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 12pm US 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.

I'm really excited to introduce our featured leader today. It's Chris Gates, Director of the Bioinformatics Core at University of Michigan. Chris, it's so nice to have you here today. I would love it if you could introduce yourself, tell us a little bit about what you do, and I will also ask you what you like to do for fun.

Thank you, Libby. It's great to be here. So yeah, I work at the Bioinformatics Core at the University of Michigan. Let me just explain that. Bioinformatics means different things to different people, so it's useful to unpack that a little bit. For us, bioinformatics lives at the intersection of three different domains. One is molecular biology, which you can think of for our purposes as DNA, genes, proteins, but also statistics, and finally computer science and software engineering, where those two overlap, AI now.

Really, what we do at the Bioinformatics Core is that we work with biological researchers, and they come to us with sequence data or proteomics data, and we help them identify patterns in DNA, then try to help them interpret what those patterns mean. So we do that a couple different ways. One way is that the algorithms that people use to analyze DNA are constantly changing, so we're always hip-deep in literature review to try to figure out what the best practices are and apply those and also share those out to the research community. Another way that we help people is that we have a series of workshops that are focused on bioinformatic analysis, and the deal there is that a lot of bioinformaticians, they might have some computational chops, but it's a steep learning curve, so we try to give them an intro to some of the basics there so they have more independence.

Then the final way we help them out is that we will do analyses on their behalf, and that happens a lot because biologists come to us and they have a deep expertise in their domain, so they really know a lot about diabetes or the human retina or adenocarcinoma, but they don't necessarily know all of the idiosyncrasies behind the informatics and all of, frankly, kind of strange data types and algorithms that can be a little bit wonky, and even if all those things work, they typically don't have the deeper training in how do you wire this up and execute it at scale in a computational reproducible way, so we help them with that.

Getting into bioinformatics

Yeah, that's a great question. I said at the outset, there's the notion that bioinformatics sits at the intersection of molecular biology and statistics and computer science and software engineering, and I feel like that's a useful kind of paradigm, if you imagine that three-circle Venn diagram, that in general, not a lot of bioinformaticians just kind of emerge from the center of that Venn diagram. What they do is they come in, and a lot of times, bioinformaticians will come in from the biology side, so that they're on the bench, they're working, they're running experiments, and then at a certain point, they start getting involved in some of the analysis, and that just kind of continues to build, to a point where they're like, I would rather be working on the computer than working on the bench.

But that's not the only way to get there. There are people that are coming in from the stats side. I personally actually came in from the software engineering side, which is, it is not a well-traveled trailhead. But yeah, the short version of this is that in 2007, I hooked up with a biotech. It was building a giant data mining platform based on microarray data, and I just kind of got the bug for the molecular biology. I couldn't get enough molecular biology, and I was a trade software engineer at the time, and they were like, why do you care about this stuff? I'm like, I don't know, but it's so amazing, isn't it? It really is this really rich, deep, complex domain.

At the time, I remember a guy explaining to me, he was trying to explain a concept that has to do with how the body translates DNA into individual genes. This is called translation. He was explaining this, and he kept getting into more and more detail. At a certain point, I just stopped him and said, Dan, that cannot possibly be the way it works. It's just too complicated. He smiled at me, he was really nice. He was like, Chris, it's not optimising for you to be able to understand it. It's this way because it makes more. That's the only reason this works. That was scary to me. It was like, oh, this is turtles all the way down. I might never understand this.

I've spent the last 17 years trying to pick up all the molecular biology that I didn't start out with.

Balancing continued education across three fields

I think that is also a really good question, because one of the kind of false aspects of this metaphor that I just drew with my hands with the three overlapping circles is that those circles are actually the center of the overlapping Venn diagram. The way I draw them and the way you imagine them, they're very static, and they're very kind of regular shapes. In real life, they are not regular shapes. They are like this bubbly, effervescent mess. Balancing in the center, it's not a flat surface. You're always having to maintain the balance in the center of those three.

I will tell you that in the last couple years, we have had a couple of interesting shocks to the system. To do this kind of analysis at scale, an important thing that we have to do is we have to put all the analyses in some kind of an automated workflow automation system, like Snakemake or Nextflow. If you're not familiar with those, those are very cool tools. We started with Snakemake, and then the Nextflow kind of community of developers, they kind of eclipsed the offerings in Snakemake. We said, okay, well, we kind of have to transition to Snakemake, or sorry, we have to transition to Nextflow. That meant we had to learn a whole new workflow automation system, which, I don't know, the first one's the hardest, but that was a challenge.

I think you're right in pointing out that this is a dynamic balance that you have to attend. If you want to master something, I would recommend ancient Latin or ancient Greek. Bioinformatics is not that. You're always one meeting away from looking kind of ignorant about something. That's the trade-off. It's like, do you want to look smart, or do you want to learn something? If you're okay sacrificing the looking smart part, you can learn a ton at every consultation that we have.

You're always one meeting away from looking kind of ignorant about something. It's like, do you want to look smart, or do you want to learn something? If you're okay sacrificing the looking smart part, you can learn a ton at every consultation that we have.

I'll also say that the diversity of the projects that we get are constantly kind of testing where we should go next. We support a lot of different sequencing platforms. The sequencing platforms themselves keep changing. There's a new sequencing platform that we're standing up right now called Ultima Genomics. The problem is that they generate such large volumes of data that we have to use our GPU high-performance clusters. We're so on the edge of what they're able to do that the techs at Ultima Genomics cannot fully answer the questions that we're putting to them. Now we've got to learn how GPUs are being allocated on this high-performance cluster. Anyway, yeah, it is a constant balance across what you're learning.

AI's role in bioinformatics

It's an interesting question. I just want to back up and say like AI as a phenomenon is definitely a disruptive innovation. And it's also very clearly disruptive at scale, in a way that people are like, oh, this is like the Industrial Revolution, or this is like the invention of the printing press, or this is like the invention of a circular saw. I haven't found the perfect analogy for this yet, but bioinformatics is definitely disrupted.

I just recently realised that I am what's called a sceptical optimist about AI. I think that AI is important and it can do good things, but I'm not taking anybody's word for it. I am watching what it's doing, and I am testing what it's doing, and I am drawing conclusions as I go. So, sceptical optimist, if you want to sign up for that camp.

It's easier for me to get to this through the software engineering perspective, because people are using AI for software creation in a couple different ways. First off, they're using AI basically as what Google should have been doing 20 years ago. It's basically giving you a better answer that's articulated more clearly. Then, I think further down the continuum, they're using agentic agents to write blocks of code, or passages of analysis.

The concern I have about this is that it's nice when you have code that works. It's even better when you understand the code that works. I feel like for many years, we were really trained in reading code and the importance of reading code as a mechanism of understanding it. Now, I think that there's a class of people that are generating code and bragging about never reading it. We just shipped this, and nobody ever looked at it. Isn't that cool? Software engineering actually has a word or a phrase for code that works that nobody understands. They just call that legacy code. That's the code that somebody else wrote that you inherited, and now you have to maintain it. Now, we figured out a way to make more legacy code faster.

All that said, I do think the stronger swimmers are going to be able to surf this wave. I think it's actually going to be pretty cool. I think that there are going to be weak swimmers that are towed out into dangerous currents. I'm a little bit worried for them. We try not to do that when we're engaging researchers, but researchers also come to us and are like, hey, I ran my data through this AI. What does it say? We're like, okay, well, let's back up. What was your prompt? Let's look at the code together.

I don't know. I think that research is really important. I think that computationally reproducible research is really important. I think an essential ingredient in that is understanding. Frankly, I think we're still on the learning curve to understand how can we deploy AI in a way where we're not compromising our human understanding and the patterns of collaboration that can really advance the domain and still get the benefits of a large language model on that existing knowledge. It's going to be cool in five or ten years to look back on this and realize we've all been using Roman numerals the whole time. This is so much easier now.

Specialist versus generalist

I really like this question. To a degree, I think that's almost a personal decision. I'll also just point out that I had no foreknowledge of how I got to here. Nothing makes sense to me in the moment. Why are you doing this? Everything makes sense in retrospect. I just feel like I've been in this beautiful canoe ride down this beautiful river and facing the wrong way the whole time. I'm a little hesitant to give you any advice on how to steer, because again, I'm generally facing the wrong way.

For me, I really like the interdisciplinary aspect of bioinformatics. I feel like I'm often at the interface of a couple fields, and also at the interface of people, in the sense that there's the people at the high-performance compute environment here that are like, oh, well, we have these kind of GPUs, but we don't have those kind of GPUs. Almost nobody really understands when they're talking, and I can step in and be like, okay, I think what they're trying to say is, which is probably a rude thing to say if they're actually on the call, but that interface, that translational aspect of, I understand the informatics, you understand the molecular biology, and together, we can have a super interesting conversation.

To get to your question, for me, I felt like I really had a deep background of computer science and engineering, so I had been scrambling to learn the statistics and to learn the molecular biology. And I found that those things just kind of had their own wellspring of enthusiasm. So I was just chasing, I was chasing what felt important and felt good at the time.

I think that if you look at the Drew Conway Venn diagram, having exposure to the different petals of that Venn diagram, I think it's super useful because it gives you the flexibility to have those different conversations and to kind of move among those different activities. Where you actually land and where your sweet spot is, this is, I think, a little bit more personal. The flip side of that is, though, I would definitely get enough exposure in some of those other domains to feel comfortable. There are times where I'm like, ooh, I don't feel like I really understand the linear mix model well enough to explain it to this PI. So that to me is a sign, I need to get over there. I need to go deeper on this because I want to be able to explain this to the PI, sorry, the principal investigator. So for me, it's like, it's almost this weird thing of like, I'm afraid of this. I should go closer to it.

If you go deeper, you're going to have that experience and that depth. But if you stay a little bit shallower, then you're going to have that breadth. I like breadth. That's just where I'm at.

If you're comfortable asking a dumb question, you will open up such an interesting conversation. And it's not just a conversation where you learn more. It's a conversation where the whole collaboration opens up, like, because I will tell you that domain experts come to me and I do not know what they're talking about. And I'll just say like, can you put a few more words around this thing that you said? And they love to talk about it. And so they're like, absolutely. I'm so glad you asked that question. So I think that the willingness or the ability to be a little bit vulnerable just creates such a more satisfying and deeper and productive collaboration.

If you're comfortable asking a dumb question, you will open up such an interesting conversation. The willingness or the ability to be a little bit vulnerable just creates such a more satisfying and deeper and productive collaboration.

Human aspects in the age of AI

Yeah. I mean, so it is, it's really interesting question. Like, like in the age of AI, what does it mean to be a good human? And in my case, what does it mean to be a good researcher or how can you support good research? I'll share a little conversation that I had with Claude, the agent, like not too long ago. I sat down, I was in a meeting and the meeting wasn't going well. And so I started a conversation with Claude and I said to Claude, how can you achieve computationally reproducible research results when you're using a nondeterministic tool like AI? And it said, well, it's harder, but there's ways to do it, but that's not really the problem. The problem is that people over-claim like the impact or the significance of their results.

And I kept going and I said, well, okay, but in a domain that rewards positive results over negative results, how is using a tool that is optimized to offer really confident, comprehensive answers, how is that going to shape the direction of the research and the science? Won't that just reinforce positive over-claiming? And it said, this was the most disarming answer I've ever gotten from an AI. It said, yeah, that's a really good question. This might get worse before it gets better, which I was like, that's such a gangster answer. I was totally unprepared for Claude to be that like vulnerable.

I think one of the problems that I have with AI as a kind of faux collaborator is simply that it isn't a collaborator, that it is not teaching us to be vulnerable. It is not teaching us to be better. It's not helping us to trust other humans more. And so I feel like it's kind of encroaching on the nature of collaborative research, which I think is built on trust. And I think trust is built on vulnerability. And almost never, except for that one moment that I just described, Claude doesn't say, I don't know.

And another thing that Libby and I talked about a little bit earlier offline was this notion, it's called Conway's law, by this guy, Melvin Conway, who was a computer scientist in the mid 1900s. Conway, he observed this thing where basically what he said was all other things being equal, the architecture of a technical solution will mirror the organizational structure that created it. And specifically what he said was the architecture of the solution is going to look like the communication structure of the organization that created it. So if you have like three chief architects and you say, you guys got to go build a compiler, they're going to build a three pass compiler, right? That pattern of like code reflects the thing that created it.

But I do wonder what relying on AI is doing to that pattern. I'm wondering like, is AI creating patterns that are easy for us to consume that fit into the organizations that are easy for us to understand and communicate to other people answers? I'm not sure. I think if I can just give the short summary of that, I rely on people to be people and I rely on people to remind me how to be vulnerable and to build trust and to collaborate. And I'm not exercising that muscle very much with AI. So I rely on people for that.

Career advice

So weird. It's, this is a weird answer, but I think a lot of people on the call might get it. Frankly, I wish I knew more linear algebra before I got into this. And I think that for me, I came into it with computer science and software engineering. So that was the, I had already checked that box. But the molecular biology was almost self-powering. It's like, I'm really interested in this. So I got to learn about this. This is fascinating. The statistics was the hardest thing for me, like in learning the statistics and learning the statistics well enough to be able to explain it to somebody that doesn't understand statistics and maybe even doesn't quite believe statistics.

So yeah, I think that the basic career advice, if you want to go into bioinformatics, you're going to want to take a tour of duty in like all three of those domains and just suss out, like, do I like this? And if you do, then I think that staying in the middle, it's hard, but it can be really rewarding. But if you get in there and you're like, I don't really like molecular biology, you still have a vibrant data science career ahead of you. If you get in there and you find out that you don't like computer science, like there's still that Venn diagram, those three overlapping circles, all of the petals have like, those are vibrant and powerful careers. So if all three are lighting up, then yeah, bioinformatics could be a really cool thing.

The basic career advice, if you want to go into bioinformatics, you're going to want to take a tour of duty in like all three of those domains and just suss out, like, do I like this? And if you do, then I think that staying in the middle, it's hard, but it can be really rewarding.

Well, thank you so much, Chris. I hope you had a good time. I had a great time. Thanks everybody. You can find Chris on LinkedIn, Chris Gates, University of Michigan, and you can also find all of us after the call on Discord. So please get on Discord. We would love to hang out with you there. Thank you for hanging out with us, everybody.