I am utilizing a benchmarking dataset to benchmark precision and recall for our pipelines implementation of mutect2 tumor only mode. We found that we can get good recalls, catching between 85- 95% of true positives, but we cannot get a precision above 8% without resulting in recall reducing down to 20-30%

With mutect2 defaults, and gold standard pipeline (learn orientation bias, contamination and filtermutectcalls), we get the following results (generated with som.py)

     type  total.truth  total.query    tp     fp     fn  unk  ambi    recall  recall_lower  recall_upper   recall2  precision  precision_lower  precision_upper   na  ambiguous  fp.region.size   fp.rate
0   indels         7990        13270  2443  10827   5547    0     0  0.305757      0.295727      0.315928  0.305757   0.184099         0.177575         0.190762  0.0        0.0      3033630367  3.568991
1     SNVs         7903        18684  1688  16996   6215    0     0  0.213590      0.204660      0.222729  0.213590   0.090345         0.086298         0.094519  0.0        0.0      3033630367  5.602528
5  records        15893        33949  4131  29818  11762    0     0  0.259926      0.253151      0.266788  0.259926   0.121683         0.118237         0.125192  0.0        0.0      3033630367  9.829147

Changing af-of-alleles not in and using fasle-discovery threshold results in a change like so: (Note this was only run on 1 chromosome but manually checked that the percentages are approximately the same for a single chromosome)

type,total.truth,total.query,tp,fp,fn,unk,ambi,recall,recall_lower,recall_upper,recall2,precision,precision_lower,precision_upper,na,ambiguous,fp.region.size,fp.rate,sompyversion,sompycmd
0,indels,667,2513,414,2099,253,0,0,0.620689655172,0.583408660615,0.656925787246,0.620689655172,0.16474333466,0.150629848617,0.179628529629,0.0,0.0,248956422,8.43119443611,som.py-v0.3.12-2-g9d128a9,/opt/hap.py-install/bin/som.py dream_files/truth_hg38_chr1.vcf SRR2020636_af3_NO_contamination_fbeta0.25.vcf -r /mnt/helomicsngs-s3/Reference/GRCh38/gatk_resource/Homo_sapiens_assembly38.fasta -o somatic_results/SRR2020636_af3_chr1
1,SNVs,676,6106,574,5532,102,0,0,0.849112426036,0.820662650549,0.874580664901,0.849112426036,0.0940058958402,0.0868782432617,0.101518314254,0.0,0.0,248956422,22.22075637,som.py-v0.3.12-2-g9d128a9,/opt/hap.py-install/bin/som.py dream_files/truth_hg38_chr1.vcf SRR2020636_af3_NO_contamination_fbeta0.25.vcf -r /mnt/helomicsngs-s3/Reference/GRCh38/gatk_resource/Homo_sapiens_assembly38.fasta -o somatic_results/SRR2020636_af3_chr1
5,records,1343,8976,988,7988,355,0,0,0.735666418466,0.711592644414,0.758725682797,0.735666418466,0.110071301248,0.103722428141,0.116671585458,0.0,0.0,248956422,32.0859367106,som.py-v0.3.12-2-g9d128a9,/opt/hap.py-install/bin/som.py dream_files/truth_hg38_chr1.vcf SRR2020636_af3_NO_contamination_fbeta0.25.vcf -r /mnt/helomicsngs-s3/Reference/GRCh38/gatk_resource/Homo_sapiens_assembly38.fasta -o somatic_results/SRR2020636_af3_chr1

How do I minimize the number of false positives that are occurring without sacrificing recall?