Primate AI

Overview

Primate AI is a deep residual neural network for classifying the pathogenicity of missense mutations. The method is described in the publication:

Publication

Sundaram, L., Gao, H., Padigepati, S.R. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat Genet 50, 1161–1170 (2018). https://doi.org/10.1038/s41588-018-0167-z

TSV File

Example

chr pos ref alt refAA   altAA   strand_1pos_0neg    trinucleotide_context   UCSC_gene   ExAC_coverage   primateDL_score
chr10   1046704 C   T   R   C   1   CCG uc001ift.3  45.49   0.849114537239
chr10   1046704 C   G   R   G   1   CCG uc001ift.3  45.49   0.795686006546

Parsing

From the TSV file, we're mainly interested in the following columns:

• chr
• pos
• ref
• alt
• primateDL_score

We also use UCSC_gene to filter out variants that don't have matching gene models in Nirvana.

Pre-processing

Converting UCSC IDs

Primate AI only provides UCSC IDs. As an initial pre-processing step, we'll need to convert these to either Entrez or Ensembl Gene IDs.

The following queries are used to download the conversions from UCSC:

mysql -h genome-mysql.soe.ucsc.edu -u genome -A -P 3306 \
      -e "select * FROM knownToLocusLink;" hg19 > ucsc_locuslink.tsv

mysql -h genome-mysql.soe.ucsc.edu -u genome -A -P 3306 \
      -e "select knownToEnsembl.name, knownToEnsembl.value, ensGene.name2 FROM knownToEnsembl, ensGene WHERE knownToEnsembl.value = ensGene.name;" \
      hg19 > ucsc_ensembl.tsv

Running the Pre-Processor

The Primate AI pre-processor can be run as follows:

dotnet PrimateAiPreProcessor.dll UGA_develop.tsv PrimateAI_scores_v0.2.tsv.gz \
     ucsc_locuslink.tsv ucsc_ensembl.tsv PrimateAI_0.2_GRCh37.tsv.gz

During conversion, 0.5% of the UCSC Ids cannot be converted to either Entrez or Ensembl gene IDs. Once the gene IDs have been acquired, we check to see which are available in Nirvana.

The following Entrez Gene IDs were not found:

399753
401980
504189
504191
100293534

Here is the output from the pre-processor:

- loading UCSC to Entrez Gene ID dictionary... 73,432 genes loaded.
- loading UCSC to Ensembl Gene ID dictionary... 76,178 genes loaded.
- loading UGA gene ID to gene dictionary... 103,277 genes loaded.
- parsing Primate AI variants... 70,121,953 variants parsed.
 
# variants with unknown gene ID: 27,253 / 70,121,953
# genes with unknown gene ID:    109 / 19,614
 
# variants not in UGA: 2,036 / 70,121,953
# genes not in UGA:    6 / 19,614

Known Issues

Known Issues

The Primate AI data set provides raw scores, but the scores are biased according to gene context. I.e. a 0.4 means something different in TP53 than it does in KRAS.

As a result, the Primate AI team provided guidance on aggregating these scores and presenting them as percentiles with respect to the associated gene. According to their research, the 25^th percentile is a good proxy for benign variants and the 75^th percentile is a good proxy for pathogenic variants.

Download URL

https://basespace.illumina.com/s/cPgCSmecvhb4

JSON Output

"primateAI":[
   {
      "hgnc":"TP53",
      "scorePercentile":0.3,
   }
]

Field	Type	Notes
hgnc	string
scorePercentile	float	range: 0 - 1.0