A fine-tuned genomic language model captures nucleotide-level information overlooked by missense variant impact...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Interpreting missense variants is crucial for clinical genomics.
• Current predictors often focus on protein-level effects, potentially missing nucleotide-context signals.
• The utility of genomic language models for this task is not fully understood.

Purpose of the Study:

• To systematically adapt and evaluate genomic language models for missense variant pathogenicity prediction.
• To determine if these models offer non-redundant information compared to existing methods.
• To develop an ensemble model incorporating genomic language model insights.

Main Methods:

• Adaptation of genomic language models using various architectures and training strategies.
• Development of GLM-Missense, a fine-tuned genomic language model.
• Creation of MetaMissense, an ensemble model combining GLM-Missense with established predictors (AlphaMissense, ESM1b, REVEL, CADD, SIFT, PolyPhen-2).

Main Results:

• Variant-position embeddings and multi-species pretraining improved model performance.
• GLM-Missense outperformed zero-shot predictions and showed low concordance with other predictors.
• The MetaMissense ensemble achieved state-of-the-art performance.
• GLM-Missense provided complementary information, particularly related to splice context and gene-level constraint.

Conclusions:

• Fine-tuned genomic language models contribute valuable, complementary nucleotide-context information to missense variant interpretation.
• GLM-Missense offers novel insights beyond traditional protein-centric or annotation-based methods.
• Ensemble approaches integrating these models show significant promise for improving variant classification accuracy.