Can Spades be made deterministic?

[andreyto]

I am using Spades in a clinical strain surveillance application. It would be highly desirable to be able to generate exactly the same outputs from the same inputs and parameters every time that I run Spades. I need a complete "repeatability".
In my tests, this does not seem to be the case.
Would you have any pointers on how much work it would take to add a --deterministic switch?
If I implemented this, would you be interested in a pull request?
Quick search for srand in v.3.11.0 code suggests this:

• Spades code itself tries to fix the seed in various places (srand(48))
• bwa and samtools from the ext directory - not clear yet, probably not deterministic. Might have to change both how they are called with the seed argument, as well as the code where identical reads are randomly placed.
• nlopt is definitely not. I can see this line nlopt_srand_time_default(); /* default is non-deterministic */. They want to have different seeds in different threads, and that function generates the seed by combining time with the thread ID.
• Is multithreading in Spades likely to be a separate physical source of randomness, regardless of the seed being used in the generator? If so, I am willing to run everything in a single thread in the --deterministic mode. Otherwise it might take too much work to implement the fixed ordering of work among the threads.

[asl]

Could you please report the cases when SPAdes output is non-deterministic (e.g. input / parameters and the observed behavior)? SPAdes is designed to be deterministic modulo the # of thread option (so, the output for 8 threads could be different compared to the output of SPAdes using 16 threads) and our tests shows that it is indeed so.

[andreyto]

@asl You are correct. Spades indeed appears to be deterministic. I must have incorrectly assigned to Spades some other source of randomness in my overall workflow. I have now ran the script below, and got back identical md5 sums from all 40 invocations.
Thanks!

$ cat run_spades.sh
#!/bin/bash
#PBS -t 1-40%20
#PBS -l nodes=1:ppn=16
echo "Host" $(hostname) "Time Begin" $(date) "Uptime" $(uptime)
cd $PBS_O_WORKDIR
set -ex
#TEST_DATA=$HOME/work/spades/assembler/test_dataset
#TEST_FQ_1=$TEST_DATA/ecoli_1K_1.fq.gz
#TEST_FQ_2=$TEST_DATA/ecoli_1K_2.fq.gz
TEST_DATA=my_data_dir
TEST_FQ_1=$TEST_DATA/my_reads_1.fastq.gz
TEST_FQ_2=$TEST_DATA/my_reads_2.fastq.gz
OUT_DIR=out.$PBS_ARRAYID
spades.py \
    -1 $TEST_FQ_1 \
    -2 $TEST_FQ_2 \
    -t 16 \
    --sc -k 33,55,77,99,127 --careful \
    -o $OUT_DIR
seqkit stats -a $OUT_DIR/contigs.fasta > $OUT_DIR/contigs.stats
## sort by sequence and renumber contigs to make the file independent of contig output order
seqkit sort -s $OUT_DIR/contigs.fasta | perl -pe 's/>.*/>$./g' > $OUT_DIR/contigs.sorted.fasta
md5sum $OUT_DIR/contigs.sorted.fasta > $OUT_DIR/contigs.md5

echo "Host" $(hostname) "Time End" $(date) "Uptime" $(uptime)

[asl]

Great! Note that there is no need to reorder the contigs. The output is deterministic as well :)