GROVER · Natural Language Processing to learn the language of the Human Genome

Natural language processing (NLP) models trained on text without explicit supervision can have groundbreaking performance. They can develop a notion for grammar, syntax, and semantics, thus learning the structure of language. However, while we have defined the rules in our language, we only have a basic understanding about the linguistics of our genome. In this project, our goal is to treat the human genome as a sequence of text and apply NLP techniques to the human DNA sequence. We will establish byte-pair tokenization to generate vocabulary from DNA sequence and analyse attention maps to see the training relationship between different “words” of the genome. We will then further investigate the language rules using methods from corpus linguistics. Together, this will allow us to explore the grammar, syntax, and semantics hidden in the genome and capture their biological meaning. For proof-of-principle, we will perform several biological prediction tasks with fine-tuning models, built on top of the pretrained model. First, we will take popular genomic prediction tasks to benchmark our approach, such as predicting genome elements, transcription, and precision of genome editing. Then we will add some novel tasks around genome stability using available multi-omics data. Throughout the project we will implement techniques for interpretable learning and strategies to observe, control, and prevent ethnic biases in our approach. We expect for large language models to change how we, as a scientific field, approach genomics data analysis and expect our models to establish how these techniques can be applied efficiently, transparently, and in a bias-reduced way. In addition to general understanding of genome biology, we plan to use our models in the future for technical improvements of data analysis, population genetics, and for translational uses with applications in cancer genomics and genome editing.