, the AADR v66 1240K release with 1.2M SNPs, cut to the populations of each analysis. In the DNA-file race, the do-it-yourself side reads the whole AADR.
- ADMIXTOOLS 2.0.10 in R, EIGENSOFT (2021 source, built with OpenBLAS), DATES v4010 in C and PLINK 1.9, against the engines Nexju runs.
- The DNA-file race starts from a 23andMe-format file made from a public AADR genome, so no one’s real DNA was used. Its steps ran once each. The Nexju steps ran in a benchmark program on the same libraries as Nexju’s workers.
- That race leaves out one step a real run needs. EIGENSOFT 2021 and ADMIXTOOLS 2.0.10 misread the AADR v66 genotype format, so the whole AADR must first be converted to the older packed format. That conversion is not counted, so the do-it-yourself total is lower than it would be.
- A run on nexju.com also waits in the queue and reads the data over the network, on servers with fewer cores than the laptop. Those times are not in these numbers.
Run your first model on Nexju
The free plan runs qpAdm, qpWave and f-statistics on the whole AADR, with nothing to install.