We present Epi-PRS, a polygenic prediction method that leverages genomic large language models to transform personal sequences to personal genomic and epigenomic features for disease risk modeling.

• Python=3.9.0
• TensorFlow == 2.8.0
• TensorFlow-hub == 0.11.0
• Java JDK == 1.8.0
• h5py == 3.6.0
• pyfasta == 0.5.2

Epi-PRS can be downloaded by

git clone https://github.com/kimmo1019/hicGAN

Installation has been tested in a Linux platform.

We provide detailed step-by-step instructions for running Epi-PRS.

Step 1: Personal Genome Construction

Given the variant call format (VCF) file that contains the genetics profile of all individuals, we first use VCFtools to remove all the indels and only keep the SNPs. Second, we use the reference-free Beagle software for phasing genotypes. Third, we use a personal genome construction tool vcf2diploid to obtain the paternal and maternal personal genome for each individual.

Phasing genotypes using

java -jar beagle.22Jul22.46e.jar gt=<Genotype> out=<Haplotype> map=plink.$3.GRCh37.map

where beagle.22Jul22.46e.jar can be downloaded from here. Both genotype and haplotype are in .vcf format.

Constructing personal genome using

java -jar vcf2diploid.jar -outDir <Per_Genome>  -id <ID>  -chr <Ref_Genome> -vcf <Haplotype>

where Per_Genome is the output personal genome in FASTA format. ID is the individual ID from the vcf file. Ref_Genome is the reference genome (e.g., chr4.fa). Haplotype is the haplotype file (.vcf) from the last step. Note that this command needs to be run multiple times to extract the personal genome for all the individuals. For a given LD, you only need to construct the personal genome for the corresponding chromosome.

Step 2: Genomic Features Extraction

Use genomic large language model (LLM) (e.g., Enformer) to extract the genomic features, including gene expression, chromatin accessibility, ChIP-seq and histone modification signals across a diverse panel of cell lines and tissues, for each maternal and paternal sequence.

Obtaining the genomic LLM features using

python3 get_enformer_feats.py <chrom> <start> <end> <parent> <fasta_path> <save_path>
[chrom] - chromosome ID (e.g., 1 to 22)
[start] - start position
[end] - end position
[parent] - paternal or maternal
[fasta_path] - path to the personal genome FASTA file in the last step
[save_path] - path to save the genomic LLM features (end with .hdf5)

Note that the first three parameters chrom,start,and end can be obtained from the LD list in both breast or diabetes folder. This program also needs to be run multiple times in order to get the genomic LLM features for all the individuals.

Step 3: Risk Prediction

A logistic regression model for binary classification setting and an elastics-net model for regression setting will be built based on the individual reduced-dimension features and the phenotype. We randomly select 80% of both case and control subjects as training set and the remaining 20% as testing set.

Risk prediction using

python3 risk_prediction.py <PATH> <pheno> <chrom> <start> <end>
[PATH] - path to the project (e.g., ./)
[pheno] - phenotype (e.g., breast or diabetes)
[chrom] - chromosome ID (e.g., 1 to 22)
[start] - start position
[end] - end position

We applied a dimension reduction using PCA for each 128 bp bin and only kept 5 PCs, then the PCs are pooled for all bins and fed to a logistic regression for predicting the phenotype.

Feel free to contact [email protected] if you have any problem in running Epi-PRS.

This project is licensed under the MIT License - see the LICENSE.md file for details