Immunomodulators are small molecules that can be used to develop more potent immunotherapies and vaccines to treat cancer.
It is challenging to identify the molecules that trigger the appropriate immune response, as there are an estimated 1060 drug-like small molecules in the visible universe—a number far greater than stars.
Machine learning guided the discovery of new immune pathway-enhancing molecules and identified one specific small molecule that could outperform the best immunomodulators available on the market, potentially setting a first for the field of vaccine design.
Prof. Aaron Esser-Kahn, co-author of the paper and experiment leader, said, They used artificial intelligence methods to guide a search of a huge chemical space. By doing this, they discovered molecules that perform at a record-breaking level and that no human would have recommended they try. They are eager to distribute this process’s blueprint.
According to Prof. Andrew Ferguson, the machine learning team leader, machine learning is heavily utilized in drug design, but it doesn’t seem to have been applied in this way for immunomodulator discovery before. It’s a good illustration of how to apply tools from one field to another.
Using machine learning for molecule screening
Immunomodulators alter the signaling activity of the body’s innate immune pathways. Specifically, the IRF pathway is critical for the antiviral response, while the NF-κB pathway is involved in inflammation and immune activation.
The PME group screened 40,000 possible combinations of molecules earlier this year to see if any had an impact on these pathways using a high-throughput screen. When adjuvants—ingredients that help boost the immune response in vaccines—were added to the top candidates, they found that the molecules increased antibody response and decreased inflammation.
The team employed an iterative computational and experimental process to identify additional candidates, utilizing these findings in conjunction with a library comprising nearly 140,000 commercially available small molecules.
Yifeng (Oliver) Tang, a graduate student, effectively navigated the experimental screening through molecular space by utilizing a machine learning technique called active learning, which combines both exploration and exploitation. This method highlights understudied areas that might contain some promising candidates while also learning from the previously gathered data to identify promising high-performing molecules for experimental testing.
Through an iterative process, the team analyzed high-throughput molecules identified by the model as potential good candidates or areas in which it needed more information. The data was then fed back into the active learning algorithm.
Molecules that perform better than others
The group discovered high-performing small molecules that had never been discovered before after four cycles, ultimately sampling only around 2% of the library. These best candidates increased IRF activity by 83%, suppressed NF-κB activity by 128%, and enhanced NF-κB activity by 110%.
When administered with a STING (stimulator of interferon genes) agonist, one molecule caused a three-fold increase in IFN-β production. STING agonists are a promising cancer treatment and they enhance immune responses within tumors.
According to Esser-Kahn, the problem with STING has been that there is either insufficient immune activity within the tumor or off-target activity. The molecule they discovered performed twenty percent better than the best published molecules.
Additionally, they discovered a number of “generalists”—immunomodulators that, when combined with agonists—chemicals that trigger biological reactions by activating cellular receptors—were able to alter pathways. In the end, these tiny molecules might find a wider application in vaccinations.
Ferguson stated that since these generalists may work well with all vaccines, it might be simpler to introduce them to the market. The idea that a single molecule could have several functions is really intriguing.
The group also determined shared chemical characteristics of the molecules that encouraged desired behaviours in order to gain a deeper understanding of the molecules discovered through machine learning. According to Ferguson, this enables them to rationally design new molecules with these chemical groups or concentrate on molecules that already have these properties.
The group plans to keep going through this process in an effort to find more molecules, and they are hoping that other experts in the field will share datasets to help with the search. They want to identify a combination of molecules that will allow them to better control the immune response, or they want to screen molecules for more targeted immune activity, such as stimulating particular T-cells.
The goal, according to Esser-Kahn, is to identify molecules that have the ability to cure illness.
The University of Chicago’s Pritzker School of Molecular Engineering (PME) team approached the issue by utilizing machine learning to direct high-throughput experimental screening of this enormous search space.