Exploring Genomic Large Language Models: Bridging the Gap between Natural Language and Gene Sequences

Abstract Motivation With the rapid development of genomic sequencing technologies and accumulation of sequencing data, there is an increasing demand for analysis tools that are more user-friendly for non-programmer users. In support of this initiative, we developed an all-in-one tool called GenomicLLM that can understand simple grammar in the question input and perform different types of analyses and tasks accordingly. Reaults We trained the GenomicLLM model using three large open-access datasets, namely GenomicLLMGRCh38, Genome Understanding Evaluation and GenomicBenchmarks, and developed a hybrid tokenization approach to allow better comprehension from mixed corpora that include sequence and non-sequence inputs. GenomicLLM can carry out a wider range of tasks. In the classification tasks that are also available in the state-of-the-art DNABERT-2 and HyenaDNA, GenomicLLM has comparable performance. Moreover, GenomicLLM can also carry out other regression and generation tasks that are not accomplishable by these tools. In summary, we demonstrated here a successful large language model with a mixture of gene sequences and natural language corpus that enables a wider range of applications. Availability and implementation Codes and data can be accessed at https: //github. com/Huatsing-Lau/GenomicLLM and https: //zenodo. org/records/10695802