Learning Patient Similarity from Genomics for Precision Oncology

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Precision oncology relies on molecular biomarkers, but many patients lack actionable targets or effective treatments.
• Patient similarity approaches can enhance decision support by analyzing comprehensive tumor profiles and clinical data from large patient cohorts.

Purpose of the Study:

• To develop a deep learning framework for measuring patient similarity using tumor genomic profiles.
• To evaluate the association of patient subgroups and neighborhoods with therapeutic outcomes in breast cancer and pan-cancer settings.
• To assess the model's utility for patients without actionable biomarkers and those with cancer of unknown primary (CUP).

Main Methods:

• Utilized real-world clinicogenomic data from a tertiary cancer center.
• Developed a deep learning model to embed tumor genomic profiles and measure patient similarity.
• Evaluated patient subgroups and neighborhoods for associations with therapeutic outcomes.

Main Results:

• The model identified clinically meaningful patient clusters with known and novel therapeutic associations.
• Derived patient neighborhoods informed therapeutic trajectories more often than expected by chance.
• Demonstrated utility for patients lacking actionable biomarkers and for cancer of unknown primary (CUP) diagnoses.
• Showcased potential for continuous learning and analysis over time.

Conclusions:

• The similarity-based framework translates complex data into actionable insights for precision oncology.
• This approach augments clinician judgment and supports patient-centered decision-making.
• Provides a foundation for a real-time learning decision support model in precision oncology.