Improving the Accuracy of Genomic Variant Calling Through Deep Learning

This project will develop new deep learning approaches to tackle unsolved problems from variant calling (e.g., SNPs and small indels in low-complexity regions with ambiguity). Unlike traditional methods, our algorithms will not only provide the best variant calling quality but also translate well across different application domains (germline/somatic), sequencing methods (WGS/exome/amplicon), and platforms (Illumina/IonTorrent). Meanwhile, our new machine-learning-based implementation will use industry-standard libraries, such as Tensorflow and STL, and target both GPUs and FPGAs for computation acceleration.