Skip to content

Latest commit

 

History

History
224 lines (172 loc) · 10.2 KB

README.md

File metadata and controls

224 lines (172 loc) · 10.2 KB

G2PT: Mechanistic genotype-phenotype translation using hierarchical transformers

Overview

Genome-wide association studies have linked millions of genetic variants to human phenotypes, but translating this information clinically has been challenged by a lack of biological understanding and widespread genetic interactions. With the advent of the Transformer deep learning architecture, new opportunities arise in creating predictive biological models that are both accurate and easily interpretable. Toward this goal we describe G2PT, a hierarchical Genotype-to-Phenotype Transformer that models bidirectional information flow among polymorphisms, genes, molecular systems, and phenotypes. G2PT effectively learns to predict metabolic traits in UK Biobank, including risk for diabetes and triglyceride-to-HDL cholesterol (TG/HDL) ratio, outperforming previous polygenic models.

Figure_1

Environmental Set-Up

conda environment file environment.yml is provided

conda env create python==3.6 --name envname --file=environment.yml

Usage

To train a new model using a custom data set, first make sure that you have a proper virtual environment set up. Also make sure that you have all the required files to run the training scripts:

  1. Participant Genotype files:
    • You can put PLINK binary file
      • --flip argument will flip ref. and alt. allele (use to recommend --flip argument, which make homozygous alt. as 2)
    • Or you can put tab-delimited file containing personal genotype data to reduce memory usage
      • Index will indicate Sample ID.
      • homozygous_a0, heterozygous, homozygous_a1 contain index of SNP by the allele
  • Example of tab-delimited genotype file
homozygous_a0 heterozygous homozygous_a1
1000909 0,1,3,5,7,9 2,4,5 6,8
1000303 1,3,6,7,8,9 2,5 4
  1. Covariates files
    • File including covariates and phenotypes.
    • same as .cov and .pheno in PLINK
      • If you want to use subset of covariates, you can put --cov-ids (i.e. with --cov-ids SEX AGE, model will use only SEX and AGE as covaritates)
    • If you do not put .cov while you put PLINK bfiles. Covariates will be generated from .fam file (Sex only)
    • If you do not put .pheno, you should include PHENOTYPE in training and validation covariate file
  • Example of covariates file
FID IID PHENOTYPE SEX AGE PC1 PC2 ... PC10
10008090 10008090 1.2 1 48 3 0.3 ... 0.5
  1. Ontology (hierarchy) file:

    • --onto : A tab-delimited file that contains the ontology (hierarchy) that defines the structure of a branch of a G2TP model that encodes the genotypes. The first column is always a term (subsystem or pathway), and the second column is a term or a gene. The third column should be set to "default" when the line represents a link between terms, (if you have nested subtree, you can put some name except 'gene'). "gene" when the line represents an annotation link between a term and a gene. The following is an example describing a sample hierarchy.

      Ontology

        • --subtree_order : if you have nested subtrees in ontology, you can set this option default is ['default'] (no subtree inside)
  • Example of ontology file
parent child interaction_type
GO:0045834 GO:0045923 default
GO:0045834 GO:0043552 default
GO:0045923 AKT2 gene
GO:0045923 IL1B gene
GO:0043552 PIK3R4 gene

  • --snp2gene : A tab-delimited file for mapping SNPs to genes. The first column indicates SNP, second column for gene, and third for chromosome

  • Example of snp2gene file

SNP_ID Gene Chromosome
16:56995236:A:C CETP 16
8:126482077:G:A TRIB1 8
19:45416178:T:G APOC1 19
2:27752463:A:G GCKR 2

There are several optional parameters that you can provide in addition to the input files:

  1. Propagation option:
    • --sys2env : determines whether model will do Sys2Env propagation
    • --env2sys : determines whether model will do Env2Sys propagation
    • --sys2gene : determines whether model will do Gene2Sys propagation
  2. Translation option:
    • --sys2pheno : Updated system embeddings are used to predict phenotype
    • --gene2pheno : Updated gene embeddings are used to predict phenotype
    • --snp2pheno : SNP embeddings are used to predict phenotype
    • if you don't put any translation option, sys2pheno will be automatically set
  3. Model parameter:
    • --hiddens-dims: embedding and hierarchical transformer dimension size
  4. Training parameters:
    • --epochs : the number of epoch to run during the training phase. The default is set to 256.
    • --val-step: Validation step
    • --batch-size : the size of each batch to process at a time. The default is set to 256.
    • --z-weight : for the continuous phenotype, individual with high absolute Z-score will be more sampled. if set as 0 (default), all population will be sampled in one training epoch
    • --dropout: dropout option. Default is set 0.2
    • --lr : Learning rate. Default is set 0.001.
    • --wd : Weight decay. Default is set 0.001.
  5. GPU option:
    • Single GPU option
      • --cuda : the ID of GPU unit that you want to use for the model training. The default setting is to use GPU 0.
    • Multi GPU option (multi-node will be supported)
      • --multiprocessing-distributed : determines whether model will be trained in multi-gpu distributed set-up
      • --world-size : size of world, default is 1
      • --rank : rank, default is 0
      • --local-rank : local rank, default is 0
      • --dist-url : distribute url, tcp://127.0.0.1:2222
      • --dist_backend : distribute backend default is nccl
  6. Model input and output:
    • --model: if you have trained model, put the path to the trained model.
    • --out: a name of directory where you want to store the trained models.

G2PT Pipeline in Overall

1. Prune Gene Ontology with your GWAS results

You can prune Gene Ontology (Biological Process) based on your GWAS summary statistics

Prune Gene Ontology based on Your GWAS results

2. Train model

You can put ontology file made from step 1.

Model Training Example (Single GPU)

python train_snp2p_model.py \
                      --onto ONTO \
                      --snp2gene SNP2Gene \
                      --train-bfile TRAIN --train-cov TRAIN.cov --train-pheno TRAIN.pheno \
                      --val-bfile VAL --train-cov VAL.cov --val-pheno VAL.pheno \
                      --test TEST --test-cov VAL.cov --test-pheno TEST.pheno \ 
                      --epochs EPOCHS \
                      --lr LR \
                      --wd WD \
                      --batch_size BATCH_SIZE \
                      --dropout DROPOUT \
                      --val_step VAL_STEP \
                      --jobs JOBS \
                      --cuda 0 \
                      --hidden_dims HIDDEN_DIMS \
                      --out OUT

Model Training Example (Multiple GPUs)

python train_snp2p_model.py \
                      --onto ONTO \
                      --snp2gene SNP2Gene \
                      --train-bfile TRAIN --train-cov TRAIN.cov --train-pheno TRAIN.pheno \
                      --val-bfile VAL --train-cov VAL.cov --val-pheno VAL.pheno \
                      --test TEST --test-cov VAL.cov --test-pheno TEST.pheno \ 
                      --epochs EPOCHS \
                      --lr LR \
                      --wd WD \
                      --batch_size BATCH_SIZE \
                      --dropout DROPOUT \
                      --val_step VAL_STEP \
                      --jobs JOBS \    
                      --dist-backend 'nccl' \
                      --dist-url 'tcp://127.0.0.1:2222' \ 
                      --multiprocessing-distributed \ 
                      --world-size 1 \ 
                      --rank 0 \
                      --hidden_dims HIDDEN_DIMS \
                      --out OUT

3. Predict with Trained Model

python predict_attention.py \
                      --onto ONTO \
                      --snp2gene SNP2Gene \
                      --bfile BFILE_prefix --cov COVAR.cov --pheno PHENO.pheno \
                      --model trained_model_dir
                      --out output_prefix \ 
                      --batch_size BATCH_SIZE \
                      --cpu N_cpu

This will generate

  • Prediction: {output_prefix}.prediction.csv, containing only predictions (Good for performance evaluation!)
  • Attention result: {output_prefix}.attention.csv, containing Nx(S+G) system and gene attention results for whole population
  • System importance score: {output_prefix}.sys_corr.csv, containing correlation between system attention and prediction
  • Gene importance score: {output_prefix}.gene_corr.csv, containing correlation between system attention and prediction

adding argument --prediction-only will make this script to predict only (no attention result)

4. Analyze Attention, Epistasis

4.1 Draw Sankey Plot from Trained Model

You can visualize attention flow from trained G2PT model.

Draw Sankey from Model Attention

4.2 Search Visualize, and Analyze Epistasis in system

You can search epistasis within system and visualize, and analyze. Please pass through example notebook

Epistais Search and Visualization Example

Future Works

  • Applying Differential Transformer to genetic factor translation
  • Build data loader for plink binary file using sgkit
  • Adding .cov and .pheno for input