Machine Learning for Building Personalized Cancer Nanomedicines: Interview with Dr. Daniel Heller

Researchers at the Sloan Kettering Institute and the Weill Cornell Graduate School of Medical Sciences in New York have developed a machine learning approach to design personalized nanoparticle therapies for cancer.

Personalized cancer therapies aim to provide a treatment that is tailored to the genetic makeup of a patient’s tumor. They can still cause side effects, however, when they accumulate in certain off-target tissues. Nanoparticles can help to increase drug accumulation in the tumor, and reduce off-target tissue exposure, helping to increase drug effectiveness and reduce side effects.

This research team has developed a machine learning approach that allows almost any class of personalized drug molecule to be incorporated into a nanoparticle. Their technique involves studying which attributes of drug molecules allow them to self-assemble into a nanoparticle, and then working backwards to design drug molecules that readily form nanoparticles.

Unlike most conventional nanoparticle delivery systems, the resulting new nanoparticles are almost 90% drug, meaning they can potentially pack a larger therapeutic punch. The team has tested their nanoparticle therapies in a mouse model of cancer, with encouraging results, and their work has recently been published in Nature Materials.

Medgadget asked Dr. Daniel Heller, senior author on the study and Head of the Cancer Nanomedicine Laboratory at Memorial Sloan Kettering Cancer Center and Assistant Professor at the Weill Cornell Medicine, some questions about the concept.

Conn Hastings, Medgadget: Can you give us some brief background on how you got interested in this area?

Daniel Heller, Sloan Kettering Institute:  We realized that many of the personalized/precision drugs on the market still have quite serious side effects, and often they need to be given at high doses to have a strong effect on tumors. This inspired us to take a deep dive into developing drug delivery systems that not only target the specific proteins/pathways involved in cancer, but that also physically target the tumor sites as well.

Medgadget: What are some of the major challenges facing personalized medicine and nanoparticle delivery systems?

Daniel Heller: Personalized medicines are diverse—there are dozens of approved drugs and hundreds in development and their chemistries differ widely. The assembly of nanoparticles usually depends highly on the chemistry of the drug and there is a lot of trial-and-error based formulation. Therefore, it would normally take quite a lot of effort to make nanomedicines out of each different precision drug. We thought this issue could be tackled by some computer magic.

Medgadget: So, how does the machine learning approach work?

Daniel Heller: My postdoc and lead author on the paper, Yosi Shamay, discovered a new method to build nanoparticles quickly via self-assembly of the drug into a core, coated by amphiphilic molecules, but we realized that this only worked for certain drugs. To figure out why, we made a ‘training dataset’ out of the lists of particle-forming and non-forming drugs. We then asked the computer which molecular properties of these drugs made them fit into nanoparticles.  Our computational chemist collaborators (John Chodera and his student Mehtap Isik) helped to make sure we were actually learning something meaningful and not just seeing a statistical fluke.  Both Yosi and I were new to machine learning, but we found that it really can help build great nanoparticles. It fits into a brand new field called “nanoinformatics” that is applying computational tools to nanotechnology.

Medgadget: The technique aids in modifying drug molecules so that they can be readily incorporated into nanoparticles. Does this mean that the activity or safety of the drug could be affected?

Daniel Heller: We identified a new set of parameters that one could use to design a drug to fit it into a nanoparticle. As medicinal chemists know, any change in a drug molecule can also change the many other factors, like target affinity and specificity, that must be tested. Luckily, these chemists can often do amazing things to balance out different factors.

Medgadget: How do the new nanoparticles accumulate in tumors? Can they be functionalized using antibodies, for instance, or do they rely on leaky blood vessels in tumors?

Daniel Heller: We found that these nanoparticles can target caveolae on blood vessels. These structures can often increase transport across capillaries and are often upregulated in tumors. We can also modify the surface of the nanoparticle to reach other targets.

Medgadget: Can you give us some insight on your result in vivo to date, and your future plans for this technology?

Daniel Heller: We found that these nanoparticles could target precision drugs to tumors to improve their anti-tumor effects while eliminating their most significant side effects. We are determined to use our new nanotechnologies and machine learning methods to improve the effectiveness of drugs in cancer patients, and we hope to test new nano-medicines in clinical trials soon.

Study in Nature MaterialsAltmetric: 44More detail Article Quantitative self-assembly prediction yields targeted nanomedicines…

Link to full article.