It’s a question both sensitive and crucial, and a personal one for Coleen Murphy, Ph.D., Professor of Genomics and Molecular Biology at Princeton University: How long can a woman have children? The answer is distinct from menopause, but begins when egg quality starts to diminish as reproductive aging ascends. And that answer is unique for each woman.
“Any individual woman wants to know, ‘How long can I have kids?’” Murphy said during the DOC 2025 session, Using Machine Learning to Build a Reproductive Aging Clock. “So we need that information at an individual level, not a population level.”
Analyzing AMH protein levels from ovarian follicles doesn’t answer this question, as this indicates egg supply — not their quality. Murphy’s team turned to data from RNA sequencing, which can reveal insights into the biological age of reproductive cells — and therefore fertility. But RNA is too dynamic, she noted, with variables that could change daily. Next, they turned to DNA methylation samples, generating what she called a “rudimentary” clock using machine learning to develop a tool sensitive enough to help women follow their reproductive timeline.
“If they got this information every year, they’d know if [they’re] rapidly reproductively aging as well,” Murphy added.
You can hear more about their progress in our video or read our lightly edited transcript below.
TRANSCRIPT:
Coleen Murphy, Ph.D.
I want to thank John and Jordan for inviting me to tell you a bit about the work that we are doing in my lab to try to understand reproductive aging. We’re using reproductive aging, tools to try to understand what how can we make life better for women? And I just want to point this out. For me, this journey was actually started as a personal thing because I became interested in this in about 2003 when I realized that even and I was already working in Cynthia Kenyon’s lab and thinking about longevity quite a bit, and I realized that even if someone gave me some sort of treatment that could allow me to live to be 150, I was still going to have to make some very important life decisions with the next 2 or 3 years. And that’s because, women’s reproductive aging is really the first human sign of aging. And I just want to make a distinction. This is distinct from menopause, which in the United States average onset is at 51 years.
Whereas what I’m interested in is how can we help people understand what’s happening early with reproductive aging? And that’s about a 10 to 15 year difference there. It’s really what we understand now is due to declining egg quality not to egg numbers. Okay. This is the kind of the question we’d like to help people with.
We know that right now there’s no noninvasive long term diagnostic for reproductive aging. But any individual woman wants to know, when am I? How long can I have kids? Right. Not what’s the average. We need that information at an individual level, not a population level. We need it on the years timescale. Not like a few weeks or a few months, which even the things that exist is really that’s what we we are offered.
Furthermore, we don’t know what rate that we’re experiencing reproductive aging at. This is what led, me and lead in my lab to try to ask a question. Can we find markers that indicate reproductive aging? That’s the question we want to ask. Many of you will know about AMH, which actually is a standard of care, particularly for IVF clinics and actually is very useful when you’re thinking about egg retrieval and success of implantation and things like that.
But in this study, for example, in 2017, even when they looked at different levels of AMH, they thought there’s there’s actually really no distinction, no difference in that prediction about how soon someone will become pregnant. That tells us that AMH is, not informative for this question. We’re asking reproductive AG.
We got very interested in trying to understand well what’s actually happening with age that affects egg quality. I’m just summarizing a lot of work from, different labs, including mine, where what we discovered was when you compare young oocytes or eggs with old ones, what you see is pretty much everything happening. You lose the ability to regulate cell cycle, take care of your chromosomes, repair damage to your DNA.
That’s not surprising. We kind of know. But it also tells you it’s not just one thing. Kind of everything declines with age. That’s not really helpful for us in trying to understand, how to fix things. It could tell us whether someone’s aging or not. But the problem there is that oocytes, as many of you know, are really hard to get.
It’s extremely invasive, and it’s not something you would want to do on a regular basis. It’s not really a great tool for understanding reproductive aging. I see a lot of women in the audience nodding. This is really a familiar problem to us. Let me try to step back and say, well, maybe we don’t look directly at the reproductive tissues themselves. Maybe there’s a systemic marker. Just to give you some insight into how we were thinking about this at the time, a lot of the work that we were doing was showing that there’s actually a relationship between how long someone can reproduce or an organism can reproduce, and how long that person lives, or how long that organism lives. There’sinteresting data that’s come out showing that this post reproductive lifespan, actually, can be modulated, but also that if you look at women who have a later than average onset of menopause, it turns out those women could have kids longer or later in life. Here’s a retrospective study done by Tom Perls back in 1997.
They looked at historical data from women from Boston, from who were born in 1896 and what they were. And this is like all before artificial reproductive technologies, IVF, all these things. They could see that when they compared to the two boxes that I’m showing you. In red are centenarians. These are women who would eventually live to be over 100 versus women who were who died by the age of 73. When they looked at that, there was no real difference early in their reproduction. But the women who had become centenarians were more than almost four times more likely to have been able to have a kid after the age of 40. That tells us that women who can live longer often can have kids later, right? There’s a systemic connection.
In other work I don’t have time to tell you about. We also found that women with extended fertility, and we’re talking about women who naturally had kids after the age of 45, actually had differences in systemic markers, including things like insulin, IGF one signaling. This brings us to our first attempts to really try to understand.
Could we find markers that would allow us to understand reproductive aging in women? This is just a pilot study. We did and only had 50 samples. This is taking blood from women of these different ages. That’s a histogram there. We did simply did RNA sequencing on that blood okay. Right away what you can see I’m not expecting to read any of the data on there. But you can see some red lines. Those are outliers right. One thing was interesting in our data. We took in information from the women as well. These are women who had been there’s before they had been diagnosed, probably with polycystic ovarian syndrome, PCOS. We could actually see it when we look back at the questionnaires that they actually had, many of the indications that would have, told us that. We removed those women from the sample and then just ordered our data by age. You can probably see on the right the ages of the women. It’s sorted from young middle and older women. Between those ages of 25 and 24 and 50, and I hope you can see right away that there’s actually some really obvious differences. You see these red blocks. Right. These are signatures of younger women.
These are genetic signatures of older women. It’s really obvious from these data right. The other thing I hope you can see if you look at that is there some outliers. Right. There’s two women who are 31 who look they make match better with the older women. There’s at least one sample that looks biologically younger.
This right away tells us everyone who’s been told you’re 35 and therefore next year, if you get pregnant, it would be considered a geriatric pregnancy. This may not be that relevant, right? There’s information in the data from the blood that we get. Problem here is that this is nice for scientists. But again it’s not really useful because RNA is really unstable. Actually it may be too dynamic like any one day you may have differences that may not be that informative. We were trying to think about what could we what else could we do. This was 2015. This is kind of driving along in my lab where we do a lot of other things, but that’s around the same time that DNA methylation clocks, which probably many of you have heard about, sort of started to become, come out in the literature. DNA methylation, these are these methylation sites on cytosine that actually basically decorate our genome. One great thing about DNA methylation is that it’s much more stable. So we want to use that as a signal. It’s much easier. Other people like Stephen Horvath and Morgan Levine and others have developed aging clocks. The whole life, aging clocks using these kinds of these kinds of, markers.
We thought, let’s try this ourselves. Fast forward a couple of years. What we want to do is now get a lot of data, and I have arrows at the part of this, reproductive aging that’s changing rapidly. We want a lot of data for those at those time points. What we did was we looked into the literature because there was no way we could get enough clinical samples. We’re not at a medical school or hospital. There’s a ton of, data in literature that can tell us about this. We got, almost more than 500 samples for men and women. These are blood samples, and we’re looking at their methylation sites. We also simultaneously got a smaller number of clinical samples from our colleagues at Columbia.
We use that to build, kind of rudimentary DNA methylation clock. What we discovered from that was that we could find, for example, oh, my markers disappeared, 25, markers that were really informative. If we look at those markers in both men and women and look at the overlaps, there’s 15 that are shared. We can actually distinguish the men from the women in these patterns. What this these little boxes are, these are with age for each one of these individual methylation patterns. You can see they actually are different between men and women. Even though we could develop a clock for each it gives us distinct information. We like this, but we knew that we needed, more data.
We greatly expanded this approach. Now that we know that it works too. Now, looking at more than 1500 samples from the, literature and making sure that it’s a little bit more distributed, and then we use, machine learning. I’m only showing this to you actually. You can see my graduate student, Sarah Dobbins actually knows what she’s talking about. The data that we see there, that’s a weight that’s given to every one of these methylation sites. This is really pretty standard for how our methylation clocks are built. What she does is she uses a 1500 sample. She pulls out 10% first and then uses that data to build a mode, and she does this iteratively.
We can take that that clock that she builds from that model and ask how well does it perform. When she tested that against the 10% that were held out, it does pretty well. You can tell that from the error. The error in a is 2.4 almost 2.5 years. But when we compared against the clinical data that we got from Columbia at the time, it does even better.
We’re actually pretty pleased with this clock as far as it does tells us about chronological aging. Now I’m backing up and saying this is not reproductive aging. I’m putting a little caveat on that because we can’t yet tell. We can only tell that this is a great clock for women within these ages. What we’re doing now, and we actually just got the IRB approved and we’ll be taking samples soon, is getting information from women from an IVF clinic where we can actually have information about whether they have reproductive issues or not, and that will help us tweak these clocks. We can really make it much more, sensitive and hopefully do what we want it to do in the first place.
The punchline, the story is we’re still not quite done, but we’re getting there.Our hope is that we can use this machine learning to, method to basically accurately, develop a reproductive clock. We think that DNA methylation, as others have shown for just general aging, is probably a good approach for that. It’s better than AMH. It’s probably better than mRNA.We’re discussing whether use proteomics or not. But I think proteomics could be an add on rather than the answer, because it could be too dynamic, just like the RNA. We’re hoping that this will give women a tool to be able to ask these kind of questions on a year’s timescale.
Hopefully if they got that information every year, you’d even know if there’s some sort of rapid, that got steeper, you know, your rapidly reproductive aging as well. Since this is not just a systemic marker, we know there’s ties to life span, we would wonder if this actually could have been funded for telling us more about general health span and lifespan.
The work I’ve been telling you about was really, primarily done by First Men, just Rachel Kaletsky, former postdoc, now scientists in my lab, along with students. Those are shown in orange. They’re Vanessa Cota, Joyce Fang, and Jiayi Zhang. Sarah Dobbins has been the real superstar in adding new machine learning methods to what we’re doing now.
We’ve had clinical collaborators at Columbia now, more recently at Yura. We’ve had wonderful funding back in 2008, the NIH actually invest in this before anyone was talking about reproductive health. And after that they kind of like shut off the tap. More recently we got funding from the GCL at the back that was really helpful in doing the second stage of these. Now Princeton Precision Health is helping fund some of that work as well. And, my colleague Trina Cuellar is here somewhere, and she can answer more questions about the company that was seen before. With that, I’d be happy to take any questions you might have. Yeah. All right.
Audience Member
Thank you. Quentin Cameto, I work with a company that does AI machine learning for fertility clinics. Briefly. I’m just curious what you make of kind of more experimental therapies that are being developed right now, namely, like ovarian tissue preservation, ovarian tissue, cryopreservation for others. Okay. So first of all, I’m not an MD, so I’m not going to put it basically.
Murphy
I think we would be interesting to take our clock and then reapply it once we have some of those therapeutics and actually be really interesting to see what happens when patients have been treated with things like metformin and other things that have been shown to reverse PCOS. I think that that was part of the testing part that we’d want to do using our clock first. Thank you.