This repository includes an implementation of a Transformer Neural Network for any-order epistasis detection. This implementation has been devised to make use of network interpretation strategies (e.g., attention scores) to identify epistatic interactions and target various AI accelerators (e.g., GPUs, TPUs).
Genome-Wide Association Studies (GWAS) analyze the influence of individual genetic markers on well-known diseases. However, this approach ignores gene-gene interactions (epistasis), which are of utmost importance to understand complex diseases. Therefore, finding new SNP associations has a high impact on our understanding of these diseases, as well as precision medicine, contributing to improve personalized healthcare.
- TensorFlow (version 2.6 or more recent)
Running the Transformer on a single GPU with a synthetic example dataset (fourth order interaction) with 1000 SNPs, 1600 samples (800 controls, 800 cases), splitting the dataset in 6 partitions, combining them in pairs, and selecting the top 5% SNPs:
$ python3 ED_Transformer.py -path Add4.txt -partitions 6 -comb 2 -top 0.05 -device GPU -n 1 The Transformer implementation requires the following arguments:
$ -path PATH a path to epistasis datasets (can be a .txt file, a folder with files, or a zip with files)
$ -partitions PARTITIONS the number of partitions
$ -comb COMB the combination order to merge partitions
$ -top TOP best SNP percentage to report after training (between 0 and 1)
$ -device DEVICE the device to use (e.g., CPU, GPU, TPU)
$ -n N number of devices
$ -sparsity SPARSITY an optional float (between 0 and 1) for the sparsity percentage on the transformer attention modules. Defaults to 0.9.
$ -epochs EPOCHS an optional int for the epochs to train the transformer. Defaults to 15.