Abstract
The convergence of natural language processing (NLP) and genomics has given rise to a new class of transformer-based models-genome large language models (Gene-LLMs)-capable of interpreting the language of life at an unprecedented scale and resolution. These models represent a revolution in the field of bioinformatics since they use only raw nucleotide sequences, gene expression data, and multi-omic annotations, leveraging self-supervised pretraining to decipher complex regulatory grammars hidden within the genome. This survey presents a comprehensive overview of the Gene-LLM lifecycle, including stages such as raw data ingestion, k-mer or gene-level tokenization, and pretext learning tasks like masked nucleotide prediction and sequence alignment. We specify their wide range of applications, spanning crucial downstream activities such as finding the enhancer or promoter, modeling the chromatin state, predicting the RNA-protein interaction, and creating synthetic sequences. We further explore how Gene-LLMs have created an impact on functional genomics, clinical diagnostics, and evolutionary inference by analyzing recent benchmarks, including CAGI5, GenBench, NT-Bench, and BEACON. We also highlight recent advances encoder-decoder modifications and the incorporation of positional embeddings, a feature specific to living organisms, which may enhance both interpretability and translational potential. Finally, this study outlines a pathway toward federated genomic learning, multimodal sequence modeling, and low-resource adaptation for rare variant discovery, establishing Gene-LLMs as a cornerstone technology for the responsible and proactive future of biomedicine.