Language Modelling Techniques for Analysing the Impact of Human Genetic Variation.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Interpreting human genome and proteome variants is crucial for disease risk analysis, medication response prediction, and personalized health.
• Natural language processing (NLP) techniques are highly applicable to computational variant effect prediction due to similarities between natural languages and genetic sequences.
• Transformer models have significantly advanced variant effect prediction, but limitations necessitate extensions and alternatives.

Purpose of the Study:

• To systematically review language modelling approaches for computational variant effect prediction over the past decade.
• To analyze the main architectures used in these models.
• To identify key trends and future research directions in the field.

Main Methods:

• Conducted a systematic literature review of over 50 language modelling approaches.
• Analyzed various model architectures, focusing on NLP techniques like Transformers and their extensions.
• Identified common themes, advancements, and challenges in the field.

Main Results:

• Over 50 language modelling approaches for variant effect prediction were identified from the past decade.
• Transformer architectures and their modifications are prominent, showing significant advancements.
• A key limitation identified is the lack of standardized evaluation frameworks and datasets, hindering model benchmarking.

Conclusions:

• NLP, particularly Transformer-based models, has greatly impacted computational variant effect prediction.
• Further development is needed to address model limitations and computational efficiency.
• Establishing shared evaluation frameworks and datasets is essential for future progress and benchmarking.