The goal of this project is to build deep learning classification models to distinguish artifactual variant calls from genuine artifacts in long-read sequencing data. The three major steps (corresponding to the three notebooks here) for carrying this out are the following:

1. Processing raw FAST5 files to variant calls which includes:

   • Downloading raw FAST5 data from the Nanopore Whole Genome Sequencing Consortium GitHub repository,
   • base calling to get FASTQ files,
   • mapping to the human reference genome to get BAM files,
   • variant calling to get VCF files, and
   • intersecting with NIST Genome-in-a-Bottle gold-standard benchmarking data to allow determination of which variant calls were correct vs artifacts.

2. Preprocessing of the VCFs to a format usable for deep learning modeling which includes:

   • extracting sequence context of all variants in the VCF files from the reference genome
   • creating a Pandas features-labels dataframe with the sequence context being the feature and artifact/not-artifact being the label
   • splitting the data into train-validation and test
   • encoding the sequences and labels into Numpy arrays for input into the deep learning models

3. Fitting and evaluating various deep learning models including:

   • multilayer perceptrons (MLPs)
   • convolutional neural networks (CNNs)
   • recurrent neural networks (RNNs) e.g. LSTMs