First human pangenome reference and Google’s AI
To contribute to the consortium’s efforts, Google engineers helped develop and apply deep learning approaches to solve genomics challenges. Engineers adapted their open-source tool DeepVariant, which uses convolutional neural networks to identify genetic variants. The consortium then used the adapted methods to improve pangenome analysis techniques and eliminate sequencing errors from the long, particularly hard-to-decode stretches of the human genome.
Google’s DeepConsensus, which uses transformers to correct errors in sequencing instrument data, helped to improve the accuracy of the data used to construct the pangenome. High accuracy is critical for a reference pangenome to ensure that it isn’t a source of error in genome analysis. Using DeepConsensus data, the consortium was able to develop a long-read assembler that achieved a final accuracy of more than 99.999%. You can learn even more about these deep learning approaches on our Google Research blog.
This breakthrough was only made possible through the collaboration of an international community of experts, including geneticists, engineers and ethicists. This demonstrates the progress made through diverse contributions — just like the pangenome itself.