image: thumbnail.jpg
The epigenomics tech stack | Varun Dwaraka | Data Science Hangout
We were recently joined by Varun Dwaraka, Director of Bioinformatics and Principal Investigator at TruDiagnostic, to chat about leveraging DNA methylation and machine learning for determining biological age, the journey from wet lab science to bioinformatics, the tech stack used in epigenomics research, and challenges in handling and analyzing genomic data.
In this Hangout, we explore how TruDiagnostic leverages DNA methylation, a biomarker that changes with lifestyle and environment, to develop machine learning models for predicting biological age and other health metrics like smoking impact. These tests aim to provide proactive health insights, allowing individuals to track trends and modulate their biological age through lifestyle changes. But the real draw of this episode of the Hangout is Varun's willingness to open up and talk about himself and his journey from not-so-great student who wanted to be a musician to finding himself as a PhD student and then joining a start-up (to relax a little, of course!).
Resources mentioned in the video and zoom chat:
π Varun Dwaraka's LinkedIn β https://www.linkedin.com/in/varunbdwaraka/
π methylclock Bioconductor vignette β https://www.bioconductor.org/packages/devel/bioc/vignettes/methylclock/inst/doc/methylclock.html
π Optuna website β https://optuna.org/
π NCBI Gene Expression Omnibus (GEO) β https://www.ncbi.nlm.nih.gov/geo/
π CRAN Task View: Handling Missing Data β https://cran.r-project.org/web/views/MissingData.html
π Human Microbiome Project (NIH) β https://commonfund.nih.gov/human-microbiome-project-hmp
π minfi Bioconductor package β https://bioconductor.org/packages/release/bioc/html/minfi.html
If you didnβt join live, one great discussion you missed from the zoom chat was about the complexities of data imputation in bioinformatics, including the challenges of handling missing data and the nuances of choosing appropriate methods. Data imputation is a hot topic! Let us know below if youβd like to hear more about it in the future.
βΊ Subscribe to Our Channel Here: https://bit.ly/2TzgcOu
Follow Us Here:
Website: https://www.posit.co
Hangout: https://pos.it/dsh
LinkedIn: https://www.linkedin.com/company/posit-software
Bluesky: https://bsky.app/profile/posit.co
Thanks for hanging out with us! π
Transcript#
This transcript was generated automatically and may contain errors.
Hey there, welcome to the Paws at Data Science Hangout. I'm Libby Herron, and this is a recording of our weekly community call that happens every Thursday at 12 p.m. U.S. Eastern Time.
If you are not joining us live, you miss out on the amazing chat that's going on. So find the link in the description where you can add our call to your calendar and come hang out with the most supportive, friendly, and funny data community you'll ever experience. Can't wait to see you there.
I am very excited to introduce our featured leader today, Varun Dwaraka, Director of Bioinformatics and Principal Investigator at True Diagnostic. Varun, thank you so much for joining us today. We would love to learn a little bit about you. If you could tell us about yourself, what you do, and what you do for fun.
Yeah, thank you, everyone. This is the first time that I'm part of this Hangout, and I'm looking at the numbers, and it just keeps getting higher and higher. So if I start going through puberty again, my bad.
But no, so my name is Varun. I serve as the Director of Bioinformatics. I'm actually maybe a little bit of a sheep in a wolf's den. I was a self-taught coder and data analytics type of person because my background was very much in molecular biology. I was trained mainly on the bench, so doing a lot of DNA extractions, PCR, ELISAs. If any of this rings a bell, I know you're paying.
And so as a biologist, starting out, a lot of where I come from is trying to understand the beauty of biology, but now taking it from the lens of data science. I'm a native of San Francisco Bay Area, which I think probably had a huge impact on why I went into the tech route. So anybody from Cupertino, what up? Mona Vista, I represent. And then through that, went to do my undergrad at UC Santa Cruz in molecular cellular developmental biology, and fell in love with bioinformatics there. If people don't know yet, UC Santa Cruz was actually the first institute to go through and build the first human assembly of the human genome. And so a lot of bioinformaticians there actually played a huge role on a lot of that development and getting the bug, and then realizing that, oh, crap, I need to learn data science. And so I had to switch gears to take more of those roles, and then did my PhD at University of Kentucky. And so I am now based in Lexington, Kentucky in biology.
So with all that said, yeah, I'm a biologist, but have spent a good deal of the past, like, I don't know, 12 to 15 years kind of learning, failing, and somehow getting to the place I'm at in data science and using R and Python specifically.
Awesome. Thank you for the background. What do you do for fun?
What I do for fun is, it's going to sound really nerdy. I just like learning, whether it be in, well, a large part of my time was coding, but now it's just like photography. I was a musician in the past before I went into the biology route. And so spent a lot of time, you know, actually working in studios during my undergrad time and even in grad school here in Lexington. So spent a lot of time there. And so, yeah, I love learning, love spending time outdoors. Lexington is beautiful for the gorge, being able to just go to the Red River Gorge hike. Lake life is a thing coming from the coast. I didn't know that that was a thing. So I've been living that a little bit. Yeah.
I don't know anything about Lexington. So that makes me want to go Google and kind of, you know, snoop around and see. It sounds beautiful.
It really is beyond bourbon and horses. There's a lot more here. Coming as a Silicon Valley native, I didn't think much of it, but now, I mean, I actively chose to live here. So yeah.
Awesome. I want to let everybody know, I'm getting a bunch of direct messages. You might be messaging just me when you think you're messaging the chat. So if everybody take a look at the little dropdown and make sure that you are messaging everyone in the group.
What TruDiagnostic does
Alrighty. Well, I would love to ask you some context building questions because I don't understand fully what True Diagnostic does. So what does True Diagnostic do? What are you building?
Yeah, absolutely. So from a large scale perspective, what True Diagnostic is, is that it's a diagnostics company which is leveraging a specific biomarker called DNA methylation. And so using this biomarker, what we're doing is we're capturing different values of your health using different types of machine learning models that we have generated, not only within the aging community, but within our company as well. Our primary focus is looking at aging in general. So we actually have biomarkers that are developed from DNA methylation data to look at your biological age and other measurements that are actually more based on your exposures.
So one way to think about this is, for example, there was a company named 23andMe, RIP, which really started this whole revolution. And I will give them all the props because they kind of got this whole personalized medicine thing down where they were looking at different types of genetic mutations, SNPs, and using that as a way to build a profile for your own, you know, what are your predispositions? The problem with that, it's a great model and I give them all the roses that they deserve, but the problem with that is that a lot of the genetic changes, genetics don't change on a regular basis. And so what we're doing is that rather than looking at something that is static throughout your entire life course, we're focusing on a biomolecule called DNA methylation, which is actually changing as time goes on, dependent on your lifestyle factors, the pollution, the environment that you live in, the stress that you go under. And so by leveraging that, we actually were able to create models that are able to capture things like your biological age, your let's say HbA1c for anybody that's in the medical field, things like your smoking impact, how much does smoking actually affect you? TLDR, it affects you a lot, regardless of whether you're, you know, good for smoking or not, that's not a thing. And also things like, you know, what type of workout actually has the most benefit based on your methylation patterns. So this goes just beyond whatever your general lab panel is that you go, when you go to see a clinician, not saying that that's not important, but this is at now capturing something that is more exposure based rather than it is, you know, something that's reactive. So really touching on more of the proactive healthcare aspect.
TLDR, it affects you a lot, regardless of whether you're, you know, good for smoking or not, that's not a thing.
I wanted to ask one follow up. Can you give us a better understanding of the direct application of what you're building? So how does this, how would this show up for me in the real world as somebody who knows nothing about biology at all as a consumer maybe or a patient?
Yeah, absolutely. So let's say somebody comes in and is interested in this whole hype of biological aging, or just wants to better understand themselves. They can take our tests. What you receive is essentially a value of here is how you are aging, just a one single composite. Some other information that you'll get is, okay, not only from the composite, this is your age based on your liver, your brain, your blood, your kidneys, your hormones, things like that. And also certain measurements of like your pace of aging. Now, these are great. And so let's say you come in and you are like, for me, I'm chronologically based on date of birth, 33 years old, but then I report as a 35. My first emotion is why the heck am I 35 years old? Why am I older? And so in the report, we actually go through and using different types of associations that we've figured out through our R&D department, ways to kind of reduce that age value. And so you take that first test, you undergo a, let's say you take some of those suggestions that are available, like maybe more certain diet changes, maybe a different type of exercise routine, maybe being more mindful about the stress that you're in. You can then maybe modulate those changes and then take another test after the fact. Now, I'm not trying to sell you a test. You could, we also have subscription type of ideas, but it's really, that is what we think is the power of taking this type of a test is the looking at the trends. Even if I'm 35 now and I'm 34 the next year, at least I came that one year back and it's kind of gamifying this whole situation of how do you take that healthcare back and focus on things that are the most important. Now, other places have this same thing. It's not, you can build a biological age test using various different cofactors, but the main point again, is going back to the biomarker being based on exposure. So now you're quantifying an age that is based on the exposures and is readily changeable based on that.
Awesome. Doesn't make me nervous because I feel like I would test as an 80 year old woman.
Well, I'm not, I get it though, because like I come back, I was really being a moderate with my age. I'm actually 38 biologically. And so, the thing though, is are you heading in the right direction? It's one of the things to know and then take steps to advance yourself. And that's what we're trying to empower.
Breaking into biology as a data scientist
I love it. This is very interesting. This is not a topic I've ever talked to anyone about before. And I think that everybody in the chat has some great questions popping up on Slido. So let's go ahead and ask some of those. The first one is from Soumita, and there is an asterisk here, so I'm going to ask this. Since domain understanding is crucial in data science, how do you see people from general data science breaking into a domain like biology or the broader healthcare domain?
Great question, Soumita. That's a fantastic question. It's something that we deal with all the time. Like for example, the word leader to describe me is something that makes me really uncomfortable because I don't feel like I'm a leader. But the thing that makes, I think that really helps within the team build, is that you have someone that is a subject matter or subject domain expert working with the data scientist. In my case, I'll be very honest, my biology trumps more of the data science, even though I am skilled and have applied a lot of these data science skills. But for example, one of the team members on my team, Saif, is an excellent machine learning expert. While he has a background in, let's say, biology, he hasn't really worked with a lot of the aging space. It's really how do you identify individuals within the team and figure out where you fit. For example, whenever we're creating these models, a model is as good as the data that you bring in, so we spend a lot of time on the cleaning side, identifying if the samples really are part of, let's say, if we're creating a biological age predictor, do we use a disease cohort? Do we use a healthy cohort? What are the predispositions for the individuals that have those methylation profiles? And so that's where the subject domain can really come into play. But working with a person that is skilled on the machine learning aspect, you don't have to be two people at once. You can have two people in their zones of genius. And I think that that's where, if you're looking to join a company that has this kind of a domain, really identifying what are the assets that you have. As in, do you have a person that really understands epigenetics? Do you have somebody that is a gerontologist who really understands aging? Do you have a machine learning guy or gal? Just whatever. Those are the things that you really need to figure out.
Awesome. Thank you. And I'm letting people know I am, when a question is being asked or answered, I'm going to highlight it in Slido, so it should show up for everybody, I think. Hopefully it shows up for you like it does on my side. And then I will mark it as answered, so it will go away. But I will try to stick them in the chat as well.
High dimensionality in genomic data
All right, everyone. Thank you so much for all of that context. I have a question from Kylie. And Kylie, if you would like to unmute and ask that, that'd be great. Yeah, sure. Hi, everyone.
So a long time ago, I used to work with microarray data in a biostatistical context. And one of the key issues that always arose was that for this genomic sequencing data, you have tons and tons of genes, but you may only have a few hundred patients. So the number of parameters you're trying to estimate is vastly more than the people who can inform it. So I'm just wondering, what kind of approaches do you use to solve this high dimensionality problem or just ensure statistical validity in the work that you do?
No, this is a fantastic question. And it really drives a lot of the ML side. So first case scenario is the sample size. One of the reasons why I joined True Diagnostic right out of my PhD was the fact that they already were amassing 5,000 to 6,000 microarray data sets from 850K Illumina arrays. And so this array is 850,000 sites of CPGs. Those are the biomarkers across 5,000 individuals. But it was the business model that was able to drive the data analytics. Because as more people wanted to take their biological age test, they consented in us being able to leverage that data to develop more models as time went on. And so specifically, when we develop a model, we look at sample size. And we are developing a lot of these data sets based on the number of samples or even looking at different established cohorts. So for aging, specifically for us, we co-developed a data set with Harvard, with Brigham's Biobank, using blood samples from there. And recently, we also are now incorporating Generation Scotland, because now we're increasing it from about, what, 4,500 samples to adding
another 17,000 of a Scottish descent, which is now increasing the exposure rate. So one way that we're trying to do that is exactly that whole thing of increasing the sample sizes, not from only biobanks, but even from our own internal cohort. Next is feature selection. We spend so much time on feature selection, whether it be using some type ofβ¦epigenome-wide association studies typically use linear regression. But that assumes the understanding that what you're comparing is a linear model, like it's a linear relationship, which in biology is the death of your entire research program sometimes. But it is a good place to start, and we've had some success with improving the features. So that actually stabilizes the model, for example, when we're doing, like, sun exposure versus non-exposure. This is another thing that we're working on, and I'm happy to talk about that. More recently, Optuna, which I think is available in Pythonβthis is what SAFE has been usingβis it actually does a data-driven approach in terms of being able to hyperparameter selection, I think, I believe. And so there's different ways of parameter tuning that we've been using. But feature selection is, like, honestly, apart from sample size, that I think has a much more impact on how those models are coming out.
So I think I might have given more of an answer, but that's really the two things that I can think of off the top of my head that really are driving the success in identifying real, meaningful biological signals. No, that's great, getting the signal from the noise. Thank you so much for all of the fantastic questions that are just pouring in, everybody. They're so good. I'm going to go through and pick some good ones. I really like Emiko's question, actually, that's on top. But don't forget, everybody, you can vote for questions. If you see one that you like, and you're like, oh, I want to ask that too, feel free to vote for it. Emiko, are you available? Let me see if I can find you here on my screen to unmute and ask your question.
Transitioning from wet lab science to bioinformatics
And the context behind why I like this is because I think that Varun is very specifically going to have some advice here from his experiences. I actually connected to one of the previous questions a tiny bit. So I was asked this not long ago. I was also a bench scientist, and I am no longer doing biology in general. But they were asking, you know, they have limited self-taught, they have very small bioinformatics experience, and they would like to enter into industry, into the bioinformatics route. But given that, like, there's so many bioinformatics graduate students with, like, a lot more experience, they don't really know how to position themselves. And so what advice would you give a person like that?
I ask, I want to, like, touch on this just a little bit. You didn't learn programming skills in grad school or undergrad, did you, Varun? I took a Python class and almost failed it in undergrad. I'll be honest. Yeah, I would love to hear a little bit more in this question and answer sequence about how you learned. Yeah, I'm going to be vulnerable a little bit. Yeah.
Thank you. So I will say this. Maybe I was fortunate in the environment that I was raised, but after undergrad, I had a gap year. And a lot of the places, especially in the Silicon Valley, like, I was able to somehow convince a professor to take a chance on me to be a bioinformatics analyst at UC San Francisco. And one of the things that I did mention is, like, hey, I don't really have a lot of bioinformatics background. Like, I know, like, certain terms. I know kind of low charts, but haven't really applied it. And that person, Dr. Sarah Aaron, still to this day, like, I still remember her saying, it's fine. Like, we're not going to pay you until you actually complete a project, but we'll give you all the data and we'll, you know, you can work with our lead bioinformatician. And I, and so this is where, like, for me, what helped was just three months of just slaving away, going, this was before ChatGPT. So, like, going through vignettes on vignettes that, like, R had, and even Python had, Stack Overflow back in the day, or I guess 2015 is not really back in the day, but Stack Overflow was where I did a lot of it. So, a lot of Google searching and working with an actual data set. I think that was the thing that really clicked in my head is that whenever I was doing projects in a Python class, it was just so tailored and I didn't really care for it. But the moment that I was in a lab where the biology really was something I wanted to understand, and I felt like I was actually applying tools that would translate to other places, that really got me motivated to fail a lot. And then the moment that the first piece of code actually worked, like, I remember I was using Top Hat to do some RNA-seq aligning. And the moment that, like, that bash script worked at all, it just, like, that high that you feel was really what drove me to keep learning.
And the moment that, like, that bash script worked at all, it just, like, that high that you feel was really what drove me to keep learning.
And so, you know, that's where, like, I was given the opportunity to, you know, like, the physical challenges and all that weren't really there. So, I do appreciate if people are facing that. But when I went into a PhD program, I knew I was, like, I got the bioinformatics bug, but I specifically told myself I didn't want to go into a bioinformatics program. I wanted to go into a biology lab, because I wanted to understand, I guess, to the last point, like, what the actual question was, this domain, what was the domain I wanted to study. And a lot of professors at the time, and maybe even now, still don't have a lot of bioinformatics students or bioinformatics help. They're really outsourcing to bioinformaticians. And so, I just pretty much, like, convinced another professor, like, hey, I have this background. I want to continue this. I just need data. And he was like, yo, hold my beer. I got all the data you need, which, again, is a specific, you know, example. But, like, I have all the data. I just don't know what to do with it. None of us know what to do with it. And so, you kind of have to go through and, like, dig through the mud and try to figure it out. Stack Overflow, again, really helped, but then also consulting with professors.
I had to look up a professor who was in the College of Medicine, even though I was in the College of Arts and Sciences for biology, who was doing a lot of microarray. He didn't know anything about next-gen sequencing, so I had to find another professor. Luckily, my professor was able to help, you know, he knew everyone on the campus, so, like, that kind of helped. But he just opened the door, you had to walk in. So, to those people, like, you're, especially nowadays with the internet, your sources of information are not that hard to find. You just have to look for it and be, and it's like a chat GPT prompt. You have to be very specific with what you're asking for. And I think that, you know, finding that, especially even here on this Hangout, like, you'll be surprised at how many people are there that actually can help you.
Absolutely, 5,000 percent. Everybody should get together. Sometimes I talk to people who have been joining the Hangout for a long time, and I'm like, hey, have you ever had a Zoom call with another Hangout member? And they always say no. I'm like, what? Why? You see them every single week, you're connected with them on LinkedIn. Maybe go have a Zoom chat. Maybe I am just a very exuberant Zoom chat asker, but I highly recommend it, everybody.
I will add to that, chat GPT nowadays is, like, insane when it comes to code building. And typically, I think a lot of people, and what I found in my experience is that pseudocode is actually, a lot of people understand logic and can actually build out pseudocode. When it comes to syntax, that's really where you get the drop off. And so, I think chat GPT nowadays has been very helpful, even for me, to build out pseudocode and structure. Obviously, you need to, or sorry, to build out syntax as long as the pseudocode is in good format. So, you have the logic trained down, you are clear with what you're wanting. And, I mean, the statistical assumptions are very important, regardless of what, you know, you're doing transcriptional data, epigenomic data, I mean, like, I don't know, SNP data, going beyond that. Sorry, I'm blanking right now. No, it's okay. I don't know what any of those words are, anyway.
Yeah, you get what I mean. It's just, there's so much data, but, like, as long as you understand this, the domain, I think chat GPT can assist with developing a lot of the syntax that really gets you started. I think that's a great piece of advice. And I, honestly, before we had LLMs helping us code, I was asking my students or my code learners to learn pseudocode and to write in pseudocode before they write their code, just because it helps you, your human brain, go through and figure out what you need to do. That is a lesson from computer science that I feel like data science could really, really apply and use, even if you are not using LLMs. I see so many people in my frame right now nodding their head. I must have, like, struck a chord. Everyone's like, yep, yep, pseudocode, pseudocode. If you don't know what pseudocoding is, go Google. Go find some YouTube videos on pseudocoding. They'll probably be computer science, but take that lesson to heart.
What genomic data looks like
All righty, what do we have? Emiko, fantastic question. Thank you for asking. Joey asked a question in Slido that we can probably answer very quickly, just to give us a little bit more context. You've been using a lot of words that I don't understand, and Joey's going to help us figure that out. Joey, are you available to ask your question? I'm going to click the little ask to unmute button here. Great, yeah. Thanks, Libby. And hey, Varun, thanks for hanging out. I have, like, an overly simplistic question. I'm sorry. I wonder, I've never worked with genomic data before. I just wonder, what does your data look like? Like, if you work with data that are in a rectangular data frame, what's in the rows and columns? Like, what are the features? What are the outcomes? Yeah, same question.
No, absolutely. I guess directly from our data, so our matrix is looking like pretty much your, depends on if you transpose or not, your columns are going to be your samples. So it'll be either an individual or, you know, the sample set that you get it from. And then your rows are going to be those individual biomarkers. And so that depends on the array or the platform that we use to quantify the biological tissue type. So most of the time we run blood, but it can be whatever you want it to be. And then, so when you get it out, it's a frame of however many samples to 850,000. I guess now it's almost a million. And then the values themselves are actually going to be from 0 to 1 in decimal format. And it's actually bimodally distributed because when you think about the data that comes out, DNA methylation really exists as a methyl mark that gets added to cytosines in the DNA. So typically, it's either a 0 or 1 orientation. It's either there or not. But the reason why we still have somewhat of a, it's not continuous, but it's, I guess, boarded by 0 and 1. I forget the statistical name for it. But it's pretty much, you have an enrichment of 0 or enrichment of 1 depending on the location on the genome. And so you're essentially seeing a lot more in either side. But then you still see some, I guess, values in the middle at 0.5, which means in the biology terms, it's hemimethylated. And we do experience a lot of these genomic locations showing hemimethylation. And that actually has implications in biology in and of itself. But typically for us, we're really caring about 0s and 1s. And also the reason, and so you can think of this as percent methylation. And the reason why you just don't get a binary is because the way that the platform works, you're actually doing replicates. It's almost like a bootstrap that's actually happening within the eye scan or the sequence. It's not really sequencing, but the scanning. So you actually get multiple replicates of that value at that position. And so you're essentially consolidating all of that into a 0 and 1. Awesome. Thank you.
Also, somebody said that Python is pseudocode. I thought that was hilarious. Is Python pseudocode for C? Joey, thank you so much for that question. I know that you think it's simplistic. I needed all of that context. And I really appreciate you being brave enough to ask it for everybody. You're not the only one. I saw Arsena saying, that's a great question. There were lots of people. I also wanted to point out that in the chat, I think it was Russ maybe had asked, is there a good example of a data set that people could see that's a public data set that would kind of represent this bioinformatics stuff? A few people had given some resources from Bioconductor and other things. If you happen to have any ideas, Varun, that you could share with me, I could maybe find them.
Public genomic data resources
But yeah, yeah, absolutely. Yeah, shout out to Rebecca for pitching methyl clock. That is actually one of the ones that we use internally. Yeah. But yeah, so for data sets, so there's multiple consortiums that have contributed a lot of genetic and genomic data. So gene expression omnibus is something that the U.S. government through NCBI has actually outputted as well. And so there, essentially, you can download because every time you publish a genetic, let's just say genomic or epigenetic type of study, you have to release your data. And so a lot of these university consortiums will release their data and deposit it into gene expression omnibus. You have ArrayExpress as well. One major source, as Rohan said, is also the Human Microbiome Project, if the NIH still has it up, especially for microbiome. You're going to typically be working with 16S sequencing, but I'm pretty sure there are panels there as well. On the European side, EMBL, E-M-B-L, is another one. I think that they have their own version of a GEO, so you can get it from there. And there are a lot of resources. I like GEO, especially during my grad school when I had to learn how to extract data from these sources. It had a lot of documentation on how to do it. And again, Stack Overflow, a lot of people on Stack Overflow really helped on that as well. So that's usually the first place I would go to. Great. All those are
fantastic resources, and I think that some people are going through and linking in the chat. I did try to find Human Microbiome Project, but I think it's down. At least the website's not working. If anybody finds a better link to that, please pop that in the chat for everybody.
That would be fantastic. All right. I'm going to call Rohan in. Rohan, I'm going to ask you to unmute and ask your question.
Can you guys hear me? Yes. Yeah. Great. All right. Hi, Varun. I also work in the bioinformatics, biostatistics space. I typically work with microbiome data. One thing that we've noticed is a lot of biotech companies have failed because they are unable to reproduce their results.
So how does your firm measure and or handle the reproducibility of analysis results?
Yeah. So a lot of the times, so I guess the multiple answers I'll kind of give is, in the context of our results, like let's say we do an interventional trial, we publish it immediately. So we use bioRxiv and if it gets accepted, whatever journal. Also for those, we are able to release all of our source code. The only thing we cannot release as of yet is our models because ultimately that is kind of what's driving our business.
In the past, I mean, in the academic sector, even the models and everything was released. So at least in terms of reproducibility from the interventional and research side, that's kind of how we're handling it is just by releasing everything. And then hopefully, I mean, we try to advertise, but hopefully somebody comes in and tries to actually use the data. Because again, we have to submit the data to GEO, download it, use the source code. If there's any discrepancies, that's a publication in and of itself that somebody can use.
So that's on the, I guess, more on the research side. In terms of our internal firm, again, the reproducibility is on multiple things. The model will stay consistent. Really, the source of variation comes from the platform. So that is very dependent on, for example, the lab side of things, making sure that they're hitting, let's say, ICC correlations, interclass correlations between, what was it, not, well, technical replicates, really, and also biological replicates, making sure that they're higher than 0.9, at least 0.96 in terms of these replicates, making sure that their quants are also highly repeatable as well, and then only using standardized pipelines. So for example, we haven't recreated on the data side how to pre-process methylation data. We use MinFi.
And then dependent on that, we also have done a lot, we follow a lot of the academic suggestions. There was an entire paper on using this particular normalization pipeline versus this one versus this one, and we went through to make sure that that's there as well. And reproducibility, we will check because any of the papers that have those normalization pipelines, ICC has been a very good correlation coefficient that we're trying to look into. So those are the ways that we touch on the data side and the lab side.
And then also, yeah, if anybody is wanting to do a research study with us, we're very collaborative because sometimes people, if they just want to access our models and they just want to run it, we're able to just run it on whatever, I think mostly Epic V2 or even MSA data, so from the methylation side. So I imagine that other companies as well, hopefully, are doing the same thing in the microbiome sector as well with 16S, which we couldβ¦or maybe there's also certain panels. Are there certain microarray panels for microbiome data, or are you guys usually using more targeted sequencing?
Oh, veryβ¦hold on one second. Rohan, I could have you unmute again. Okay, thank you. I'm sorry, Varun, can you repeat your question? Yeah, yeah. Are you guys using mostly like 16S sequencing, like is it sequencing data, or do you guys actually have a targeted, like array panels?
We're typically using 16S data.
That's all right. That was a great question, Rohan. Thank you so much for asking. Yeah, thank you, Varun. Yeah, of course.
Tech stack overview
Yeah, I love it. And while we're on this sort of topic of highly technical questions, I have seen a couple of people putting lots and lots of resources in the chat for data. Can we talk a little bit more about tech stack stuff? Like what languages do you use, what are the different languages do you use, what types of models do you use? And as far as other things in your tech stack, like I think you use Posit products. How do you use them? And like give us a little quick context builder on that.
Absolutely, yeah. So kind of going back to what Rohan was mentioning, in terms of reproducibility, we're using a lot of bioconductor packages. So R is our native like pre-processing, and also quantification steps. So like MethylClock that was mentioned is actually from a collaborator of ours, Juan Gonzalez, who I think manages that repository. I think it was MethylClock. And so pretty much like R will handle all the pre-processing just because MinFi has been the one that seems like it translates over to a lot of academic labs and all that.
But the moment that it gets into, let's say, the beta matrix format, which is that format that I was describing, I think it was Joey that asked that question, then that's where we branch off. So if, let's say, we're doing a lot of what I handle are actually the academic collaborations. So if we are running these academic research and commercial collaborations, if we're running this on a data set where people are trying to understand how does, you know, people that eat vegan food versus carnivore food, like how do they differ in their epigenetic ages, then we'll stay in R because a lot of the statistical packages are just perfect there.
And so for this, RStudio has been like, that's how we started our entire bioinformatics business, was just using RStudio. And then now we've gotten into Posit, I think the UIUX via AWS, just for a more decentralized connection into S3, and S3 being the source. So decentralization is huge there.
And so for this, RStudio has been like, that's how we started our entire bioinformatics business, was just using RStudio. And then now we've gotten into Posit, I think the UIUX via AWS, just for a more decentralized connection into S3, and S3 being the source. So decentralization is huge there.
However, for the modeling, we used to use a lot of like CARET, C-A-R-E-T, GLMNet for elastic net regressors, and then I think for feature selection using LIMA to do kind of those, like differential methylation identification, and also things like, what is the other one? There's another one that I'm thinking of, which, mutual information, sorry, yeah. So using mutual information specific packages, and then gradient boosting and all that is typically handled there. I know that tidy models was one that we tried to use, but because of things like, we just weren't able to include it, but I got a shout out to tidy models because that really alleviates a lot of the moment, the time that it takes.
However, Safe, who is now the actual ML guy, has shifted everything over to PyTorch and Keras to do all of the neural net, elastic net, random forest regressors, and light GBM, everything like that. So that's why we've kind of, we still use Posit a little bit there, but I think his kind of attack is SageMaker on AWS, just because we're not running one model at a time. Typically, for example, this one project that we're working on is, we're trying to use DNA methylation as a way to predict metabolite levels, and the results are actually pretty striking, but we're doing this for thousands of metabolites. And if we did it serially, it would, I mean, we'd still be working on it. And so he's been using SageMaker as a way to just kind of distribute a lot of those trainings so that we're not waiting like, you know, decades to have it finish.
So yeah, so that's kind of how the tech stack looks on that end. And then the engineering team uses Python to kind of wrap it all up and integrate it into our sub stack.
Data privacy and security
I have some questions that I would love to get to, because we have so, so many, and we have about 15 minutes left. There are a fair number of questions that are on a similar topic. So I'm going to summarize. That is like, hey, 23andMe thing is happening. So aging data is great. It's powerful. It can be helpful for people. But what privacy precautions are you taking for safeguarding information and making sure that there's informed consent and stuff like that?
Yeah. I mean, this regulation is a little bit over my pay grade, I'll be very honest. But this is a question I had. And so HIPAA compliance is huge, making sure we never store anything on a localized server, because I had to move from HPC, like high performance clusters, at my academic institute to S3. And so S3 actually has a lot of privacy backings that allow us to be able to keep everything there.
Anyone who wishes to, AWS S3, it's essentially their data storage recall, I guess, service, if you will. Yeah. Kylie just asked in the chat, so that's a great answer. Yeah, I saw the chat. It's really hard not to laugh, because there's some funny stuff going in there too.
Look away from the chat room. But yeah, so AWS S3 kind of helps with that. We are also, like anytime that we run through CLIA certification and CAP certification, which is the laboratory type of certification, they also check to make sure that the servers are also protected in order to make sure that patient data is not being, there are no loopholes or anything. So that's been there as well.
And then anytime somebody wants to delete their data, we'll delete it. We'll delete their entry, we'll delete their IDAT files, which are the raw format of these. And then we also have their, if we have their blood samples, we will destroy it as well. Destroying blood samples is very easy. You could do it even when you don't want to.
Handling imputation
Uh-oh. All right. There's probably some stories there. I have a question from Noor. And Noor, it looks like part of your question has already been answered, but you have a second question that I think would be great. Let me see if I can find you to make sure you can unmute.
Can people hear me? Yeah. Okay. Awesome. Let me look up my question again so I remember what I said. Okay. Okay. I'm assuming the second to me talking about the secondly question, which is imputation is always a complex thing. I did a master's in statistics. So I know like that can be your whole entire dissertation, but how do you handle imputation, if at all, in a decision medicine context?
Oh, imputation is the bane of our existence. That's the worst question you wanted to hear. I'm also a bioinformatician, so I understand. No, it's- I picked the spicy question. I did it on purpose. No, I love this. This is exactly how, no, this is a fantastic question. And imputation has been the bane of our existence within the bioinformatics department.
Okay. So imputation is essentially, if you have missing data, in our case, if a probe is not as reliable in quantifying that CPG site, or let's say, I don't know, metabolite level or whatever, you have a certain detection that has to pass. If it doesn't pass it, then it just goes as missing. Or, you know, sometimes you just get missing data. And so, yeah, I guess everyone's describing it.
But how we handle imputation is, so initially, if a data set has more than, I mean, like commercially, if a data set has more than 85% missingness, we will just rerun the sample. That's actually one of the best things to do, especially because we have a lab. That's one of the perks of working at TruDiagnostic is that we can actually just rerun the sample and get a new data set. In the cases that we can't, some of the ones that we've used in the past are impute.knn in R, just knn imputation.
But we've also, and actually by we, I have to really shout out Natalia Carreras and Laura, who are on the bioinformatics team, who've actually went through and using our, let's say, I think it was like, what, 17,000 patient data set, actually tracked average methylation levels across sex, age, and I think a few of race and everything, and just replace that with the median of that. Now, is that the best model? No, absolutely not. But this is where, like, if as long as there's like an 85 to 90%, what we're trying to do is just be able to replace it with something that is possibly logical.
Now, in the future, or not even in the future, right now, especially in methylation, individuals are using transformer models to create more, to create these imputations. So, for example, I think it was called CPGPT. That's literally what it's called. I'll put in the bioarchive maybe soon, or I'll just type it in here. Hold up. GPT. There's a bioarchive for this. MICE is not something that I've used in the past. I would love to. And then methyl GPT. So, these are the two that have been coming out. One from Vadim's lab, and then others, Lucas, who works for Shift Biosciences. And they've been doing a pretty good job. But the problem there is, again, array-to-array compatibility. So, it's a very hand-wavy thing of, we haven't figured this out, but we try to use the best that we can, presumably. And so, impute.knn is typically the one that we use most of the time. If not, we always base it off of a reference dataset there. I'm not sure that MICE is something that we've used in the past.
for Jared, rerunning a sample does not ensure reproducibility, though. Well, it's, this is where it kind of comes to making sure that the lab process is, you know, we do test for, if you take the same biological sample, we're hoping that the ICC is the same, like, let's say 0.96 or 0.97.
It's not the best method, but, I mean, it's better than having, like, you know, 20%, let's say, what, 50% missingness. It's just a, I think in statistics, it's like, you just are trying to minimize the error as much as possible. That batch correction is something, like, we're really well aware of how to deal with versus, like, imputation, which is a whole other, basically, I mean, basically, for context, when you rerun a sample, you are taking into account, you may have batch issues, because when you run a bunch of samples, that's basically a batch, you run another bunch of samples, there may be some technical variation that they may knocked in, ask me how I know I may have residual trauma.
Oh, we're getting vulnerable here, Nour. All right. I will say this. Sorry, I did want to say, a lot of what I was talking about was commercially, but whenever we're running, like, interventional data, we have a lot more control with accounting for variation. So, like, for example, we can calculate the principal components, the first three to five of their control probes between each of the B chips. We can also account we can look at the correlations between, let's say, B chip ID. So, like, we run eight at a time, and so each B chip will innately have some level of variance, and so we can account for that as an adjustment factor using some type of generalized linear model or generalized bimodal, you know, whatever limo runs the assumptions behind that negative binomial distribution model.
So, I think it's a moderated t-test at that point. So, there's a lot more control on the research side where we actually have a finite number of samples, but when we're running it commercially, we just have to account for this a little bit more on the fly.
Biological age questions
Okay, great. Awesome. All right. I'm going to do another sort of summarized question, because I have a bunch in Slido that are sort of on a similar track, and they're all anonymous, I think. So, most of them are questions about biological age, right? So, one question is, like, let's say you're 40. Could your biological age be less than 40? Like, if you lived a super optimal lifestyle, what's, like, the best age that you could possibly achieve when your actual age is 40? And then, also, I think the other question is more, like, have the links between these age measures been validated to correlate to things like quality of life or other positive or negative clinical outcomes?
Yeah, yeah. I think, for number one, we see this a lot. Actually, like, a lot of our team, like, CEO and everyone, like, you know, they take this test pretty routinely, and we see that a lot of the times they're scoring less. For me, I'm scoring higher, but they're scoring less, and I've asked them why. And a lot of the times, the qualitative answer, I guess, that they say is mindfulness, being able to actually, like, reduce stress in a healthy way, which might be exercise and all of that. Now, with all that said, I more am on the interventional side, so I'll maybe spend... So, to answer the first question, yes, there are a lot of ways. That kind of bleeds into the second question, which is, yeah, the reason why we've been using DNA methylation is because what was noticed when it first was developed by Steve Horvath is that the margin of error, he couldn't actually create a model that was, like, 100 percent on the dot for chronological age, because the way he was developing this is essentially chronological age on the left of the tilde, and then putting all of your biomarkers on the right, and then running a penalized regression model and selecting for which of the CPGs were able to predict this with the lowest mean absolute error.
And so, he was trying to minimize it because it would be a forensic tool, right? You could use... Now, we'd get, like, a blood sample and then just be able to run it, and then we get a very precise age measure. But the theoretical limit that he was always hitting a little bit was three, but what was really cool about that, the fluctuation is three years, is that individuals who were scoring higher or lower had higher hazard ratios or lower hazard ratios to different disease outcomes. And this was validated on these data sets, such as Framingham Heart's FHS, I think Women's Health Initiative, I think UK Biobank is another one that was also validated there. And so, that right there was where this became a more clinical, you know, anomaly rather than a more of, like, a, oh, it's just cool to get your, you know, biological age, which really, like, what do you do with it, right? And so, ultimately, for us, we don't care about the number. What we care about is the relationship of that number to the outcome of disease, which is why DNA methylation was our, like, you know, biomarker of choice.
What we care about is the relationship of that number to the outcome of disease, which is why DNA methylation was our, like, you know, biomarker of choice.
And nowadays, we don't actually train to chronological age because it really doesn't make sense. Why do you want to train to chronological age when you're trying to predict biological age? Rather, we've been training a lot of this to things like Charleston Comorbidity Index, which is another score that's clinically developed to look at, you know, I think it was, like, frailty outcome. It's an index for 10-year mortality risk, I believe, paraphrasing. And so, we already have these scores, but, like, the idea of training it to here is to capture those molecular diagnostics, which then, in turn, turns into much more of a relationship to disease outcome.
Career advice
Fantastic. Those are all amazing questions. Some people in the chat have helpfully put some spelling and stuff in there, including Azuka, thank you for the Framingham study. All right. What else can we summarize or ask in our three minutes that we have left? There are several questions that are basically asking for advice, and they're anonymous, so I would love to sort of wrap them up in a career advice question, because there are a lot of questions that are sort of like, what would you have done differently as a student? And I know that you, from previous conversations with you, were you a fantastic student? Did you always think you were going to go get a PhD? Tell me about your life lessons, Varun. Give us some career advice.
I know, man. I mean, my GPA in undergrad was, like, at 2.8, and so, like, it was pretty bad. But I think the, I don't know, I don't think I would have, you know, I mean, like, hindsight is 50-50. I would have loved to have taken more, you know, computer science classes or, like, actually statistics classes. It's not even the computer science. It's more of statistics, because that's really where I fell in love with numbers and understanding numbers. But honestly, no, I don't think I would have changed much, just because the failures taught a lot, and I think that that was the reason why I was able to even succeed at a PhD, which succeed meaning I finished it. And so, like, I see this a lot with new PhD students nowadays, is, like, you know, the brightest, like, the ones that had, like, valedictorians or 4.0s or whatever, like, no knock on them. Like, I wish I was one of them. My parents wish I was one of them.
But, like, you know, there's something about failing a lot and being comfortable in that failure that really allows you to kind of finish that degree. And I think that, you know, the confidence then gets instilled of, like, you know, what are the fundamentals? Where is the starting point, and how do you kind of expand upon that starting point? And I think that's helped a lot in my current position, because when I came to True Diagnostic, the bioinformatics was not developed at all. Like, we didn't actually have a division for bioinformatics. And so, a lot of the times, like, I had to learn what the hell AWS was. I had to figure out what modeling was, like, in the context of, like, how do I actually deploy GLM net? And then later, it's, like, how do I make this more efficient, and then be able to do a little bit more of a reproducible aspect? So, I think the failures allowed me to be more comfortable when, like, I loved that background of, I believe it was Luke, where everything's on fire, and you're still, like, okay, this is fine. Like, so, I think it really, it all just kind of comes down to, like, that. So, yeah, accept failure as a really good teaching tool.
But, like, you know, there's something about failing a lot and being comfortable in that failure that really allows you to kind of finish that degree.
Wonderful. Thank you. Yeah, there, Vern, you said something hilarious to me the last time we spoke, which was something like, yeah, I was kind of, like, burned out after my PhD, and I thought maybe going to join a startup would, like, get me away from that and help me relax. I was like, you thought that joining a startup would be relaxing?
I was just convobulated. I also thought about moving to Nashville and, you know, just being a barista at some, like, you know, trendy, like, coffee shop and then trying to make it into the music business, and so. I think you're among a lot of other musicians here. Everybody who's a musician, raise your hand really quickly in the chat or in your camera, because for some reason, musicians and data science go together. Well, Vern, thank you so much for joining us. I had a wonderful time in this talk. I would love to chat with you again, and I'll probably ping you on LinkedIn later to see if I can ask you some more questions and get some more resources. Thank you, everybody, for joining us today. I'm going to hang out for another few minutes and leave the chat open, so anybody can save it if they want to. There's a little three dots in the top right-hand corner of your chat screen where you can save the chat if you want, so you don't lose all those great resources. And next week, we have Harvey Lieberman, Associate Director of Data Science at Novartis. You're not going to want to miss this one. If you liked talking with Farouh and you are going to love talking with Harvey, so please join us. And, yeah, thanks, Farouh, and thank you so much, everybody. Thanks, guys. It was a lot of fun. Nice to see you all. Thank you. All right. Bye, everybody.