Discovering Synergistic Drug Combinations for COVID with Biological Bottleneck Models
This work addresses the challenge of finding effective drug combinations for new diseases like COVID-19, where data is scarce, by leveraging biological information to improve predictions and enable experimental validation.
The researchers tackled the problem of identifying synergistic drug combinations for COVID-19 with limited data by proposing a biological bottleneck model that learns drug-target interactions and synergy, achieving a test AUC of 0.78 and discovering two novel combinations with strong synergy in vitro.
Drug combinations play an important role in therapeutics due to its better efficacy and reduced toxicity. Recent approaches have applied machine learning to identify synergistic combinations for cancer, but they are not applicable to new diseases with limited combination data. Given that drug synergy is closely tied to biological targets, we propose a \emph{biological bottleneck} model that jointly learns drug-target interaction and synergy. The model consists of two parts: a drug-target interaction and target-disease association module. This design enables the model to \emph{explain} how a biological target affects drug synergy. By utilizing additional biological information, our model achieves 0.78 test AUC in drug synergy prediction using only 90 COVID drug combinations for training. We experimentally tested the model predictions in the U.S. National Center for Advancing Translational Sciences (NCATS) facilities and discovered two novel drug combinations (Remdesivir + Reserpine and Remdesivir + IQ-1S) with strong synergy in vitro.