Neural network helps identify many more potential targets for cancer drugs
Click Here to Manage Email Alerts
As many as half of cancer-signaling proteins once thought to be resistant to drug treatment due to lack of targetability may, in fact, be treatable, a study conducted at Perelman School of Medicine at the University of Pennsylvania showed.
The findings, published in Nature Communications, identified targetable “cryptic pockets” in these proteins using a Penn-developed neural network called PocketMiner.
Key protein structures in which cryptic pockets likely exist, according to the researchers, include WNT2 in the Jak/Stat signaling pathway and PIM2, an enzyme implicated as a driver of lung, prostate and breast cancer, as well as leukemia and myeloma.
“I hope the field as a whole will revisit several of the proteins we know are essential to various forms of cancer but that we’ve shied away from because the available structural information makes them seem too difficult,” Gregory R. Bowman, PhD, professor of biochemistry and biophysics and bioengineering at Penn, told Healio. “We can reconsider whether some of these ‘undruggable’ proteins are actually viable targets and could be targeted in a way that could help people.”
Bowman spoke with Healio about his inspiration for studying cryptic pockets and what the findings could mean for the future of targeted cancer therapy.
Healio: What inspired you to conduct this study?
Bowman: I am legally blind due to a deleterious mutation, so I have had a longstanding interest in how we restore protein function. This led me to a passion for allosteric communication between distant sites on a protein and a phenomenon called cryptic pockets. The structure of a protein as observed by crystallography looks like a compact globular thing with no pockets for small molecules to bind. However, if you could watch the dynamics, you would see things spontaneously open up and create unexpected or cryptic binding pockets. I have been working on this for the last 10 years, and we’ve made some really good progress using computationally expensive, atomically detailed simulations to find these cryptic pockets. We followed that with experimental tests of whether the pockets exist and whether putting molecules in them can impact function.
We have had some nice successes with a number of viral and bacterial proteins that are important for combating infectious diseases, and have been looking around for where else this could be applicable. One of the things I’ve been challenging my students with for a while is how we could develop an algorithm that you could give a single crystal structure, for example, that would predict if and where cryptic pockets are likely to form, such that we could triage collaborations or prioritize potential targets in the signaling pathway.
Healio: How does the PocketMiner neural network function?
Bowman: We realized a lot of important work is happening in machine learning right now. The basic takeaway is that if you have enough data to train these neural networks on, they often perform very well in real predictions. One of the challenges is that there are very few data points when it comes to cryptic pockets, from an experimental perspective. If you look in the protein data bank, there are 200,000 different structures, but only in about a hundred of these cases are there structures of a protein without ligand and then a protein with the ligand capturing the open state of some cryptic pocket. That’s not much data to train these networks on. However, at this point we have a lot of simulation data, so instead of training on the experimental data, we trained the network to take a structure from some random point in our simulations as input and predict where cryptic pockets are likely to form in the next 50 nanoseconds of simulation. This worked very well on simulation data and has turned out to work well when we pivot and apply the same algorithm to ask, given a crystal structure, where the cryptic pockets are likely to form.
Healio: Does finding these cryptic pockets suggest some of these proteins may be druggable?
Bowman: Yes. We took 10,000 different protein structures that are important in cancer, roughly half of which lack obvious pockets for targeting with small-molecule drugs, according to previous work. We ran our PocketMiner algorithm on them, and for half of the purportedly “undruggable” proteins, we found clear signals that there are likely cryptic pockets. We then ran computer simulations and saw pockets open up at those sites. This was nice corroboration that the neural network is predictive.
Healio: What is next in your research on this topic?
Bowman: I’ve been obsessed with these features for the past 10 years. We’re constantly trying to determine what the biggest hurdle is in making these therapeutically viable. So, we’re shifting a lot of our attention to getting better at targeting these pockets with small molecules. Some of that is happening in my laboratory. I recently started a company called Decrypt Bio that is doing some parallel work in an industrial setting.
Healio: Could this work potentially help with your legal blindness?
Bowman: Yes. It’s still a way off, but my blindness results from a transporter protein that malfunctions due to a mutation. So, one of my dreams is to be able to take dysfunctional proteins like this and design drugs that would restore function.
Healio: Is there anything else you’d like to mention about this research?
Bowman: Our simulation data was all generated with the Folding@home distributed computing environment, which empowers anyone with a computer and an internet connection to become a citizen scientist and help our research by running computer simulations of protein dynamics on their personal computers. We have about 100,000 active volunteers right now, and have gotten as high as 1 million volunteers during the COVID-19 pandemic, when we were doing a lot of work on SARS-CoV-2.
References:
- Meller A, et al. Nat Commun. 2023;doi:10.1038/s41467-023-36699-3.
- Penn Medicine research suggests more cancers can be treated with drugs than previously believed (press release). Available at: www.newswise.com/articles/penn-medicine-research-suggests-more-cancers-can-be-treated-with-drugs-than-previously-believed. Published March 14, 2023. Accessed April 3, 2023.
For more information:
Gregory R. Bowman, PhD, can be reached at Penn Medicine, 3620 Hamilton Walk, Philadelphia, PA 19104; email: grbowman@seas.upenn.edu; Twitter: @drGregBowman.
Collapse
Click Here to Manage Email Alerts