Fig. 4: Predicting response to therapy using a composite machine learning model.

a, Schematic of the machine learning framework. CV, cross validation. b, Feature importance calculated as the average z-score resulting from dropping each individual feature from the three components of the model and calculating the new area under the receiver operating characteristic curve (AUC). The importance of chemotherapy sequence features have been averaged into a ‘therapy sequence’ row for simplicity. ES cell, embryonic stem cell; TMB, tumour mutation burden. c, Receiver operating characteristic curves for the clinical (dashed) and fully integrated (continuous) models applied on the external validation cohort. The dotted line indicates random performance. FPR, false positive rate; TPR, true positive rate. d, AUCs for models with increasing levels of data integration. The continuous line on the foreground corresponds to the AUCs obtained from the external validation cohorts (filled markers), with bands representing the standard deviation estimated with bootstrap. The filled band on the background corresponds to the standard deviation of the AUCs obtained using cross-validation on the training dataset, with mean values represented by a dashed line. DigPath, digital pathology. e, Potential clinical impact of the pCR model, using data from the external validation confusion matrix (left). Bar plots show the number of patients that would be identified to be chemoresistant using operating thresholds of 0 and 2 false negatives (FN), using either the clinical or fully integrated models, respectively (right). ML, machine learning; NAT, neoadjuvant therapy.