Prediction of human pathogenic start loss variants based on self-supervised contrastive learning

Area of Science:

• Genomics
• Computational Biology
• Molecular Genetics

Background:

• Start loss variants disrupt translation initiation, affecting protein production.
• Accurate pathogenicity assessment is vital for understanding disease mechanisms and clinical genomics.
• Currently, only about 1% of human start loss variants are classified due to data limitations.

Purpose of the Study:

• To develop a novel computational method for predicting the pathogenicity of start loss variants.
• To address the challenge of limited labeled data in classifying genetic variants.

Main Methods:

• Introduced StartCLR, a prediction method integrating diverse DNA language model embeddings.
• Employed self-supervised pre-training with supervised fine-tuning to leverage both unlabeled and labeled data.
• Utilized contrastive learning to enhance the utilization of unlabeled data.

Main Results:

• StartCLR demonstrated strong generalization and superior prediction performance on various test sets.
• The method effectively captures variant context information from multiple dimensions.
• Even when trained solely on high-confidence labeled data, StartCLR maintained or improved prediction accuracy.

Conclusions:

• Integrating self-supervised contrastive learning with unlabeled data effectively addresses the scarcity of labeled start loss variants.
• StartCLR shows significant potential for improving the classification of pathogenic genetic variants.
• This approach enhances the utility of genomic data in clinical practice.