Scientists Use Machine Learning to Pave the Way for the Design of Antivirals
Why are there no drugs that can cure Covid, SARS, or the flu? Why is it that if you have a strep throat, you get prescribed an antibiotic, but with a virus, you are told to sit it out? In short, why are there no antivirals in similar quantity, variety, and effectiveness as antibiotics?
Unlike antibiotics, antiviral drugs are typically designed to target one virus. This narrowness of scope makes it uneconomical for drug companies to invest in developing new antivirals. As a result, non-HIV therapeutics comprise less than 1% of the total therapeutics market, which stands in direct contrast to the dominance of infectious diseases in everyday human experience. While the obvious solution to this problem is to develop antivirals that can be used to treat multiple infectious diseases, finding such broad-spectrum drugs has proven to be a highly elusive goal for the scientific community.
A groundbreaking study conducted in collaboration between the laboratory of Prof. Roee Amit of the Technion – Israel Institute of Technology’s Faculty of Biotechnology and Food Engineering and the group of Prof. Yaron Orenstein from the School of Electrical and Computer Engineering at Ben-Gurion University of the Negev provides a feasible pathway forward to achieving this goal. The study, led by Dr. Noa Katz, demonstrates that a combined synthetic biology and machine learning approach can result in the discovery of molecules which can bind proteins from two distinct viruses.
The traditional method for identifying therapeutics is to apply a low-throughput and labor-intensive screen for molecules that might perform the required function. In contrast, the synthetic biology and machine learning approach seeks to map the “space” of potential interactions so that molecules with the desired properties can be reliably predicted. This is done by first generating a large and high quality experimental dataset from a library (i.e. collection) of various known and suspected potential virus-protein-binding molecules. The dataset is then used to train a neural network to allow it to form a multidimensional mathematical function representing the collective’s protein-binding capability.
Once computed, such a function can then be used in reverse. Namely, it can be used to identify regions of high binding capability, and extract “predicted” molecules not previously tested. These unseen or predicted molecules can then be synthesized
and tested for the desired biological functionality. The researchers applied this approach to first map out the binding space of two distinct coat proteins from two different bacteria-attacking viruses, and then synthesized and validated RNA molecules that were predicted to be at the interface between the two spaces, and which therefore possess both functionalities. This achievement provides the scientific community with a blueprint
for an approach that can be used to identify novel RNA sequences that could potentially become key ingredients of broad-spectrum anti-viral drugs.
Click here for the paper in Nature Communication