Using machine learning to predict subclone evolution and response during chemotherapy

End of project report

Main messages

• Breast cancer is the most common cancer in women in Wales; outcomes for metastatic disease remain poor due to treatment resistance.
• Circulating tumour DNA (ctDNA) provides a minimally invasive biomarker to track tumour evolution in real time.
• Adaptive therapy - adjusting treatment based on tumour dynamics rather than continuous maximum tolerated dose (MTD) - may delay progression but requires reliable biomarkers such as ctDNA.

What we did

• Collected baseline and progression ctDNA samples; sequencing pipeline established at the All Wales Medical Genomics Service.
• Developed a mathematical model of subclonal dynamics and ctDNA variant allele frequency (VAF) changes.
• Built a virtual cohort of 500 synthetic patients to compare adaptive therapy vs MTD and generate predictive features.
• Early outputs presented at international meetings and published; findings leveraged into national policy and leadership roles.

Key insights

• Recruitment slower than planned, but modelling + virtual cohorts mitigate sample size limits.
• Adaptive therapy can prolong tumour control in specific profiles compared to MTD.
• Early ctDNA dynamics (e.g., VAF slopes) show promise as predictors of treatment benefit.

Next steps (to March 2026)

1. Complete ctDNA sequencing of all samples.
2. Calibrate the mathematical model with real-world data.
3. Retrain machine learning models with integrated real + simulated data.
4. Report predictive performance for subclone evolution and progression.
5. Translate findings into trial design: ctDNA-guided adaptive therapy in breast cancer.