Identification of host gene transcripts by machine learning and their application to predict outcome in Ebola virus disease

Abstract Background In humans, Ebola virus disease (EVD) has a high mortality rate. While viral load is a known prognostic factor, it is not effective in predicting outcomes across all patients. Exploring host transcriptomic signatures may provide additional predictive value. Methods Here, the transcript abundance of four candidate host genes (VCAM1, TUBG1, MS4A4A, ISG15) was quantified by RT-qPCR in blood samples from hospitalised patients with Ebola virus (EBOV) collected during the 2013–2016 West African outbreak. Samples were taken for diagnostic purposes during the acute phase of infection. Associations between target gene expression, viral load, and outcome (survival/death) were assessed, and machine learning models were trained and tested to predict patient outcome. Results EBOV Ct values were significantly higher in hospitalised survivors compared to fatalities. During the acute phase, the transcript abundance of VCAM1 and MS4A4A was higher in fatal cases, while TUBG1 expression was elevated in survivors; ISG15 expression did not differ by outcome. In support vector machine models, EBOV Ct alone classified patient outcomes with 86.9% accuracy, while MS4A4A alone reached 82.6%. Combining EBOV Ct with MS4A4A improved prediction to 91.3%. The best-performing model, integrating EBOV Ct, ISG15, and MS4A4A, achieved 100% accuracy in the test set. Discussion Host gene expression markers, particularly MS4A4A, in combination with viral load, improved outcome prediction in EVD beyond viral load alone inambiguous cases with mid-range Ct values. These findings highlight the utility of host transcript-based assays to inform clinical management and for benchmarking the effectiveness of medical countermeasures during outbreaks.