Machine Learning Based Prediction of COVID-19 Mortality Suggests Repositioning of Anticancer Drug for Treating Severe Cases

Abstract

Despite available vaccinations COVID-19 case numbers around the world are still growing, and effective medications against severe cases are lacking. In this work, we developed a machine learning model which predicts mortality for COVID-19 patients using data from the multi-center 'Lean European Open Survey on SARS-CoV-2-infected patients' (LEOSS) observational study (>100 active sites in Europe, primarily in Germany), resulting into an AUC of almost 80%. We showed that molecular mechanisms related to dementia, one of the relevant predictors in our model, intersect with those associated to COVID-19. Most notably, among these molecules was tyrosine kinase 2 (TYK2), a protein that has been patented as drug target in Alzheimer's Disease but also genetically associated with severe COVID-19 outcomes. We experimentally verified that anti-cancer drugs Sorafenib and Regorafenib showed a clear anti-cytopathic effect in Caco2 and VERO-E6 cells and can thus be regarded as potential treatments against COVID-19. Altogether, our work demonstrates that interpretation of machine learning based risk models can point towards drug targets and new treatment options, which are strongly needed for COVID-19.

Keywords: Covid19; Drug repositioning; Explainable ai; Machine learning; Precision medicine.

Fig. 1

Kaplan-Meier plot of COVID-19 patients in LEOSS. The plot shows the estimated survival function according to the well-known product limit estimator, see section “Methods” . The gray area depicts the 95% confidence interval.

Fig. 2

(a) Model prediction performance measured via Uno's C-index on held out test sets (COX = elastic net penalized Cox proportional hazards regression; WEI = elastic net penalized Weibull accelerated failure time regression; XGBSE = XGBoost Survival Embeddings; RSF = Random Survival Forest; DEEPSURV = DeepSurv); (b) model calibration error measured via Integrated Brier Score (IBS) on held out test sets; (c) model prediction performance as function of time on held out test sets with 95% confidence interval, with integrated AUC (iAUC) denoting the mean (standard error) AUC over time. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

Feature importance using absolute SHAP values: (a) top 10 predictors; (b) cumulative influence per feature modality. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4

Partial dependence plots for most influential predictors. Boxplots show the distribution of patient specific hazard ratios per variable category. The red horizontal line defines the reference. The hazard ratio describes by which factor the median lifetime is expected to change compared to reference. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 5

Regorafenib (panels A and C) and Sorafenib (panels B and D) activities measured in different cell lines (Vero-E6 cells upper panels; Caco2 cells lower panels) as percentage inhibition of viral cytopathic effect normalized to Remdesivir as positive control (100%). Cells in wells were treated with SARS CoV-2 virus, and drugs were administered after 48 or 96 h after infection. Subsequently, cells were stained, washed and counted if alive. Some signs of toxicity on Caco2 cells (lower panels) started to surface at higher drug concentrations and this might be the reason for the higher observed variance of triplicates. The slightly negative relative inhibition shown in panel D is caused by plate control differences within plates.