Saige v1.4.3 producing Malformed output: some rows have incorrect number of fields

[aokulabasile]

Hello,

I am running single-variant association using SAIGE v1.4.3 and am encountering an issue where some rows in the output files have an incorrect number of columns (fields).

To investigate, I ran the same association analysis three times. Each run produced an output file containing malformed rows (both fewer and more fields than expected). Notably, the affected row corresponds to a different variant/site in each run.

I have attached the three output files along with the code used to generate them.

Has anyone encountered this issue before or have any insight into what might be causing it?

Rscript /sw/saige/saige_1.4.3/SAIGE/extdata/step2_SPAtests.R \
    --vcfFile=${vcf} \
    --vcfFileIndex=${vcf}.csi \
    --vcfField=DS \
    --minMAF=0 \
    --is_imputed_data=TRUE \
    --minInfo=0 \
    --dosage_zerod_cutoff=0 \
    --dosage_zerod_MAC_cutoff=0 \
    --idstoIncludeFile=interval_file.txt \
    --chrom=${chr} \
    --nThreads=5 \
    --GMMATmodelFile=${outdir_step1}/${step1_output_prefix}.rda \
    --varianceRatioFile=${outdir_step1}/${step1_output_prefix}.varianceRatio.txt \
    --SAIGEOutputFile=${outdir}/saige_analysis_run<number>.txt

interval_file.txt
saige_analysis_run1.txt
saige_analysis_run2.txt
saige_analysis_run3.txt

[aokulabasile]

Follow-up: in our SAIGE v1.4 (and also saige v1.5) Step 2 runs, malformed output rows appeared to be caused by multithreading.

With --nThreads=5, all tested output files had rows with inconsistent field counts (expected 14 fields, observed rows with too few or too many fields).

Rerunning the same test with --nThreads=1 resolved the issue completely. This suggests a possible thread-safety/output-writing issue in Step 2, at least for our setup.

Our workaround is to force --nThreads=1 for Step 2