Fig. 5

On treatment ctDNA dynamics associate with survival benefit. Kaplan Meier (KM) curves of overall survival (OS) for patients in a training set (n = 240) with ctDNA levels at baseline in patients ctDNA negative (blue, zero mutations detected), ctDNA levels ≥ 64 mean tumor molecules (MTM; median, red), and < 64 MTM (black) (A). KM curves of OS for patients in a training set (n = 190) with ctDNA levels at C3D1 below the assay limit of detection (LOD; < 1 MTM, ctDNA low risk, blue) versus ≥ 1 MTM (ctDNA high risk, red) (B). A machine learning model was built in the training set based on 5 ctDNA features (MTM at C3D1, # of pathogenic mutations at C3D1, change in # of mutations detected from baseline to C2D1, total cell-free DNA concentration at C3D1, and area under the curve (AUC) for ctDNA level from baseline to C2D1). This model predicted the stratification of patients by molecular risk that associated with OS in a hold-back test set (n = 192) (C) and an external validation cohort (OAK, n = 73) (D). Patients in (C) and (D) predicted to be in the high-risk group (mPD, red) have a worse OS than patients predicted to have a molecular response or stable disease (mResp + mSD, blue). Modified from Assaf, ref. [104].; copyright © 2023, Springer Nature