Genomic language models (gLMs): Emerging applications, challenges, and future directions in computational genomics

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Genomic language models (gLMs) adapt natural language processing to decipher biological sequence data.
• gLMs aim to capture complex patterns like regulatory grammar and evolutionary constraints.
• Established methods like alignment-based or motif-based approaches have limitations in detecting subtle sequence features.

Purpose of the Study:

• To systematically review recent studies on genomic, RNA, and DNA language model development.
• To evaluate model architectures, training objectives, biological tasks, and benchmarking strategies.
• To assess the utility and limitations of gLMs in various genomic applications.

Main Methods:

• A systematic literature search was conducted across PubMed, Scopus, and Web of Science.
• 58 primary model development or adaptation studies were included after rigorous screening.
• The review focused on model design, biological tasks, and performance evaluation.

Main Results:

• gLM utility improved with biologically informed designs (e.g., evolutionary alignments, RNA structural priors).
• gLMs did not consistently outperform specialized supervised models or established bioinformatics tools.
• Generative gLMs show potential for sequence design but often require in silico validation.

Conclusions:

• The field of gLMs is rapidly evolving with diverse applications in genomics and RNA biology.
• Broader impact necessitates standardized benchmarks, clearer reporting, and experimental validation.
• gLMs are valuable tools but require further development to surpass or complement existing bioinformatics methods.