Machine learning reveals powerful immunomodulators for vaccines and immunotherapy

Small molecules called immunomodulators can help create more effective vaccines and stronger immunotherapies to treat cancer.

But finding the molecules that trigger the right immune response is difficult – the number of drug-like small molecules has been estimated at 10⁶⁰much higher than the number of stars in the visible universe.

In a possible first for the field of vaccine design, machine learning guided the discovery of new molecules that enhance the immune pathway and found a specific small molecule that could outperform the best immunomodulators on the market. The results are published in the journal Chemical Science.

“We used artificial intelligence methods to guide the search of a vast chemical space,” said Professor Aaron Esser-Kahn, co-author of the paper who led the experiments. “By doing this, we found molecules with record-level performance that no human would have suggested we try. We’re excited to share the blueprint for this process.”

“Machine learning is used extensively in drug design, but it doesn’t seem to have been used in this way before for the discovery of immunomodulators,” said Professor Andrew Ferguson, who led the machine learning. “It’s a nice example of transferring tools from one field to another.”

Machine learning for examining molecules

Immunomodulators work by changing the signaling activity of the innate immune pathways within the body. In particular, the NF-κB pathway plays a role in inflammation and immune activation, while the IRF pathway is essential in the antiviral response.

Earlier this year, the PME team conducted a high-throughput screen that looked at 40,000 combinations of molecules to see if any affected these pathways. They then tested the top candidates, finding that when these molecules were added to adjuvants -? ingredients that help boost the immune response to vaccines – the molecules increased the antibody response and reduced inflammation.

To find more candidates, the team used these results in combination with a library of nearly 140,000 commercially available small molecules to guide an iterative computational and experimental process.

Graduate student Yifeng (Oliver) Tang used a machine learning technique called active learning, which combines exploration and exploitation to efficiently navigate experimental control through molecular space. This approach learns from previously collected data and finds potential high-throughput molecules to be tested experimentally, while also highlighting underexplored areas that may contain some valuable candidates.

The process was iterative. the model suggested potential good candidates or areas where more information was needed, and the team conducted a high-throughput analysis of these molecules and then fed the data into the active learning algorithm.

Molecules that rise above the rest

After four cycles – and finally sampling only about 2% of the library -? the team found highly efficient small molecules that had never been found before. These top-performing candidates enhanced NF-κB activity by 110%, increased IRF activity by 83%, and suppressed NF-κB activity by 128%.

One molecule caused a threefold enhancement of IFN-β production when administered with what is called a STING (stimulator of interferon genes) agonist. STING agonists promote stronger immune responses within tumors and are a promising cancer treatment.

The challenge with STING was that you can’t get enough immune activity in the tumor or you have off-target activity. The molecule we found outperformed the best published molecules by 20 percent.”

Prof. Aaron Esser-Kahn, co-author of the paper

They also found several “generalists” -? immunomodulators capable of modifying pathways when co-administered with agonists, chemicals that activate cell receptors to produce a biological response. These small molecules could eventually be used in vaccines more widely.

“These generics could be good in all vaccines and therefore could be easier to market,” Ferguson said. “That’s pretty exciting, that one molecule could play a multifaceted role.”

To better understand the molecules found by machine learning, the team also identified common chemical characteristics of the molecules that promote desired behaviors. “This allows us to focus on molecules that have these characteristics or to logically create new molecules with these chemical groups,” Ferguson said.

The team expects to continue this process to search for more molecules and hopes that others in the field will share data sets to make the search even more fruitful. They hope to screen the molecules for more specific immune activity, such as activating certain T-cells, or to find a combination of molecules that gives them better control of the immune response.

“Ultimately, we want to find molecules that can cure diseases,” Esser-Kahn said.

A team from the Pritzker School of Molecular Engineering (PME) at the University of Chicago tackled the problem by using machine learning to guide high-throughput experimental screening of this vast search space.