This repository can be used to create embeddings from a a set of genes using HyenaDNA.
Clone the repository.
Create a mamba/conda environment and setup with all packages described in requirements.txt/.yaml.
Download the humane genome ref GTF (and unpack zip) from (https://ftp.ensembl.org/pub/release-114/gtf/homo_sapiens/) for ensemble gene ids, clinvar_pathogenic.tsv, gnomad_high_impact_variants.tsv, and the set of patient genes e.g. 'sample' (mygene2_5.7.22.txt) into the ./data folder.
Model: create ./model directory, cd run 'git lfs install' Clone the HyenaDNA model from HuggingFace or their own repo: git clone https://huggingface.co/LongSafari/hyenadna-medium-450k-seqlen-hf -- for portability, we then rename and replace '-' with '_' in the directory name: 'hyenadna_medium_450k_seqlen_hf'
We then have the following pipeline to create the gene embeddings, run these source scripts;
1_generate_unique_genes_file.py (to create a list of unique gene IDs from your sample) 2_generate_gene_sequences.py (to create the gene FASTAs) 3_generate_gene_tokens.py (using local HyenaDNA to create the tokens using the FASTAs) 4_generate_gene_embeddings.py (using HyenaDNA to create the embeddings from the tokens, using chunking and average pooling to fit the 450k model)
Current issues: -HyenaDNA 450k officially needs a GPU? -Ch38.p14 used as reference genome with GenomeKit