The global health crisis triggered by Covid-19 has underscored the urgent necessity for developing effective therapeutic solutions against emerging viral threats. One remarkable advantage that artificial intelligence offers us is the capability to rapidly adapt our strategies in real-time—provided we can maintain pace with viral evolution and access appropriate datasets.
As with novel medical challenges, data collection often lags behind the immediate need, while viruses demonstrate no such delay in their relentless spread. This creates a significant obstacle as pathogens can quickly mutate and develop resistance to existing medications. This dilemma prompted researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Jameel Clinic for Machine Learning in Health to explore a critical question: How might we identify optimal synergistic drug combinations to combat the rapidly proliferating SARS-CoV-2 virus?
Conventionally, data scientists leverage deep learning algorithms to identify promising drug combinations using extensive datasets for conditions such as cancer and cardiovascular disease. However, these established approaches prove inadequate when addressing emerging illnesses with limited available data.
Faced with insufficient data, the research team devised an innovative approach: a dual-function neural network. Recognizing that drug synergy typically occurs through the inhibition of biological targets (such as proteins or nucleic acids), their model simultaneously learns drug-target interactions and drug-drug synergy to discover novel combinations. The drug-target predictor component models interactions between medications and known biological targets associated with the specific disease. Meanwhile, the target-disease association predictor learns to evaluate a drug's antiviral efficacy by determining viral yield in infected tissue cultures. Together, these components can effectively predict the synergistic potential of two drugs.
Through this methodology, researchers identified two promising drug combinations: remdesivir (currently FDA-approved for Covid-19 treatment) paired with reserpine, and remdesivir combined with IQ-1S. Both combinations demonstrated potent antiviral activity in biological assays. The research findings have been published in the Proceedings of the National Academy of Sciences.
"By modeling interactions between pharmaceutical compounds and biological targets, we can substantially reduce our dependence on combination synergy data," explains Wengong Jin SM '18, a postdoc at the Broad Institute of MIT and Harvard who recently completed his doctoral studies in CSAIL and serves as lead author on the research paper. "Unlike previous methods that utilized drug-target interactions as fixed descriptors, our approach learns to predict these interactions from molecular structures. This presents a significant advantage, as a substantial portion of compounds possess incomplete drug-target interaction information."
Employing multiple medications to enhance therapeutic efficacy while minimizing adverse effects represents a standard approach for treating cancer, cardiovascular disease, and various other conditions including tuberculosis, leprosy, and malaria. Utilizing specialized drug combinations can critically mitigate the serious—sometimes public—threat of drug resistance (consider methicillin-resistant Staphylococcus aureus, commonly known as "MRSA"), since many drug-resistant mutations are mutually exclusive. A virus faces considerably greater difficulty developing simultaneous mutations to resist multiple drugs in combination therapy.
Significantly, this model's application extends beyond a single SARS-CoV-2 strain—it could potentially be adapted for the increasingly transmissible Delta variant or other concerning variants that may emerge. To enhance the model's effectiveness against these strains, researchers would only require additional drug combination synergy data specific to the relevant mutation(s). Furthermore, the team successfully applied their methodology to HIV and pancreatic cancer research.
Looking ahead, the team plans to incorporate additional data dimensions—including protein-protein interactions and gene regulatory networks—to further refine their biological modeling capabilities.
Another future research direction involves implementing "active learning" methodologies. Many drug combination models exhibit bias toward certain chemical spaces due to limited training data, resulting in prediction uncertainties. Active learning can help guide the data collection process and improve accuracy across a broader chemical spectrum.
Jin authored the paper alongside Jonathan M. Stokes, Banting Fellow at The Broad Institute of MIT and Harvard; Richard T. Eastman, scientist at the National Center for Advancing Translational Sciences; Zina Itkin, scientist at National Institutes of Health; Alexey V. Zakharo, informatics lead at the National Center for Advancing Translational Sciences (NCATS); James J. Collins, MIT professor of biological engineering; and Tommi S. Jaakkola and Regina Barzilay, MIT professors of electrical engineering and computer science.
This project received support from the Abdul Latif Jameel Clinic for Machine Learning in Health; the Defense Threat Reduction Agency; Patrick J. McGovern Foundation; the DARPA Accelerated Molecular Discovery program; and partially from the Intramural/Extramural Research Program of the National Center for Advancing Translational Sciences within the National Institutes of Health.
