Machine learning–based liquid biomarker analysis from the NEONAX trial for response prediction to perioperative (PO) and adjuvant (A) gemcitabine/nab-paclitaxel in resectable PDAC (rPDAC).

Abstract

Listen

Translate article

755 Background: The phase II NEONAX trial examined PO and A gemcitabine/nab-paclitaxel (G/nP) efficacy for rPDAC pts with a comprehensive biomarker program. This translational study aimed at identifying liquid biomarker for early prediction of G/nP success in each treatment group (PO and A) and for the combined group using machine learning (ML). Methods: Pts from both study arms were stratified into groups based on Short- or Long-DFS ( n=80 pts total; PO n=42 pts; A n=38 pts). Blood plasma from selected G/nP naïve pts was analyzed by multiplexed ELISA (mELISA, 80-Plex ProcartaPlex Human Immune Response) and mass spectrometry (MS; feature list generation) and clinical data were acquired for each patient. For ML pts were divided into training (80%) and validation (20%) datasets. The Weka-based algorithm WrapperSubsetEval (WSE) with 8 different classifiers was used for feature selection. The best performing (highest accuracy, best ROC-AUC, minimal signature) classifier with the corresponding feature panel was selected by a 10x10-fold cross-validation (CV). Respective panels for all features combined as compared to clinical features and Ca19-9 blood levels were tested for performance via CV and bootstrap aggregating for training and validation datasets (10x with replacement). Results: The feature generation process generated 579 features from G/nP-naive pts (mELISA: 80; clinical data: 99; MS: 400). For response prediction of the whole rPDAC group to G/nP (S-/L-DFS), the ML-based flow identified a panel of 8 proteins from MS (SERPINA1, C1QB, KRT1, C4B, UBB, VCAM1, HPD, LYZ) and 3 proteins from mELISA (Galectin3, IL34, CCL4) combined with a logistic regression classifier. The predictive panel with the best performance for determining PO G/nP success was established using a kernel logistic regression classifier and included 2 proteins from mELISA (CXCL2, IL17A), 4 proteins from MS (C3, IGLV3-25, HLA-B, SHBG), and 2 clinical data (WHO grade, Na/Mg blood level ratio). The best performing biomarker for predicting success in A included 2 clinical data (tumor size at staging, hematocrit) and 1 protein from mELISA (IL22) and was established with a random forest classifier. All panels represented minimal feature signatures (rPDAC: 11; PO: 8; A; 3) and showed a high performance with ROC-AUC (training) > 0.90, ROC-AUC (CV) > 0.85, and ROC-AUC (validation) > 0.90. All panels described were superior compared to similar predictive signatures (with corresponding ML classifiers) for all clinical data or Ca19-9 blood levels. Conclusions: We show that minimal liquid biomarker signatures for early prediction of G/nP success in the NEONAX trial based on mELISA, MS and clinical data can be established by ML. The study shows the potential of ML for biomarker panel development and the value of mELISA/MS for generation of feature lists to use in ML. Clinical trial information: NCT02047513 .