AD-genomicLLM

Associating genotype to imaging and clinical phenotypes of Alzheimer’s disease by leveraging genomic large language model

In this work, we propose a novel computational framework that leverages genomic large language models (LLMs) to enhance the association analysis between genetic variants and Alzheimer's disease (AD)-related phenotypes, including imaging and clinical features.

Enviornment

• Python==3.9.0
• java==1.8.0
• TensorFlow==2.12.0
• TensorFlow-hub==0.12.0
• pyfasta==0.5.2
• scikit-learn==1.0.2

Apart from the above softwares/packages, please also make sure the following softwares are installed properly: 1) vcftools for processing WGS .vcf file. 2) Beagle for genotype phasing. 3) vcf2diploid for constructing personal genome. 4) FreeSurfer for processing sMRI images.

Instructions

We provide detailed step-by-step instructions for running our pipeline.

Processing genotype data

In our study, we downloaded whole genome sequencing (WGS) data from ADNI database, which provides .vcf file (gzip compressed) for each chromosome. Here, we take the chr19 and use gene APOE as a demonstration case study.

Step 1: remove indels

vcftools --gzvcf ADNI.808_indiv.minGQ_21.pass.ADNI_ID.chr19.vcf.gz  --remove-indels --recode --recode-INFO-all --out SNPs_ADNI.808_indiv.minGQ_21.pass.ADNI_ID.chr19
[gzvcf] - input compressed vcf file
[out] - output file name (prefix)

Step 2: genotype to haplotype

java -jar preprocess/beagle.22Jul22.46e.jar gt=SNPs_ADNI.808_indiv.minGQ_21.pass.ADNI_ID.chr19.recode.vcf out=SNPs_ADNI.808_indiv.minGQ_21.pass.ADNI_ID.chr19.recode_hap map=plink.GRCh37.map/plink.chr19.GRCh37.map
[jar] - path the the beagle java program
[gt] - input vcf file from Step 1
[out] - output file name (prefix)

Note that the beagle .jar file is from here and the plink full map files are from here.

Step 3: personal genome construction

java -jar vcf2diploid_v0.2.6a/vcf2diploid.jar -outDir fasta/chr19  -id  003_S_1057 -chr hg19/chr19.fa -vcf SNPs_ADNI.808_indiv.minGQ_21.pass.ADNI_ID.chr19.recode_hap.vcf.gz
[outDir] - output directory
[id] - personal ID
[chr] - reference genome for a chromosome
[vcf] - input vcf file from Step 2

Note that the above command can only construct the personal genome (both maternal and paternal) per individual per chromosome. For the construction of multiple individuals, the above command should be iterated over all individuals.

The vcf2diploid.jar was downloaded from here. The reference genome for a chromosome can be downloaded from here.

Above the above three steps, one should get chr[CID]_[PID]_maternal.fa and chr[CID]_[PID]_paternal.fa in the fasta/chr[CID] folder where CID,PID denotes chromosome ID and personal ID, respectively.

In the above example case, it is chr19_003_S_1057_maternal.fa and chr19_003_S_1057_paternal.fa.

Extracting genomic LLM features

python3 get_llm_feats.py --gene_name [gene_name] --fasta_path [fasta_path] --refGene_path [refGene_path] --output_path [output_path]
[gene_name] - gene of interest, e.g., APOE
[fasta_path] - path to the fasta in the last step, e.g., fasta/chr19/chr19_003_S_1057_maternal.fa
[refGene_path] -path to the refGene file, e.g., refGene_hg19_TSS.bed
[output_path] - output path

The refGene file records the TSS information for each gene. The output_path will be created if not exist. A python .npy file with the same prefix (e.g., chr19_003_S_1057_maternal.npy) will be generated under the output_path folder with the shape (3,896,5313). It represents the 5313 features in 896 bins for 3 genomic LLM input regions.

Note that the Python script is designed for per fasta file per gene. For extracting genomic LLM features for a large number of individuals, GPU is recommended to accelerate the process.

Processing MRI image data

The sMRI images of 246 individuals were downloaded from ADNI database (entry name: ADNI1_Complete_1Yr_1.5T) in .nii format. We put the .nii raw data from each individual each time point in a separate folder. Note that some individuals may have multiple .nii files from the same time point. The image data are organized as the structure below

 MRI/
    |-- 003_S_1057_bs/
    |   |   |   |   |--I52821.nii
    |-- 003_S_1057_m06/
    |   |   |   |   |--I81339.nii
    |-- 003_S_1057_m12/
    |   |   |   |   |--I96202.nii
    |-- 007_S_0128_sc_bs/
    |   |   |   |   |--I118683.nii
    |   |   |   |   |--I36640.nii
    |-- 007_S_0128_sc_m06/
    |   |   |   |   |--I121135.nii
    |-- 007_S_0128_sc_m12/
    |   |   |   |   |--I59863.nii
    ...

Step 1: Processing sMRI images

export SUBJECTS_DIR=[path-to-project]/img_output
recon-all -s 003_S_1057_bs -i MRI/003_S_1057_bs/I52821.nii -all

One can set an environment variable SUBJECTS_DIR to specify the output path. Note that the above per individual per time point command needs to be iterated over all individuals and all three time points (baseline, m06, and m12)

Step 2: Extracting imaging phenotypes

aparcstats2table --subjects 003_S_1057_bs ... 007_S_0128_sc_bs  --hemi lh --meas thickness --parc=aparc --tablefile=parcstats_thickness_lh.txt --skip
[subjects] - subject IDs separated by space
[meas] - imaging phenotype (thickness, area, volume etc)

We provided the extracted imaging phenotypes (thickness, area, and volume) of 246 individuals across 3 time points in the MRI folder.

Associating genotype to imaging phenotypes

python3 get_img_association.py --gene_name [gene_name] --llm_path [llm_path] --refGene_path [refGene_path]
[img_feat_type] - MRI image feature type. e.g., 'thickness'
[gene_name] - gene of interest, e.g., APOE
[llm_path] - path to the folder containing genomic LLM feature .npy files
[refGene_path] -path to the refGene file, e.g., refGene_hg19_TSS.bed
[res_path] - path to save the association results

res_path will contain the association metric (e.g., Pearson's correlation) between a gene and a specific brain region of interest (ROI).

Associating genotype to clinical AD phenotypes

python3 get_AD_association.py --img_feat_type [img_feat_type] --gene_name [gene_name] --llm_path [llm_path] --refGene_path [refGene_path] --res_path [res_path]
[gene_name] - gene of interest, e.g., APOE
[llm_path] - path to the folder containing genomic LLM feature .npy files
[refGene_path] -path to the refGene file, e.g., refGene_hg19_TSS.bed

The auROC will be calculated for binary AD trait.

Contact

If you have any questions regarding our code or data, please do not hesitate to open an issue or directly contact me ([email protected]).

Cite

If you used our work in your research, please consider citing our paper

Qiao Liu, Wanwen Zeng, Hongtu Zhu, Lexin Li, Wing Hung Wong. Leveraging Genomic Large Language Models to Enhance Causal Genotype-Brain-Clinical Pathways in Alzheimer’s Disease [J]. medRxiv. 2024.

License

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