GenotypeTensors: Efficient Neural Network Genotype Callers

Abstract

We studied the problem of calling genotypes using neural networks. A machine learning approach to calling genotypes requires a training set, an approach to convert genomic sites into tensors and robust model development and evaluation protocols. We discuss each of these components of our approach and compare four types of neural network training protocols, two fully supervised and two semi-supervised approaches. Semi-supervised approaches use unlabeled data to supplement limited quantities of labeled data. Random hyper-parameter searches identified highly performing models that reach indel F1 of 99.4% on a chromosomes 20, 21, 22 and X of NA12878/HG001. We further validate these models by evaluating performance on HG002, an independent sample used in the PrecisionFDA challenge. We apply GenotypeTensors to evaluate the impact of (1) training with small datasets, (2) training models only with sites inside confidence regions, or (3) training with improved true label annotations. A PyTorch open-source implementation of GenotypeTensors is available at https://github.com/CampagneLaboratory/GenotypeTensors. DNANexus cloud applications are provided to help process new datasets both to train model or call genotypes with trained models.