OS1.7 Plasmatic differential scanning fluorimetry profiles discriminate glioma patients from healthy controls

Abstract

Abstract BACKGROUND differential scanning fluorimetry (DSF) has been recently proposed to be used to perform high throughput biofluids profiling by following protein denaturation. Our objective was to discriminate patients with glioma from healthy controls using plasmatic DSF profiles. MATERIAL AND METHODS We included 78 glioma patients and 44 healthy controls. Plasmas were collected using EDTA tubes and analyzed in duplicate using nanoDSF Prometheus NT.Plex instrument (Nanotemper). The following DSF data were analyzed: protein fluorescence at 330 and 350 nm, first derivation of the ratio of fluorescence at 330 and 350 and the absorbance at 350 nm. Then we ran several machine learning algorithms to differentiate gliomas from healthy controls: Logistic Regression (LR), Support Vector Machine (SVM), Neural Networks (NN), Random Forest (RF) and Adaptive Boosting (AdaBoost). All these methods have been tested using a leave-one-out approach where each datum is used once as test while the other are used to train the automatic classifiers. RESULTS We included 78 patients with a median age of 57.8 years (range, 21.8 - 89.9). Thirteen patients (17%) presented with a 1p/19q codeleted IDH mutated oligodendroglioma, 12 patients (15%) with an IDH mutated astrocytoma and 53 patients (68%) with an IDH wild-type astrocytoma, including 26 patients with a recurrent grade IV IDH wild-type astrocytoma. DSF data were independent from classical prognostic factors or patient characteristics: age, Karnofsky Performans Status, IDH mutation status, 1p19q codeletion, grade, initial or recurrent setting, steroid doses, patient size and weight, tumor size and location. The different datasets of the DFS output were tested independently and in combination as input of the machine learning algorithms. Results were obtained using the tuned best parameters on 157 data. The best obtained accuracy was 95.54% with 2% of fake positives and 5% of fake negatives (algorithm: SVM). Others achieved ≥ 90% of correct classification: LR accuracy was 89.17%, NN accuracy was 92.99%, RF accuracy was 91.72% and Adaboost accuracy was 92.36%. CONCLUSION DSF profiles analyzed by machine learning algorithms could allow us to identify glioma patients from healthy controls with an accuracy of more than 95%. These results suggest that DSF of biofluid could be a useful and non-invasive tool to monitor glioma patients. Further investigations, including longitudinal profile analyses are ongoing.