Deep Learning Model for Tumor Type Prediction using Targeted Clinical Genomic Sequencing Data.

Area of Science:

• Oncology
• Genomics
• Bioinformatics

Background:

• Histology-based tumor diagnosis presents challenges in clinical oncology.
• Genomic alterations are highly diagnostic of tumor type, offering potential for improved classification.
• Existing genomic tumor type classifiers often require whole genome sequencing (WGS) or are limited in scope.

Approach:

• Developed Genome-Derived-Diagnosis Ensemble (GDD-ENS), a deep neural network ensemble model.
• Utilized genomic features from a large dataset of 39,787 solid tumors sequenced via clinical targeted cancer gene panels.
• Implemented a hyperparameter ensemble strategy for robust tumor type classification.

Key Points:

• GDD-ENS achieves 93% accuracy in predicting tumor type across 38 cancer types.
• The model's performance rivals that of WGS-based methods but uses more clinically feasible targeted sequencing data.
• GDD-ENS demonstrates utility in diagnosing rare tumor types and cancers of unknown primary.
• Incorporation of patient-specific clinical information further enhances prediction accuracy.

Conclusions:

• GDD-ENS provides a clinically feasible and accurate method for tumor type classification using targeted sequencing data.
• The model can assist in real-time clinical decision-making by providing rapid, reliable tumor type predictions.
• Integration into clinical sequencing workflows can significantly improve cancer diagnosis and treatment guidance.