✨PrimateAI-3D

Overview

PrimateAI-3D is an AI annotation model developed by the Illumina Artificial Intelligence Lab to predict missense variant pathogenicity.

It uses a 3D convolutional neural network with structural and primate variation data to score missense variants. Scores range from 0 to 1, where lower values suggest more benign variants and higher values suggest more pathogenic variants.

It is trained on data from 233 primate species and has enabled the reclassification of approximately 4.4 million variants of unknown significance (VUS) as likely benign, improving the accuracy of variant interpretation. By learning from conserved sequences across primates, PrimateAI‑3D avoids bias tied to any single human ancestry.

For more details, refer to:

Publication

Gao, et al. The landscape of tolerated genetic variation in humans and primates. Science (2023). https://doi.org/10.1126/science.abn8197

Professional data source

This data source requires a Professional license. Contact [e-mail] to request access.

Parsing

TSV File

chr	pos	non_flipped_ref	non_flipped_alt	gene_name	change_position_1based	ref_aa	alt_aa	score_PAI3D	percentile_PAI3D	 refseq	prediction	 per_gene_percentile_PAI3D	hgnc
chr1	69094	G	A	ENST00000335137.4	2	V	M	0.6169436463713646	 0.5200308441794135	 NM_001005484.1	pathogenic	 0.699207135777998	OR4F5
chr1	69094	G	C	ENST00000335137.4	2	V	L	0.5557043975591658	 0.4271457250214688	 NM_001005484.1	benign	0.6053022794846382	 OR4F5
chr1	69094	G	T	ENST00000335137.4	2	V	L	0.5557043975591658	 0.4271457391722522	 NM_001005484.1	benign	0.6053022794846382	 OR4F5
chr1	69095	T	A	ENST00000335137.4	2	V	E	0.8063537482917307	 0.8032228720356267	 NM_001005484.1	pathogenic	 0.9202180376610506	 OR4F5
chr1	69095	T	C	ENST00000335137.4	2	V	A	0.5795628190040587	 0.4631329075815453	 NM_001005484.1	benign	0.6442021803766105	 OR4F5
chr1	69095	T	G	ENST00000335137.4	2	V	G	0.7922330142557621	 0.7834049546930125	 NM_001005484.1	pathogenic	 0.900396432111001	OR4F5

Illumina Connected Annotations parses all columns from the file:

• chr
• pos
• non_flipped_ref
• non_flipped_alt
• gene_name
• change_position_1based
• ref_aa
• alt_aa
• score_PAI3D
• percentile_PAI3D
• refseq
• prediction
• per_gene_percentile_PAI3D
• hgnc

The gene_name and refseq fields provide the Ensembl and RefSeq transcript IDs. These IDs are passed through as-is, so some can be obsolete or unrecognized by current Ensembl or RefSeq releases.

GRCh37

For GRCh37, a lifted-over file is provided. The file is not sorted, so sort it before use. Some RefSeq transcripts were not mapped during lift-over.

Pre-processing

Sorting for GRCh37

gzcat PrimateAI-3D.hg19.txt.gz | sort -t $'\t'	-k1,1 -k2,2n | gzip > PrimateAI-3D.hg19_sorted.tsv.gz

Generate supplementary annotation files

This step converts the raw PrimateAI-3D source file into supplementary annotation files using SAUtils that Illumina Connected Annotations can load during annotation. You only need this step if you are generating the data source manually. If you use prebuilt annotation files, you can skip it.

dotnet SAUtils.dll \
PrimateAi \
--r "${References}/Homo_sapiens.GRCh38.Nirvana.dat" \
--i "${ExternalDataSources}/PrimateAI/3D/PrimateAI-3D.hg38.txt.gz" \
--o "${SaUtilsOutput]"

Known Issues

Known Issues

Some transcript IDs in the source file are obsolete, retired, or updated. Illumina Connected Annotations does not remove or modify them.

Example:

ENST00000643905.1 transcript is retired according to Ensembl

NM_182838.2 transcript is removed because it is a pseudo-gene according to RefSeq

JSON output

"primateAI-3D": [
  {
    "aminoAcidPosition": 2,
    "refAminoAcid": "V",
    "altAminoAcid": "M",
    "score": 0.616944,
    "scorePercentile": 0.52,
    "genePercentile": 0.7,
    "classification": "pathogenic",    
    "ensemblTranscriptId": "ENST00000335137.4",
    "refSeqTranscriptId": "NM_001005484.1",
    "geneSymbol":"OR4F5" 
  }
]

Field	Type	Notes
aminoAcidPosition	int	Amino Acid Position (1-based)
refAminoAcid	string	Reference Amino Acid
altAminoAcid	string	Alternate Amino Acid
ensemblTranscriptId	string	Transcript ID (Ensembl)
refSeqTranscriptId	string	Transcript ID (RefSeq)
scorePercentile	float	range: 0 - 1.0
genePercentile	float	range: 0 - 1.0
score	float	range: 0 - 1.0
classification	string	pathogenic or benign classification
geneSymbol	string	HGNC gene symbol

Interpreting scores

PrimateAI-3D scores range from 0 to 1.

• Lower scores suggest more benign variants.
• Higher scores suggest more pathogenic variants.

Resources

• PrimateAI-3D GitHub
• Download URL