2ND PLACE, ORIGINAL RESEARCH TRACK: DROID: Dual-headed Representation for Omics-based Identification of SARS-CoV-2 Infection and Cell type via Deep Learning

The COVID-19 pandemic has claimed millions of lives globally, demonstrating the need for computational/diagnostic models that can detect infection and reveal cell type signatures at single cell resolutions. Predictive models usually optimize for only one objective, such as making normal versus disease predictions without considering cell type specificity, leaving them blind to many shared transcriptomic patterns. Here, we present DROID, a novel dual head multilayer perceptron that leverages a shared 256-dimensional latent space, splitting into a softmax head for cell identity and a sigmoid head for COVID-19 detection. Trained on 647,366 peripheral immune cells across 126 donors, DROID achieves superior performances in predicting cell types (85.6% accuracy) and disease states (91.5%), outperforming XGBoost and SVM baselines despite learning on skewed cell class data. The disease head achieves an area under the ROC curve and AUPR curve of 0.984 and 0.998 respectively. Using SHAP Analysis for interpretability of DROID, the disease head emphasizes canonical interferon stimulated genes such as IFI27, IFITM3, S100A8/9, while also discovering more novel and under-reported marker genes including CLU, DDIT4, VCAN. We further found that ablating the top 30 disease markers lowers accuracy by an average of 7.6 percentage points (Wilcoxon p < 10 −8 ), demonstrating that the model’s findings are statistically indispensable for accurate disease prediction. Overall, DROID delivers accurate and interpretable multitask classification, discovering both overlooked and novel marker genes for diagnosis of COVID-19 with potential for other pathogens. DROID’s improved diagnosis could stop many pathogens and save lives.