EPISODE 1638

[INTRODUCTION]

[00:00:00] ANNOUNCER: Proteins are nanomachines inside cells and perform the incredible array of tasks required for cells to function. They are composed of a chain of hundreds to thousands of amino acid building blocks. Peptides are similar to proteins but have only about 3 to 30 amino acids. Their smaller size gives them distinct properties that are useful in therapeutic applications. 

Menten AI is using cutting-edge generative methods to engineer new peptide therapeutics and are backed by Y Combinator, Khosla Ventures and others. Patrick Finneran is the Associate Director of Biochemistry at Menten. He joins the show to tell us about the drug development process, handling noisy biological data, building a hybrid team of software engineers and biologists and more. 

This episode of Software Engineering Daily is hosted by Sean Falconer. Check the show notes for more information on Sean's work and where to find him. 

[INTERVIEW]

[00:01:07] SF: Patrick, welcome to the show. 

[00:01:08] PF: Hey. Thanks, Sean, for having me. I'm super excited to talk with you guys and talk a little bit about the biology and the software engineering space. 

[00:01:15] SF: Yeah. Thanks so much for being here. Before we dive too deep, let's have you start by introducing yourself. Who are you? What do you do? And how did you get to where you are?

[00:01:22] PF: Yeah. My name is Patrick Finneran. I'm Associate Director of Biochemistry at Menten AI. Title aside, it's a startup like any other tech startup out there. We all wear many hats. One of my jobs is the biochemistry. Making sure the wet lab testing work is done. But I'm also a developer and engineer for the company. 

How I got here was a very convoluted tale. While I was working on my PhD in biochemistry, I was very much focused on wet lab work and performing those studies on the bench. But I had a large amount of computational work and analysis of large data sets. Near the end of my PhD, Hans Melo, who's the CEO at Menten AI, actually reached out to me. Told me he was starting a company and he was looking for some help on the science side and the biology side. And I offered to give my opinions and kind of talk to him about the work he was doing with quantum computing and machine learning in the space. And a lot of what I've learned from machine learning and quantum computing comes from those initial discussions. Yeah, then joined full-time after getting my PhD. And I've been here ever since. 

[00:02:24] SF: Yeah. That's awesome. I think that one of the really exciting and also sometimes stressful things about being at a startup is that you do get to wear many hats. There's no such thing as swim lanes. I founded a company once upon a time and I probably did everything from writing a lot of the original source code to essentially taking out the garbage and getting snacks for people. I have a lot of empathy for the position that you're in. 

I did once upon a time actually work in bioinformatics. I was primarily on the data analysis and tool design side. But I did a stint in my post-doc work at Stanford University in Stanford Medical School in bioinformatics. And I really enjoy this space of combining computer science and traditional physical sciences like biology. 

But I think probably the majority of folks listening probably aren't – they're not working in this world. I want to start with some basics. The drug discovery process is uh pretty complicated from the little bit that I understand about it. Maybe can you give a brief overview of the overarching stages of drug discovery just to start? 

[00:03:28] PF: Yeah. Definitely. I guess the most shocking thing to bring up with people when we talk about drug development is the actual timeline. Many people are blown away when they hear that most drugs can take about 10 to 15 years to be developed. And, of course, this is over a decade of research. And we can further break that down. 

And I think the FDA has done a really good job at writing their guidelines to help demonstrate this process. And the way they depict it is there's an initial phase where you kind of identify your target proteins. This is what's going wrong in the disease. Or this is just something that we can kind of take advantage of to help treat someone. And that's like a protein or a gene, something very biological. 

The first step of the process is taking that target, identify it and validate that this can be a target and that it's a way to drug someone. The next stage is where a lot of the work takes place for scientists like myself. And that's the preclinical development. This is just finding molecules that are drugs. This is improving certain properties while kind of leveraging those good properties while weighing down the bad properties that we want to ignore. 

It's a lot of guess, test, revise until you finally select your candidate that is safe and is effective that you can bring into clinical setting. This brings us to the third stage where it's the clinical development. This is where we can actually test the drugs in humans. Very high-level safeguards and something I personally try to avoid myself. Much more so for the medical doctorates and the PhDs. 

From there, we have the largest portion for a lot of my friends and everything. This is the FDA review. You kind of have FDA review at a few different points. Before going to the clinic, you need the FDA to kind of approve and say, yes, you can go into a human. But then also after you have your clinical data, large amount of work. A lot of machine learning being brought in here to help speed up those timelines to getting approval from the FDA. 

And the last is one that's the least talked about, which is actually the post-market studies. Even after approval, you can get a contingent approval or even like, "Yeah, we're going to give you the approval but like just collect this data for us and be certain that everything is safe." All highly valid, useful areas to focus on in drug discovery. 

[00:05:38] SF: I realized that you're not necessarily focused on all those stages. And there's probably not necessarily one company that's focused on all those stages. But as a startup that's focused on any part of that and given that things can take 10 to 15 years to go essentially from the very beginning to all the way where people are actually using this, is it hard as a startup to be in the space to show progress or to feel like you're moving in the right direction? Because so much about being successful as a startup is the speed of learning and these learning cycles. But a 10 to 15-year full learning cycle is a very long cycle to learn from.

[00:06:14] PF: It's definitely very different. Compared to the normal tech startups that you chat with and we all hear about, we don't have the same luxuries as them where we can just put out an MVP and everyone knows, "Hey, we have market validation. We're good to go." Our MVPs tend to be very different. They're very focused on the molecule, the drug that we're making. And it's more of a proof of concept to show, "Hey, the platform itself works." But just having a platform is not our MVP. The code is the minimum requirement. 

Luckily, there's been a lot of research and understanding what makes a good hit molecule or a good lead molecule. These are the early drugs that aren't perfect just yet. And so, we have benchmarks to say whether our platform is putting us in that space. We don't need to wait the full 10 years to show, "Hey, we're definitely doing great." But it changes what your fundraising rounds look like. And there's a lot to talk about in that space. 

[00:07:05] SF: Yeah. I see. And so, Menten AI is focused on the preclinical development space or stage of those various stages that you've discussed. Can you share a little bit more background on what's involved with that particular stage before we sort of dive deeper into what you're doing at Menten? 

[00:07:25] PF: I could easily spend a two-semester course talking about what goes into preclinical development. Each little section I'm going to talk about has its own bachelor's level degree course requirement that goes into it. And I can kind of summarize this into some very short sentences that I love to use to explain it. 

We start off with pharmacology. This is what the drug does to the body. Does it have the correct effect? Is it engaging your target? Does it work in animal models? From there, we go to pharmacokinetics. This is what the body does to the drug. This is actually way more important than a lot of people give it credit for. Because you need to see, when you take that drug, will it absorb through your intestines? Will it get into your blood? Once it's in the blood, will it stay in the blood? These compounds break down over time and we need to know that it's going to be at a high enough concentration in your blood for a long enough time. 

From there, we move into toxicology, which is now switching the other direction. What's the drug doing that we don't want it to do? And then the last one is the one that more biology people like me try not to talk about as much, and it's the chemistry manufacturing and controls. This is our actual chemistry. The synthesis. Inactive ingredients and formulation. What's the shelf life of your actual chemical matter? 

But all four of these areas are areas that companies like Menten AI and any preclinical machine learning company is looking into all four of them. And it's because each of these four properties matter. And it's a complicated loss function that we're actually looking at. Because to improve your pharmacology, you don't want to take a hit on your pharmacokinetics. And it creates these very fun, multivariable problems that we love to solve in this space.

[00:09:07] SF: We talked about how maybe the full life cycle of a drug it could be 10 to 15 years. How long does just the preclinical development stage typically take? 

[00:09:17] PF: This is highly varied. I would probably give some rough numbers of preclinical development in total is going to probably be at minimum a year to about four years. Highly varied. I don't even want to give a good number on this. But there's just so many moving parts to it. And there's even these funny little anecdotes of just the hit discovery part. The first part where you just want to find a molecule that at least has an effect. And you don't care that it's a good effect. It's at least something. 

There's stories of finding that molecule within 6 months. But then the companies will spend another six months just to decide they found their molecule halfway through that year. That's not unheard of at all in these procedures.

[00:09:59] SF: Mm-hmm. Yeah. Even though it's not 10 to 15 years, it's still a pretty large, long-running project in the grand scheme of things. 

Outside of the new world of computational drug discovery and design that you're working in, how does that compare to sort of the traditional process of drug discovery of yesteryear? 

[00:10:20] PF: Yeah. A lot of people talk about how drug discovery hundreds of years ago was accidental. You just find a molecule and you'll find out, "Hey, this makes my headache not as bad." 

We eventually moved into the era of big, large pharma companies that can screen millions of compounds to be like, "Oh, hey. Out of these millions, we found five compounds. We can make some random changes and test another thousand compounds and eventually find our molecule." 

What's going on nowadays is the super exciting realm of using computer science, using smart experimental design to kind of limit our search space. Instead of looking at those thousands and thousands of compounds, we can remove that search space. With Menten AI, instead of testing thousands of peptide therapeutics, we only test about 50 to 100 in the lab. And that's that acceleration that we're seeing today. 

There are also people on the other side which are actually testing larger and larger libraries in silico. And that's also a whole exciting premise. Everyone's doing something different to speed up the same areas. 

[00:11:24] SF: A lot of it is about speeding up essentially the existing process. Either reducing the search problem or being able to search faster. 

[00:11:31] PF: We want to take that 10 to 15 years and bring it down much closer to five to 10 years. 

[00:11:36] SF: Yeah. Which obviously would have massive impact not just from a revenue perspective on the businesses that are operating in the space. But also, hopefully, for the benefit of people who need whatever the drug is that is going to help them.

[00:11:49] PF: Yeah. The faster we can find cures and the faster we can help people, I think that's hopefully the motivation at least behind the scientists. 

[00:11:56] SF: You mentioned peptide therapeutics. Can you talk a little bit about what exactly that is for software engineers like myself that are listening that have no idea what that is? 

[00:12:04] PF: Yeah. In biology, we talk about proteins a lot. Especially as a biochemist, that's my bread and butter. Proteins, we hear about it all the time. We have the protein powder when you're lifting weights. But what exactly is that? Proteins do everything in your body. They're the machines of the cell. And a protein is actually just a chain. It's a bunch of little Lego stacked on top of each other of these building blocks called amino acids. A lot of people in the fitness world hear these terms all the time. They love talking to me about biotech. 

These amino acids, you can have anywhere from 50 to over a thousand amino acids in a protein. And the sequence of those amino acids, because we gave them as little tiny letters, we have everything from A to Z. Well, there are five letters missing. But we can take those letters and whatever sequence we generate can be a novel protein with a new function or a new shape. 

Peptides take these very large molecules and we actually only do a very short chain. The amino acids that we work with in peptides are anywhere from three amino acids to about 30 amino acids. And it's highly debated when you become a peptide versus a protein. But for the most part, they're just a lot smaller and they do different things. 

[00:13:20] SF: What are some of the benefits of working with the peptides directly?

[00:13:23] PF: Yeah. Normally for therapeutics, we have two main classes we talk about. The first is small molecules. This is like your ibuprofen. It's a chemical structure. It's not really biological in nature with how it looks. But then you have your proteins and your biologics. This is also your gene editing and all those new technologies. 

The biologics are expensive to make. They're large molecules. You need a lot of material. It's really limiting that world. Also, these biologics aren't permeable. You can't swallow them as a pill and have them go into your blood. Whereas the small molecules, you can swallow them as pills. We all know how Ibuprofen works. It's really much easier to take a pill over an injection. 

These small molecules are what we're trying to push the biological space into when we utilize peptides. Peptides have been shown to have these properties of small molecules. And the more the field learns, the more we're realizing that we can kind of hack this biological capability of peptides to fit right in between these two very large modalities that are used.

[00:14:32] SF: What are some of the challenges of working with peptides versus working with the larger molecules? 

[00:14:39] PF: Yeah. Everyone you talk to in the field is going to have a different answer for this. And it all depends on what their job is and what their day-to-day is. I've met a lot of people that have issues with solubility of peptides. A good peptide drug won't necessarily dissolve in water for you. Things like that can be a problem. And we try to design and take that into account to make it easier to test these. 

Additionally, they're not as well understood as the biologics. Antibodies are a large class of drugs right now. The FDA approval process and the chemical synthesis of these larger molecules have actually been pretty well established and understood with FDA guidelines. We're kind of in more of a Wild West area. And those are just two to name a few some of the issues you come across. 

[00:15:27] SF: I want to start to talk a little bit about Menten AI and some of the work that you're doing there. As I understand it, Menten AI is a platform for molecular modeling. Can you talk a little bit about how does this work? When I think about platforms as an engineer, I maybe think about Twilio or, I don't know, AWS. What does it mean to be a platform for molecular modeling? Am I designing molecules with the product? 

[00:15:53] PF: In drug discovery, we use the word platform very specifically. And it actually clashes sometimes with the software engineer approach of a platform. But at the end of the day, it's the same definition. The platform of Menten AI is a computer-based code which can help us generate new peptides. 

I have friends at other companies that have their own platforms. These could be code-based or they could be an experiment. They have a certain type of experiment that generates a library. But at the end day, your platform is what gets you your drugs. 

How it works at Menten is we have a generative AI approach where we can design novel peptides. Usually, we can either do an internal program. Or a lot of our work lately has been with partners. We work with some of the big names in pharma. And we will take a target of theirs. And then what we do is we can actually use our algorithms along with our scientists to help kind of guide and build these new molecules. 

It's definitely not as easy and it's not as fresh out of the box as some other software engineers is working on. And it does create this funny thing where you remember a lot of us are biochemists pretending to be software engineers. 

[00:17:07] SF: Yeah. Actually, I wanted to talk a little bit about the makeup of the team or like what essentially the people working on the software are. You mentioned that sometimes you're taking people like yourself that have more of a biochemistry background and then you're working as a software engineer. Do you also marry that with software engineers that you're having to train on the sort of the biology of it? How do you sort of balance the technology needs versus the biological needs? 

[00:17:36] PF: Yeah. This is a multifaceted question. It all depends on our need at the time. There are times where we bring in contractors. And very early on in the company, we brought in a contractor specifically for data engineering and setting up a SQL database for us. No one in the company had ever had to do that before from scratch. And so, we will find outside resources. That tends to work better for a lot of things than bringing someone in in-house to handle it. It's easy for us to learn a lot of the computer science once it's up and running. 

There's other cases where we've cut back on the quantum computing research. But when that was a big deal for our company, we brought in many quantum computing experts which were not biology experts. And it worked really nice where we had the synergy where we could teach them the biology and even help set up the problems. And we put it in a machine-readable format for them where they don't even know that it's biology half the time. And I think it's really fun to find ways to set up those problems to help people.

[00:18:39] SF: Yeah. It sounds like, for specific types of problems, the one that you mentioning there, setting up a SQL Server from scratch, you'll leverage contractors. Does that also the case where I imagine if you're working – if you're not that experienced necessarily building in engineering systems at scale, you might run into scale issues or other types of sort of classical engineering problems that someone who's spent a decade-plus of their life engineering systems at Uber or something like that has figured out how to solve those types of things. How do you address that? Is that again leveraging contractors? Or do you have in-house talent that can solve some of those problems? 

[00:19:16] PF: A lot of it ends up being leverage in-house talent. And the issue for us is that so much of the time is spent on the biology of the science and the computation that's very closely related to the biology. And developing specifically for computational biology tools that the scaling and the improvements of this, it usually isn't enough to have someone full-time just come in to take care of that for the size of our company. I think this works a lot better at large companies and when you get larger initial seed funding, I think we could have benefited from having those people. 

But instead, we maybe had a week or two delay where someone on the team had to learn it we had to talk with someone. We had to figure out what those solutions were. And that's where kind of the difficulties come across where the learning to be a software engineer. But luckily, with the right support group at your company, they'll get you the resources to learn and find the people to talk to you to get us going. 

[00:20:13] SF: Mm-hmm. Yeah. Okay. That makes sense. And clearly, the people who are working on this are bright individuals that are well-educated. They, hopefully with the right resources and the right motivations, can kind of learn some of these tricks of the trade and so forth. 

It sounds like we haven't dove into all the generative AI and some of the things that you've done historically around quantum computing. But this we're talking about the team structure, it sounds like there's a lot of expertise involved. You need people who have experience and knowledge of drug discovery and have that kind of scientific training. And then you also are mentioning generative AI. You have regular software engineering needs. Quantum computing, how do you actually acquire all this talent? It's quite the Avengers that you're assembling there.

[00:20:59] PF: It's interesting to think back on. It's very interesting for me because I didn't have any of this before working for Menten AI. And when we hire people, we look for flexibility more than anything else. If you can come in and talk about your work, cool. I don't care. That's the bare minimum to talk to me that you can talk about your own work. 

When you can start giving suggestions of why your work matters to the bigger picture. How does this interact with other things that aren't what you work on? That's what we need to see to be like this is the type of person that's going to work well with us because we know they can learn what's adjacent to them. 

There's also just a lot of understanding of the company of flex to help others when you can. And raising your hand when you're the one that should jump in. Because at the end of the day, no one has all five of those skills. Even the quantum computing, I've struggled to pick up more than just the basic understanding to know what types of problems can be applied on there. But smart, flexible people is what you want. 

[00:21:57] SF: And then how are the teams structured across these sort of different areas of expertise? 

[00:22:01] PF: Yeah. What we typically try to do is keep people grouped up within their domain and allow for a highly interconnected matrix structure. You see this at a lot of – at least at a lot of biotech companies, I'm less familiar with how tech companies usually operate. But you'll usually have your department head of quantum computing, department head of protein design. And you'll kind of pick one person from each place to kind of work across a project across a matrix. But yeah. 

[00:22:28] SF: I know your CEO and founder has a background in machine learning from the University of Toronto. And maybe not everybody knows this, but unless you're in the machine learning and academic world, but University of Toronto is actually one of the world's leading institutions in machine learning research, especially around neural networks. Basically, they were doing neural network research for years long before deep learning and it became this hot thing. And now, the world generative AI that we're talking about. But in terms of what Menten is doing with generative AR, are you able to essentially leverage existing tools in the space and sort of all the development that's gone on? Or is a lot of it just built custom? 

[00:23:09] PF: Yeah. It's really hard for me to give a good answer on that because I've never worked at any other company doing AI. I want to say that we probably do a lot more custom work than other places would. I've talked to a few people just in tech companies and everything and my experience tends to be very different from theirs. 

A lot of building from the ground up to make sure it works with our specific data type and data sets. 

[00:23:34] SF: Can you comment on what the ML toolchain looks like? What is the sort of components in terms of like how you're training these things and how you're actually using the generative AI? 

[00:23:45] PF: I don't think I can comment on that if I understand what the ML toolchain is. Sadly, I'm not allowed to talk about any models or anything like that. We can give very vague, big picture. But a lot of the basic building blocks, we're using the same as everyone else.

[00:24:01] SF: What is the training material? Or is that something also that you – 

[00:24:04] PF: The training material, I can talk about. This is actually one of the things I enjoy talking about and warning people about if they want to come to biology. We have the worst data in the world. Biological data is the rawest, dirtiest data you will ever see. You can't even trust data from publications. You'll have one group aggregating data from 20 different papers all using completely different experiments to get the same type of data. And the error on these data sets could be 100% error. It could be double to even triple the value and it's still not right. 

What we've realized works best is either classifying the data and working it like that or working with computationally generated data. These are the areas that I'm personally most excited and interested in working on. And with the computational data, even if you don't have a deterministic approach to get that result, usually running it a few times enough even if it's non-deterministic will still be a smoother value than you get from the actual experiments. And so, that's the data we like to try and focus on. But every company's different and everyone's working in a different area. 

[00:25:20] SF: Mm-hmm. In terms of computationally generated data, are you talking about synthetic data? 

[00:25:23] PF: I guess I would probably call it synthetic data. In the biology world, we have a lot of computational biology tools. We have programs for designing proteins. We can use data from that. We have programs which look at the proteins and look at how they move, and flex and how pieces come apart. And we can watch the atoms in these simulations and we can use data from that. We can calculate force fields to show how two atoms would interact in a molecule and we can use that data. And so, we're using all these computational tools that we've been developing over the past 50 years in the chemical computational chemistry space.

[00:26:03] SF: And then you mentioned also another approach to classify the data. What do you mean specifically by classifying the data? Is that something that a human is doing? Or are you talking about using another machine learning approach to essentially automatically classify the data? 

[00:26:18] PF: Yeah. Typically, when I say classifying data, I mean something as simple as you have a range of solubilities. How much drug can you dissolve in water? Because that's a feature we care about. And instead of taking the raw number value, like, "Oh, 10 millimolar drug can fit in here." It's in my mind usually better with this experimental data to just say low average, high solubility. And computationally easily assess that. You just put filters at arbitrary data points and you allow for three. That way, you have a nice little in between and it's not just binary. But I think handling data like that works a lot better especially as the data gets noisier. 

[00:27:01] SF: Mm-hmm. I see. And then where does the quantum computing piece play a role? 

[00:27:07] PF: This we've decided to drop off over the past year. And it's not a major focus for us. We are still doing a few academic long-term collaborations, especially with our academic co-founder, Vikram Mulligan. And the goal here is we know quantum computing is going to be great for drug discovery. The quantum computers just aren't there yet. They're still a little noisy. We're still able to get solutions with classical computers. 

What we have noticed is that the quantal computers give different results. They're just as good. They're just different results at finding a solution. And that actually excites us in drug discovery. Because we've realized, over centuries now of drug discovery, our drug molecules started looking more and more similar. And we need to start throwing in random, new chemical matter. Anything to really diversify is great. And I think quantum computing is going to help us all out there. If not finding the best solution, finding new random solutions that are – not random but good solutions that are different from what we're used to. 

[00:28:09] SF: Mm-hmm. If I understand this correctly, and please correct me if I'm not understanding this, but, essentially, you're using generative AI to propose essentially new types of peptide therapeutics that then scientists can use to figure out is this something that's going to work as a drug to address some sort of human need. 

[00:28:30] PF: Correct. And I will expand on that a little bit. Our generative ey approach is like any other hit discovery out there. They're just hits. They're the initial drugs that have drug-like properties. Ours tend to encapsulate more of those other things where we try to predict what the body is doing to the drug. This is important. Because you take that molecule and you have to optimize it. It's never great after the first round. And so, you have to go back. You can apply computational tools. You can do traditional tools. But the job's not over when it's right out of the computer. 

[00:29:01] SF: Yeah. It sounds like there's still a lot of work. But this helps essentially shortcut the process and limit the amount of human work that might be involved.

[00:29:09] PF: Correct. 

[00:29:11] SF: What is the sort of unmet need from a healthcare and treatment perspective that Menten focuses on? Is it primarily speeding up, essentially, this process so that you don't need to leverage humans quite as much only for certain difficult tasks? 

[00:29:26] PF: It's actually everything. Speeding up the drug development process is great. It's exciting. It's fun. It's what we want to do. But with the generative AI approaches that we're implementing, we're finding novel chemical matter. We're finding drugs that don't look like drugs that anyone has ever tried before. 

And so, this allows us to go after a space of – in biology, we call it the undrugable targets. These are targets that we've been wanting to go after for the past decades and no one has produced a drug yet. And you don't go after those with what everyone else has been doing. And so, really, we can make a difference in untreated diseases. And I think that's more important than just speeding up the timeline sometimes. 

[00:30:09] SF: Yeah. It sounds like, essentially, by leveraging this technology, you might be able to come up with something that like no person was ever going to think about that's completely novel and different that allows us to have a major breakthrough that we weren't able to have before. 

That kind of reminds me of some of the work that Google Brain did on their Go program that essentially was able to beat the world's champion in Go. And it turned out that I think they used reinforcement learning for this. But, essentially, the Go program was doing moves that no human would ever think of doing. And then that actually allowed human Go players to get better because they could see the move and try to reverse engineer it and figure out why would they make that move. I never would have thought of that. And it actually is like a learning tool now to get even better human players. 

[00:30:57] PF: This is actually my new favorite analogy because of you, Sean. If you don't mind, I'm going to be stealing that moving forward. Because that's exactly what we try and say at Menten. I mentioned earlier how we don't just have the platform available for anyone to come in and use themselves. We need to have the scientist there alongside it. 

The computer's going to give us some random stuff sometimes that doesn't make sense but it will tell us what direction it's heading in and our application specialists can help direct the AI algorithm and kind of put it on pause, move it into this area and tell it, "Okay. Now do some exploring over here." And I think that's what we've been needing. 

[00:31:33] SF: And one of the challenges, and we certainly see this now in the LLM space and particularly like around ChatGPT, is understanding like why generative AI is producing the output that it's producing. It's very black box. Is that something that you also run into? Essentially, you could spit out some molecule but then it's going to take a person that has a lot of domain expertise to kind of reverse engineer how did this result arrive at this place. 

[00:32:00] PF: A lot of that matters to us. And these are the very fun questions in drug discovery. Because once you have a molecule, that's actually not enough to just optimize it. We need to know why is this one working better than the other one. Which areas on the surface are doing more of the interaction? And we've been developing more of our platform to interrogate and investigate these effects. Because you're absolutely right, that these are the things you need to be looking at. The why behind it. And no machine is ever going to tell us why it chose to do that.

[00:32:34] SF: Yeah. At least not yet. Maybe we can develop another AI that analyzes the result that tells you why it came to – 

[00:32:41] PF: I mean, I'm sure. It wouldn't be hard for drug discovery. I think having AI tools to quickly assess which regions of a drug are contributing to the energetics. I know that's an active field of research in computational biology. I imagine there has to be someone working on AI in that region. If you are, let me know. Because it'll be cool to see what's going on there.

[00:33:02] SF: In terms of the Innovation that's happening in the space, is it largely coming from academic research? Or is industry really pushing the space forward?

[00:33:13] PF: That is a solid question. I would give credit to everyone. Academics are great. They're solving amazing problems. They're finding new implementations. A lot of the startups in the past few years have been really blowing my mind and changing my opinion. And then even big pharma is using AI. Everyone is using different tools. And I think over the next 10 years, we're going to kind of come to a convergence where we start deciding what actually worked better. And I think this is the exploration phase. 

[00:33:41] SF: Yeah. It seems like there's this area. I mean, still feels very new. I'm very much an outsider but it feels new. But there's a ton of activity in a lot of startups that are being built up in this space. How close are we getting to drug discovery essentially feeling like an engineering discipline where you can sort of engineer the design of it versus it being a little bit some randomness or some luck involved? 

[00:34:09] PF: Instead of answering your question, which is impossible to answer, I kind of want to follow up with an idea of why we'll never be perfect. Because, yeah, we are currently engineering. That's kind of what Menten's doing, is engineering these drugs. A lot of companies engineer antibodies. But why is it never going to be as simple as just design what you want and put it in a human? And what I like to bring up with people is the data that's missing. 

We can do animal models. We can do purified biochemical systems. What we can't do is put random drugs into random people every day of the week. That's immoral. It's a little messed up. And it's also expensive. And for that reason, we're never going to have a large collection of human-centric drug discovery data, which is what we need. 

I can show you 100 different cases in the past 10 years of drug companies, create a molecule. They go to clinical trials. The first clinical trials just show it safe and healthy people. And then you go to the second clinical trial where you need to show it actually works in patients and it fails and no one knows why. There's no change at all in the disease. And everyone's just like, "Well, we thought it was going to work." 

And because we still see that, that's the bridge that we need to fix. And I don't think we're going to get there unless we get more human data, which I cannot rightfully vouch for. I think we need to rely on new biochemical techniques. There are companies that are developing organs on a chip. If I can name-drop very heavy bio companies, I like Emulate a lot. I think they're really cool. And they're trying to bridge this gap of generate human-centric data without the need to go into a human.

[00:35:44] SF: Yeah. It sounds like there's – essentially, we have a big testing problem when it comes to any kind of drug that's going to need to be ingested by a human because we don't want to harm people. And it takes a lot of work to get a drug to the point where you can even test it with people. But even then, it might fail. How can you essentially, we talked about at the very beginning, speed up that learning cycle in a way that we can essentially improve the tests earlier on without impacting or harming an individual? 

[00:36:14] PF: No. That's totally exactly what it is.

[00:36:16] SF: And then what are your thoughts on sort of the future of the space? You talked about some of these companies that are trying to solve this problem around testing. What are your thoughts on the future of the space? 

[00:36:29] PF: The future of this space is going to be limiting that chemical search space as much as possible. The more testing we can do in silico and the faster we can do it, the quicker we can give even less compounds to the wet lab to test. When we get to a point where we just give them five compounds and they all work flawlessly, that's going to be the future. And I'm hoping that happens in the next decade. 

Actual timelines, who knows what that will end up looking like? But I think a mix between generative AI and other predictors are what we need. Because it's not just do you bind to a protein? But are you safe? Are you going through the blood? Are you staying in the blood for long enough? And once we get to a point where we can answer all those questions with high fidelity. And I think that's what we're working towards. 

[00:37:18] SF: Mm-hmm. And then for people who are coming at this, maybe someone's listening to this, has more of a traditional software engineering background but they're really interested in this. And it sounds like something that, I don't know, it's easy to feel passionate about working on something that's hopefully helping people. How much biology do they need to really understand to work on sort of the engineering side of this?

[00:37:39] PF: I would say it's highly company-dependent. I told you from my experiences at Menten, we try to hire biologists more so than software engineers. Sometimes I reflect on that. I wonder if we made the right choice. I hope maybe other biotech people listening today will heed my advice and maybe bring in one of those software engineers early on. 

When you work with a smart biologist, they should be able to work with a software engineer. We should be able to help each other as much as possible without needing to be experts. But I will say, if you have zero interest in learning the biology, maybe stay away. 

[00:38:15] SF: Yeah. I mean, that's probably the case with any company that you're working on the product. If you have zero interest in the users of the product, then maybe it's not the right company for you. 

[00:38:24] PF: But if you're interested and you want to learn, do it. Those biologists are all going to be super friendly and helpful. And they're going to love talking to you.

[00:38:30] SF: Mm-hmm. Yeah. Awesome. Well, as we start to wrap up, Patrick, is there anything else you'd like to share? 

[00:38:36] PF: Not really. I really do want to stress that I think biology is a great space to move into. If you are a software engineer and you want to learn more about biology and how to move in this space, reach out to me and let me know. I'll gladly answer any questions. And I love chatting with people about random ideas. And if I can motivate one more person to jump into the biotech space, I'll feel very proud about myself. Because to me, this is saving lives. This is helping people. Hopefully, everyone working in the space can make a difference eventually. 

[00:39:06] SF: Awesome. Well, Patrick, thanks so much for being here. I really enjoyed the conversation. It's a super fascinating space. I think there's a lot, hopefully on the Software Engineering Daily podcast front, that we'll cover related to this. Because I think it's something that does feel like it's really exploding. And there's a lot of people doing really innovative, interesting work. 

And I think if you're listening to this and that, like you said Patrick, this is something that like does sound interesting, I think there's a real opportunity for people who have an engineering background who may be also interested in biology to really be a unicorn in terms of skill set that is going to really set yourself apart in the industry.

[00:39:44] PF: And I guess, on that note, if you are a biologist listening to this because you're getting software development, do it. Go for it. Everyone's going to love you. It is very hard to find people like me who enjoy software engineering. 

[00:39:56] SF: Awesome. Well, thanks so much. And cheers.

[00:39:59] PF: Hey, have a great day, Sean.

[END]