Using artificial intelligence (AI) to model clinical variant reporting for next generation sequencing (NGS) oncology...

Area of Science:

• Genomic Medicine
• Computational Biology
• Oncology

Background:

• Targeted next-generation sequencing (NGS) is crucial for oncology diagnostics.
• Genomic analysis of NGS data is a bottleneck for patient assessment.
• Machine learning (ML) can streamline variant reporting.

Purpose of the Study:

• To develop ML models for predicting reportable variants in clinical oncology.
• To address the bottleneck in expert genomic analysis for NGS assays.
• To improve consistency and reduce inter-reviewer variability in variant reporting.

Main Methods:

• Utilized data from three somatic assays (2020-2023) comprising 1,350,018 variants from 10,116 patients.
• Trained ML models (Logistic Regression, Random Forest, XGBoost, Neural Networks) using 211 variant annotations and sequencing features.
• Employed expert-curated reportable variants as ground truth for model classification.

Main Results:

• Tree-based ensemble models achieved high performance (0.904-0.996 PRC AUC) in predicting reportable variants.
• Model explainability was enhanced through waterfall plots showing feature contributions.
• In-house sequencing assay statistics significantly contributed to model performance, limiting generalizability.

Conclusions:

• Longitudinal NGS data supports ML models for clinical oncology variant selection.
• ML models offer a framework for consistent reporting and reduced inter-reviewer variability.
• Enhanced model transparency through individual variant prediction visualization aids reviewer workflows.