Deepvariant error: deep_common2/vcf.cpp:1801] Allele ->A,A not found in probability dictionary

Healthcare & Life Sciences Parabricks
,
1

I ran a large data set through deepvariant_germline and some samples fail with this error when deepsomatic runs. Our goal is to align and call against the T2T reference

[PB Info 2026-Aug-18 09:58:16] ProgressMeter -	chr1:236597000	2.4
[PB Error 2026-Aug-18 09:58:18][deep_common2/vcf.cpp:1801] Allele ->A,A not found in probability dictionary., expected iter != flattened_probs_dict.end(), exiting.
terminate called after throwing an instance of 'thrust::THRUST_200802_SM_750_800_860_890_900_1000_1200_NS::system::system_error'
  what():  trivial_device_copy D->H failed: cudaErrorCudartUnloading: driver shutting down
terminate called recursively
terminate called recursively

I had a look at the CRAM file that deepvariant_germline produced before it failed: samtools doesn’t have a problem with it. If I run prun haplotypecaller the run succeeds.

After some experiments on this sample, the run actually fails near the beginning of chromosome 2 but with several hundred reads in chr2 before the failure. Through trial and error it is the addition of this record to the sam file that causes the failure (this is not the first record at position 64)

A01298:522:H5M22DSXC:4:1671:25391:20008 113     chr2    64      17      46S23M2D82M     =       242695097       242695037       CTATCCCTATACGTAACGCTCTCCCTTACAATATCCATATCGCTTTCCCTAACCCTAACCCTTACACTATCCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCC ,F:,F,,F,,,:,F:,F,,F,,,,,,,:,,,,,,,,,FF,,,,F,,,,:FF,:,,F,,F,:F,::,,FF,,:,FF,F,FFF,FFFFFFFFFFF:FFFF:FF:FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF XA:Z:chr5,-1479,46S88M1I16M,5;chr21,+45086567,5M1I21M1D50M1D28M46S,6;   MD:Z:16A2C3^AC82        PG:Z:MarkDuplicates     NM:i:4  AS:i:87 XS:i:79 RG:Z:MRN1361

The error produces is slightly different (the Myelin warning does not appear in the original)


[PB Info 2026-Aug-18 09:49:26] ProgressMeter -	Current-Locus	Elapsed-Minutes
[PB Error 2026-Aug-18 09:49:27][deep_common2/vcf.cpp:1801] Allele ->A,A not found in probability dictionary., expected iter != flattened_probs_dict.end(), exiting.
Internal Error: MyelinCheckException: device_utils.cpp:202: CHECK(find_iter != kGpuToSmCntMap.end()) failed.

I am running Parbaricks 4.7.1-1. I tried earlier versions – they all failed but the error message was a little different (e.g. failing about decoding cram files).

This is an example of a call that fails

version=4.7.1-1
ref=chm13v2.0.fa
DIR="${DIR:-./}"
cd $DIR
base=AGS0360_t2t
#base=s2
num_gpus=2
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-2,3}"



version=4.7.1-1
ref=chm13v2.0.fa
base=s2
num_gpus=2

singularity exec --nv -B  /local/scott/t2t  \
	  /bioinf/parabricks/clara-parabricks_${version}.sif \
	  pbrun deepvariant --ref t2t/$ref --num-gpus=${num_gpus} --num-cpu-threads-per-stream 8  \
	  --in-bam ${base}.cram --out-variants ${base}.vcf.gz

Any help would be wonderful – thank you

2

Hi Shaze,

Thank you for isolating and sharing the triggering record.

We tested that record by itself using Parabricks 4.7.1-1 and the NCBI T2T-CHM13v2.0 chromosome 2 reference. The run completed successfully and produced no variant candidates, so the record alone is not sufficient to reproduce the failure. It appears to trigger the issue only in combination with the surrounding reads in your minimized dataset.

Could you please share:

  • The complete minimized SAM that reproduces the failure, including its header.
  • If available, the working version of that minimized SAM before the triggering record was added.
  • The source or download URL for the exact chm13v2.0.fa reference you used.

We do not need the complete sample. If the minimized files cannot be shared publicly, we can arrange a private transfer.

Once we have the reproducer, we can confirm the cause and validate the fix against the same input.

Thank you

3

Many thanks for the prompt response

I am happy to share the data privately – Scott.Hazelhurst@wits.ac.za

I have a data set of ~1,300 records that fails. Removing the record I shared previous allows it to work. (But it also seems to require a lot of the preceding records too – I tried to make a smaller example that failed)

FYI: If I convert this data set to BAM it fails with the same error in 4.7.1-1 but succeeds in 4.7.0-1.

If I convert to CRAM and run with 4.7.0-1 I get

[PB Info 2026-Aug-19 09:17:42] ProgressMeter -	Current-Locus	Elapsed-Minutes
[PB Error 2026-Aug-19 09:17:42][src/PBCramFile.cpp:416] cram_decode_slice_pb failed with a status of -1, expected sts == 0, exiting.
Pr

(4.7.1-1 gives the different error message I reported)

4

Thank you for providing the reduced dataset. We reproduced the reported error with Parabricks 4.7.1-1 using the matching CHM13v2 reference.

We also tested the same input with Google DeepVariant 1.9, the corresponding upstream baseline. It terminates at the same site, but reports an earlier sanity-check error:

Alt allele indices ... [[0], [0,1]], which is invalid.
ValueError: `call_variants_outputs` did not pass sanity check.

The underlying issue is an incomplete set of intermediate predictions for a multiallelic site. Google DeepVariant 1.9 detects this immediately, whereas Parabricks 4.7.1 continues further and subsequently reports the missing A,A probability. This does not indicate corruption of the BAM or reference.

We are updating Parabricks to detect this condition earlier and provide a clearer error consistent with the upstream baseline.

For the currently released Parabricks 4.7.1-1, we confirmed that the following option allows the supplied test case to complete:

--multiallelic-mode product

Please note that this option changes probability merging at all multiallelic sites and may therefore change genotype calls, QUAL/GQ values, or filters in the resulting VCF. In the supplied reduced case, the ambiguous genotype was conservatively left uncalled. Any whole-genome output produced with this option should be validated as using a different multiallelic-calling configuration.

Thank you again for helping us reproduce and understand the issue.