The Serendipitous Off-Target Explorer – Researcher Interview with Jason Sheltzer
, by Peggy I. Wang
Targeted therapy—a modern type of cancer treatment promising better efficacy and less side effects than traditional chemotherapy—is a pillar of precision medicine. Targeted therapies, such as pembrolizumab and imatinib, work by taking away a particular protein needed by a cancer cell to grow, divide, or spread.
Developing targeted therapies rests on the ability to identify these proteins needed exclusively by cancer cells, also known as genetic dependencies. But while many targeted therapy drugs have entered clinical trials in the last two decades, only a tiny fraction of them have gone on to receive FDA approval.
Dr. Jason M. Sheltzer is an independent fellow at Cold Spring Harbor Laboratories who set out to explore new ways to understand cancer and seek out new dependencies and new drugs to target them. However, his research seems to suggest that many genetic dependencies reported in the literature are in fact not dependencies at all. I spoke to Dr. Sheltzer about his research and what it may mean for cancer drug development.
Peggy I. Wang: How did you find yourself doing this type of study—discovering off-targets?
Jason M. Sheltzer: Well, it wasn't a straightforward path! My original interest was trying to find genes that are associated with cancer patient outcomes. I.e., what genes tend to be upregulated in very aggressive deadly cancers and what genes tend to be upregulated in benign cancers? The thinking was an upregulated gene might be a very good target for therapeutic development.
We had done some computational work to identify genes overexpressed in deadly cancers, and the next step was to knock them out using CRISPR to see if they were essential for cancer growth.
As a positive control, we chose the kinase MELK—it was highly overexpressed in deadly cancers, it had been reported to be essential for cancer growth, and there was also a small molecule drug targeting MELK that was in clinical trials at the time.
We were really surprised to find that we could hit MELK with CRISPR and as far as we could tell, the cancer cells didn't care! Because this was supposed to be the positive control, for a long time we thought we were screwing our experiments up. But eventually, we came to the conclusion that the earlier research on MELK must be incorrect.
PIW: So your experiments told you MELK was not essential, at least not in the way in which it had been reported. And things took off from there?
JMS: After our study on MELK, we took a step back and asked, is this an aberration that we stumbled upon, or is this perhaps indicative of larger challenges with how cancer drugs and drug targets are characterized?
We went back to the literature, found other proteins reported to be cancer dependencies to study, and after verifying that CRISPR was working properly and knocking out the genes with several different ways, the cancer cells continued to grow.
So our research suggests that it is in fact a common problem among new cancer therapies—this incorrect characterization of drug targets and their mechanism of action.
PIW: Why do you think so many of these dependencies turn out to be false positives?
JMS: Many of these dependencies were initially identified with experiments using RNA interference (RNAi), which was the state of the art for studying cancer dependencies many years ago. So to some extent, it's a consequence of using the best technology that was available at the time. And CRISPR-Cas9 (the newer, more sensitive and specific genetic approach) is giving us different answers.
PIW: One of the genes you found to be non-essential was HDAC6, a purported dependency studied in one of our programs, among others. Is there still hope to learn something about essential function in cancer from these types of situations?
JMS: We did find that HDAC6 is dispensable for cancer cell proliferation in more than a dozen cancer types. Also, ricolinostat (a drug developed to target HDAC6) kills HDAC6-knockout cancer cells. So any anti-cancer activity of this compound is independent of HDAC6.
Learning from mischaracterized drugs is indeed something we’re trying to do. For example, we found that anticancer activity of the drug OTS964 comes from inhibiting CDK11, a member of a very important family of cancer enzymes called the cyclin-dependent kinases. We think that OTS964 is a great tool for investigating the consequences of CDK11 inhibition.
PIW: So what is a more reliable way to confirm genetic dependencies? Is it just to use CRISPR instead of RNAi?
JMS: CRISPR is a really powerful tool for making precise manipulations in cancer cell genomes. But no experiment is flawless, no methodology is perfect. So we try to use independent techniques to demonstrate the same thing to get more confidence that what we’re observing is a real biological phenomenon and not an artifact of the particular way in which we asked a question.
E.g., we use multiple independent guide RNAs (sequences that control where to cut the DNA) that target the same gene, use multiple independent CRISPR systems such as standard cutting CRISPR and CRISPR interference to repress a gene without fully knocking it out, or analyze some of the robust whole genome CRISPR screens that have been published.
Verifying on-target drug activity is the other half of the work that needs to be done, mainly by showing that mutations in a drug's target confer resistance to the drug.
This type of validation was probably first shown by Charles Sawyers in the 90s—I don't want to take credit for making this a gold standard! He was studying leukemia patients treated with the Bcr-Abl tyrosine kinase inhibitor imatinib. Dr. Sawyers showed that instances of imatinib drug resistance could be traced to point mutations in the Bcr-Abl gene, and these mutations blocked the ability of imatinib to bind to the protein.
PIW: CRISPR Is a major improvement over RNAi, but are there shortcomings to CRISPR as well?
JMS: Though powerful, CRISPR does have a few limitations. While CRISPR has fewer off-target activities than RNAi, it still does have some, and these could complicate the interpretation of results.
An important caveat to remember is that with standard CRISPR, you're literally damaging the genome by inducing a double-stranded break in a gene of interest. That genomic damage can have significant consequences, including activating p53—a key tumor suppressor that senses and responds to genomic damage.
Also, ideally you're making a null mutation in a gene of interest. I.e., you're cutting a gene and inducing a frame shift or stop codon. So you go from having the 100% of the gene to having 0% of the gene. But cancer drugs are not surgically removing a gene from the genome. Instead, they are likely suppressing gene activity 90, 95, 98%, which could be biologically very different.
PIW: I imagine there are difficulties to this type of work—uncovering mischaracterizations in the field. Like how many months did you spend convincing yourself your interpretation of your data was right?
JMS: Uh, every month from the time I started this position to today! We are always trying to make sure that that everything is right. I must also say that the co-first authors on the paper, both undergrads at the time of the work, Ann Lin and Chris Giuliano are absolutely brilliant, amazing, talented students!
Yes there are difficulties, but I'm really thankful for the support and positive feedback we get from many people. It’s important to remember a lot of effort, time, and money has been spent characterizing these drugs and drug targets.
We try to be very careful about accurately describing and interpreting the experiments that we've done. E.g., we have specifically tested the hypothesis that the proteins that we've knocked out are essential for cancer proliferation in a cell autonomous manner. That is, we've tested whether these genes are required for a cancer cell to go from one cell to two cells to four cells to eight cells in a petri dish.
It’s possible that some of the genes have other important roles in cancer proliferation, such as angiogenesis (the formation of a blood supply) by a tumor. The knockout experiment we described does not test for that.
So when we say our results challenge what was previously reported in the literature, we can't rule out all cancer-related phenotypes for a gene. It's impossible to prove a universal negative.
PIW: With such narrow interpretations of an experiment, and so many ways for a particular experiment to not work, how do you stay optimistic about finding cancer dependencies and believing that a drug is working in the way that you think it is?
JMS: One answer is that false positives are probably the most common problem, and as I discussed earlier, the best controls are using orthogonal approaches and carefully-designed guide RNAs.
False negatives (where there should be a phenotype but we did not observe it) are much less common with CRISPR, but they can occur. For example, cells could use a translational start site downstream of the lesion that you introduced or use alternative splicing to splice around the lesion. To try and rule that out, we perform western blots using antibodies that recognize multiple distinct epitopes within a protein of interest to try and be sure that there isn't any functional protein fragment.
Additionally, we design guide RNAs (based on work from another group) to target exons in a gene of interest that code for functional protein domains and make it less likely that cells can just splice around the lesion.
PIW: In addition to identifying genetic dependencies and understanding some of these mischaracterized drugs, what else are you looking forward to working on next?
JMS: Our work suggests that some cancer drugs may have multiple cellular targets. Maybe a drug was designed to inhibit protein A, but it turns out that it actually functions by inhibiting proteins A, B, C, and D together—a concept called polypharmacology. And we’re working to see what rules we can uncover or characterize about how polypharmacology works and whether we can leverage it to make better cancer drugs.
PIW: Is there any other aspect of your work that you’d like to talk about?
JMS: I believe that publishing negative results is important. When I published some of my work, some people reached out to me and said, thanks for publishing this, we saw the same thing.
A lot of researchers in academia and industry are sitting on mountains of unpublished, negative data and I think that if everyone were a little more willing to share that negative data and take the time put it into the literature, that could improve our cancer research infrastructure.
PIW: Why is it important to understand the mechanism of action? Is it okay if there’s a drug that works against cancer but not in the way that we think?
JMS: At the end of the day, what matters is how well the patients do and if you have a drug and aren't quite sure how it works but it shrinks tumors and prolongs survival, that's terrific.
At the same time, our goal is to establish a regimen of precision medicine in cancer, where a patient can get their tumor sequenced and then clinicians look at it and say, all right, it has an amplification of this gene, a mutation in this gene, a deletion in this gene, and based on this it'll respond best to this drug combination.
In order to get to a precision therapy regimen like that, you need to have a really precise understanding of how drugs are operating. Drugs that work through some unknown mechanism aren't getting us closer to that goal.
A study found that out of all fields of medicine, clinical trials in oncology have the absolute lowest success rate—97%. I think that a 97% failure rate suggests that we need to approach these trials with a better understanding.