Biosimulation for Drug Development

Jo Varshney is the founder and CEO of VeriSIM Life, an AI-driven bio-simulation platform that enables pharmaceutical companies to accelerate drug development. She has a PhD in Comparative Oncology and Genomics.

In this episode, we cover a range of topics including:
- State of play in AI-powered drug discovery
- Why is it so difficult to know which drug candidates to pursue?
- The founding of VeriSIM Life
- Role of AI in drug formulation
- Role of AI in drug repurposing
- Role of AI in drug-drug interaction
- How does AI help in evaluating potential toxic effects of a drug candidate?
- What should ML developers know about biochemistry? And about biochemists?
- Determing the efficacy
- De Novo drug design
- Patient stratification
- The future of AI powered drug discovery

Jo's favorite book: Built to Last (Author: Jim Collins)

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Copy Transcript Copied

Prateek Joshi (00:01.002)
Joe, thank you so much for joining me today.

Jo (00:04.398)
Likewise Pratik, it's an honor to be here.

Prateek Joshi (00:07.586)
Let's get right into it. What's the state of play in AI powered drug discovery? Meaning, what can we do well today and where are the active areas of research?

Jo (00:23.11)
Yeah, good question. So the state of AI and drug discovery has evolved significantly in the last five years. It started with AI being known to only a few folks to basically AI being known to everyone, thanks to Chad GPT and of course Gemini is another big thing that healthcare folks are now catching up on. And there are certain aspects of drug discovery that

AI has significantly shown massive improvements. So for example, alpha fold. How do we identify basics of protein chemistry and protein folding? That can implement how best you can build a new drug with a specific disease target. Because the big challenges in drug discovery are, how do we first identify a novel?

or a very well-known target, then how do we create a drug that affects the target in a very specific manner? Because the target's chemistry is quite similar to many other targets in your body. So what AI has done in today's field, and a lot of companies have really worked significantly on finding the best what we call computational chemistry. It could be either like how well the binding occurs.

between a drug candidate and a target, how well we can identify new targets, or how well we can use omics data to find new targets and find potential drug candidates. And there's a lot of in-betweens in that where AI has been significantly used.

Prateek Joshi (02:10.446)
amazing. When you look at all the drug candidates that a large company can pursue, why is it difficult to know which ones are likely to succeed versus the ones that are likely to fail?

Jo (02:27.674)
You know, it's a great question. I spent 15 years, you know, answering that question. And I was really obsessed, you know, why this is why I started Verisome Life is the drug discovery part. It's just the beginning. There are so many steps and the steps get more and more complicated because ultimately they have to work in humans. So you start with a very high throughput, lots and lots. You think of millions of compounds. And then now you have to narrow it down.

Prateek Joshi (02:30.967)
Hahaha

Jo (02:57.102)
without having to test in humans, what's the best drug candidate? Then it comes, the complexity gets added. And this is the propagation of error situation when it comes to animal tests. Like if I told you the Tylenol we take, we all are taking based on what rat body weight was, and it got scaled to humans and a human, which is a 40 kilogram, a 45 year old white male. I think we will all pause. It's like, that's ridiculous.

We are not that standard human model, and we're also not like rats. So that complexity creates a lot of errors, which we call the translational gap. So how do we translate what we see in animals, in experiments, and drug discovery to work in humans? It's a, in my opinion, a multi-billion dollar caution, and it's not more. Because ultimately, if we can...

figure that magic out, the time to market will significantly improve, and of course, the cost will come down. And that is one thing I'm gonna shamelessly insert what we do is we are enabling that.

Prateek Joshi (04:12.05)
Yeah, I think it's a good stopping point to just quickly talk about the SM life. So for listeners who don't know, can you just quickly explain what the company does?

Jo (04:22.298)
Right, so Verisim Life has built a hybrid AI software platform to create what we call a credit score for drug development. So just like a credit score that you have tells you the financial health, similarly, we have what we call a translational index score. It tells the drug developers what's the likelihood of clinical success at an earlier stage before the drug enters into the clinic.

Prateek Joshi (04:51.69)
Right. And there's so many parts to this. So I want to touch upon all of them. Let's start with drug formulation, right? And it helps determine the best way to deliver an active ingredient or a molecule to patients. Why is it challenging and how does AI help in this step?

Jo (04:55.366)
Great. Yep.

Jo (05:05.755)
Absolutely.

Jo (05:12.49)
Right. So let me first answer why is it challenging. The biggest challenge here is formulation is an afterthought in drug development. What that means is drug discovery folks and even like early stage drug developer are not thinking about formulation. Their objective is to find a molecule that works. Now how it works, how much you have to put it in a human is not the questions that they are answering. So for example an active ingredient could be active.

But if you have to take kilos of that drug as a powder, it's impractical. So that's what the formulation is. So the formulation folks will always tell you, and this is usually being a relevant challenge in around phase one, which is around when the patient trials and the dosing questions are really big and problematic. So they have to figure out what's the best carrier, like excipients, and the way...

the AI can really help us identify the right choice of excipient with the active ingredient. So there's like library of excipients, you know, available. And we are first focusing only on small molecules. So with that library, what's the probability of which excipient has a better chemistry with the active ingredient is where AI can be really helpful and something, you know, we actively pursue. So that an earlier stage, before you have to really start thinking about formulation challenge, you can think about it.

And then you can test that out, you know, in a petri dish, which is much cheaper and more time efficient so that you have the knowledge, okay, what would be that combination and how that would be relevant in increasing, let's say bioavailability in your body, in your gut, and then how it will reduce sort of the toxicity or the clearance of the drug. So that's really what we believe that AI can be helpful at.

Prateek Joshi (07:12.174)
Another component here is drug repurposing. And large companies can, they do this, there's a lot of value to be derived from it. So can you explain drug repurposing and also in what ways can AI help here?

Jo (07:30.478)
Right, I think this is where AI can really truly shine because first of all, what is drug repurposing? It's like a drug that is already approved for a specific disease. Now you take that drug, which basically has gone through the ringer of all the validations and the FDA approval. Now you take that, basically you have a lot of data and now you apply it for unknown disease so that now you have enough data and information

disease that could have similar implications for a new disease, or it could be a rare disease. So the benefits of using AI is one, you know, you can, it's a transfer learning approach. Number two is really you could apply with more confidence since there's a lot of data and a lot of validation from the regulatory body that how well this drug would interact in that other target that you of interest.

And then another big aspect of drug repurposing is the FDA has a shortcut path. Like they look at repurposed drugs more from a lens of like, how quickly can we get to phase two, which is a massive opportunity for drug repurposing drug developers to not have to do the phase one and directly test out how safe and effective the drug would be. And we call it the direct to phase two.

applications.

Prateek Joshi (08:59.934)
If you think about drug repurposing, now, so let's say there are two paths. The first path is you start from scratch, you develop the best drug possible for this given scenario versus drug repurposing where you're taking an existing thing and trying to repurpose it. Now, will that, will drug repurposing, will that reduce the effectiveness of the drug? Meaning like if you had started from scratch, would that lead to a better drug by any chance? Or is it the same?

Jo (09:30.511)
It's a fantastic question. The question here is, question here that we are trying to address is the intellectual property. So, many companies and many investors invest in companies who have novel drugs rather than repurposed drugs because you have significant IP. When it's a repurposed drug, you don't get the matter of composition IP.

There are ways to still enhance your IP portfolio by creating a new formulation or delivering the drug in a new route of administration. So we have a subsidiary called PulmoSympTherapeutics. We used AI to repurpose a drug, which is very well known and a very different disease. Now we are repurposing it for an inhaled form. So it solves a lot of those IP issues.

And then the question that you ask, will that be better or worse? It really comes down to the disease you're going after. If there is a significant need, for example, a rare disease population where there's no drug to cure it, it's a really important thing for us to think about time, because creating a new drug means you have to invest a lot of time and you have to have a lot of data to convince FDA to make, put it, test it in humans.

But with repurposed drugs, you significantly reduce that time. And with AI, there's an added bonus, right? You can do a lot of those iterations virtually or in silico and then do an experimental method to validate it. So I think it really comes down to what type of disease you're going after and how many different drugs are, what's the competition and the market looks like. And of course, the interest of IP expansion.

Prateek Joshi (11:27.01)
Right? Let's talk about drug interaction, which is defined as two or more drugs interacting in such a way that the effectiveness or toxicity of one of those drugs is altered. Now, just quickly, why is this a problem? And two, what are the ways that people are using to figure out that, hey, what's the problem and how do we fix it?

are available to detect this problem.

Jo (11:59.266)
Right, you know, really important problem and very less talked about challenge. Like any adult in today's time, I would be surprised if they're not taking supplements or, you know, a drug in a consistent manner. And of course, the supplements are not drugs. But if you think of food and drug administration, there's a reason why they call it food and drugs. So it's the interactions that you want to know before you take any drug. And you know, you probably know like you don't take

a drug with milk and they will tell you all that in the prescription and there are real reasons because one, you reduce the bioavailability of a drug. So that means it becomes less effective for that disease. And if you look at cancer patients, they need to be put on, you know, they go through chemo, then there is a, then they may be put on different types of drugs. And if you know enough about what those two different types of drugs, how they will interact or even like when they should be taken and at what time during the day.

that is a significant reduction of any kind of adverse outcome, which could be toxicity or like some major side effects, because these patients have serious implications of taking two drugs that may not be friendly with each other, or we call it synergistic. So I think biggest thing that we have found is like there are FDA databases that talk about existing drug and what's then known.

drug-drug interactions. So that's a good start for ML engineers to really dive in deeper. And I would say, have a friend or make a friend who understands pharmacology or the drug-drug interactions so that you are applying machine learning in a more comprehensive manner and understand the biology a little bit so that you could combine what the chemistry of the drugs are, what understand...

what the kind of data that being done and how that can relate in terms of simulatory matrix to other drugs that we may not know the drug-drug interactions. So I think that's a really good start for machine learning engineers if they're interested in drug-drug interaction.

Prateek Joshi (14:13.614)
This is actually a good segue into my next question. You know, you've been, you know, spent your entire professional life in this field. What should ML engineers know about biochemistry and also about biochemists?

Jo (14:31.638)
Ah, interesting. So biochemistry, I'll spare the biochemist at the moment, is a specific field that looks into the cellular processes. So the machine learning folks need to understand the difference between biochemistry, chemistry, systems biology, and then computational chemistry. They're all very different field and they serve very different purposes in the drug discovery and drug development.

Prateek Joshi (14:37.282)
Hahaha.

Jo (15:00.53)
So when it comes to biochemists, their job is to look at, for example, how is the energy being utilized in a cell? And I don't know if you probably recall Krebs cycle. And it's a fateful thing that we all had to learn about, but the reality is that's how we know, how within the cell, a chemical entity, if it enters, how it's gonna get processed, what's the...

cellular energy process or change is going to be and those perturbations would be very important for computational chemistry folks to understand in terms of chemistry. So if you know I'm happy to expand anything further if there are any follow-up questions what ML folks should know about.

Prateek Joshi (15:44.907)
Yeah, yeah. I think it's also important because there are so many AI tools being developed for many different, it's a vast field like drug discovery, biology. So I was just curious about if an ML engineer that would spend their entire life just building ML systems and now obviously this area of biochemistry is of interest to them, but they don't know anything about this field.

Jo (15:53.114)
Right. Absolutely.

Prateek Joshi (16:13.998)
usually in situations like that, they just approach it without any domain knowledge. So it was more from that angle. I think as you said, the best thing to do is make a friend who knows a lot about biochemistry in general and absorb everything. Okay. Let's move to drug candidates and evaluating potential toxic effects. It's a very important field for any drug candidate, drug assets, and big companies. They have to do this before.

Jo (16:18.566)
Right. OK.

Jo (16:24.444)
Right.

Jo (16:27.698)
Great.

Jo (16:36.158)
way. Absolutely.

Prateek Joshi (16:43.486)
anything happens in real life. So how can AI help in evaluating these potential toxic effects?

Jo (16:48.862)
Yeah, absolutely. Very important question and a lot of drugs fail when it comes to toxicity. And so they're first of all for animal folks, they should understand there are different types of toxicities that are important or the FDA really cares about. And also we should care about before putting testing the drug into in humans. So one is like, for example,

Prateek Joshi (16:57.366)
Alright.

Jo (17:16.77)
It could be specific cellular toxicity. So there's a lot of different elements. There are several FDA has released several different types of datasets to talk about these different toxicity. So there's Tox21, there's ToxGaAs. They also have like free webinars and workshops that you can attend because they're working on a report where they will release more data, which is really great starting points to understand how machine learning can be applied to these different datasets. And so that's number one. Number two is,

Applying machine learning to predict different types of toxicities is more straightforward, but how they are related to each other is a little more complicated process because they are not standalone toxicities, right? Because body is interconnected, so is toxicities interconnected. For example, the other aspects that machine learning folks should be considering is the drug dosage. So it's, as you know, like if a specific dose is being prescribed by a doctor, there's a really big reason because the

the, if you go over that dose, you can have potential toxicities and that could be detrimental to your health. So having that kind of knowledge and understanding how what we call drug exposure within your body and how that implies to different types of toxicity is important. The other aspect that is again very little talked about in unfortunately in our field is off-target effects. It is a really hard problem. What that means is

You did all the right work. You know, you found the best models, you've tested it, you validated it, you did experiments, FDA was happy, you go into humans. And then you're like, wait a minute, my patients are like throwing up and having massive seizures. That wasn't something we thought about or we predicted. Now that's where, and then I'm gonna shamelessly insert our company too, but I think that's where we have to significantly spend time understanding.

what are the reasons of these off-target effects and are there ways that we could catch it earlier before the clinical trials occur. So that's another field that I think again not overlooked but really if we can have more folks to take deeper interest in understanding the off-target effects I think it could really shape the future of clinical trials.

Prateek Joshi (19:39.514)
Right. An important metric for any new drug candidate is its efficacy, which is defined as the medicine's ability to produce a desired effect. Now, I have two questions here. One, how do you measure efficacy objectively? Like, how is it measured? And two, how can AI help increase the efficacy of a given drug asset?

Jo (20:09.298)
Efficacy is very specific to the disease of interest. So for example, if you're measuring efficacy for a cancer drug, you are looking at how much it shrinks the tumor without having any toxic effects, for example. If you're looking at a pain medicine, then you're looking at how quickly... It's an absurd experiment, unfortunately, where if you put a heat pack and how quickly the animal moves the...

far from that heat band. And that kind of pain, because pain is very little understood, is the metric for efficacy. So it really is context specific. Unfortunately, there the cross translation is, it's not very straightforward. And you have to really understand the disease of interest and how, what kind of test they are going to be doing in humans. So

Like for example, in our rare disease, pulmonary arterial hypertension, the patients are going to ask to walk after taking the drug. So how long can they walk? What's their impact on the heart? Because it's the lung and heart condition is the measurement. So how can AI help? So I think one is we have a generalistic approach of understanding the targets. So the targets could be, you know, we have a target library. We can, it's a good start.

and then looking into applying machine learning and understanding how these targets could impact the efficacy of the drug binding. So that's a more of a generalistic scalable approach. And then you have to go into creating specific models based on the data from the experiments and also from the lab Petri dish experiments on how that binding can impact

For example, tumor growth inhibition or walk test in patients. So there are two different approaches we can take to solve or predict the efficacy of a drug using AI.

Prateek Joshi (22:17.03)
All right, amazing. Because obviously, I think for a given situation, if we can increase the efficacy, everybody would. It's a winning situation. So I think it's a good tool to have. Let's talk about de novo drug design. And it's not a term that you come across, unless you're in a domain, it's not something you come across on a daily basis. Can you explain what that is?

Jo (22:45.05)
Yeah, so the de novo means creating a new structure, which does not exist, right? So it's like a new structure and the combinations are 10 to the power 63. So creating a structure that could help impact a target, disease target, is what really is de novo drug design. Most of it is inspired by existing knowledge. Now there's generative AI where, you know,

it's more like a global approach, like, hey, let's just go out there and really span out the space and see what comes out of it. So we do a combination of global and local approach and in a manner which is a little, actually it's quite different than what drug, do you know what drug design is. We take a reverse translational approach. So for example, we know what the disease is, we know what are the

efficacy markers are, we know what the patient responsiveness needs to be like, and then we build models to predict from a seed structure what's the best drug feature combination could look like for a new molecule, which could address all those quotients and run that cycle millions of times to find the best molecule and then filter it out based on

the efficacy, potential toxicity, potential drug exposure and all that. So most of the de novo in today's time is only looking at the chemistry. We try to make sure it's taking into account all different aspects and elements from the human standpoint and then coming back to chemistry. But ultimately the goal is to develop a new drug structure that has not been developed or created before.

Prateek Joshi (24:36.046)
Another important pillar in this whole process of assessing drug candidates and bringing them to market is patient stratification. Can you talk about what that is and why it's important? And also obviously, how is AI helping here?

Jo (24:57.058)
So patient stratification is basically segmenting patients based on it could be age. So for example, pediatric population versus adult versus elderly. It could be depending on gender. It could be very specific to the enzymes you have in your body. So if you have specific enzymes, liver enzymes.

that could metabolize a drug more efficiently versus a subset of population don't have those enzymes or those enzymes don't function quite well, that's what the field of patients ratification is really doing. So how can AI help? So let's double click on liver enzymes. So there's a very well-known understanding of different liver enzymes in humans. So you could take a look into that and predict how your drug

could metabolize with the chemistry, that known chemistry of the drug could metabolize within that population versus the population does, which has a malfunctioning or an absent enzyme. And that could itself help in better clinical trial protocol because when you're designing a clinical trial, you wanna be able to present like, hey, we wanna test the drug with different doses, with different types of representation because that gives...

statistical significance, but more importantly, efficacy is significant, right? Because if you know, oh, I have to increase my drug dosage because it would poorly metabolize, but it shouldn't be toxic. And if we present that kind of work to the FDA, they, it's really great for them to really represent equally, if not better, the patients that, where the drug is gonna be tested. So I think AI can significantly improve that process.

Prateek Joshi (26:49.666)
Right. I have one final question before we go to the rapid fire round. And it's about the future of drug discovery. Recently, the CEO of NVIDIA, Jensen Huang, he said, in an interview, somebody asked him, like, where is the next big revolution in AI? And he said, like, straight up biology, like digital biology using AI is going to make a huge impact in that field. So,

What's the next big breakthrough we can expect in the field of AI-powered drug discovery in the next three years?

Jo (27:29.07)
Yeah, I think there are going to be massive shifts in the way FDA and other regulatory bodies review the package that comes for them to understand AI application. I think that's the biggest thing that we are not talking about, but ultimately we can do all kinds of innovations, right? Our biggest hurdles in drug development is the regulatory bodies to approve those methods and processes.

And working with the FDA now, we know they're open to it. They are very excited about this. And in the next three years, I really am pretty optimistic that we are going to see better clinical trial protocols, faster time to market, and better ways to test the drug so that we are avoiding the research and development based and unnecessary animal testing. And I think that's really where I am.

putting my body.

Prateek Joshi (28:27.806)
Amazing. All right. With that, we are at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. All right. All right. Question number one. What's your favorite book?

Jo (28:32.463)
Right, let's do it.

Oh

Jo (28:42.912)
Built to Lost by Jim Collins.

Prateek Joshi (28:46.702)
Amazing. All right, next question. What has been an important but overlooked AI trend in the last 12 months?

Jo (28:56.258)
explainable AI in regulated sectors.

Prateek Joshi (29:02.41)
Right. What's the one thing about drug discovery that most people don't get?

Jo (29:09.714)
It's just the beginning.

Prateek Joshi (29:12.895)
Amazing. All right, next question. What separates great AI products from the good ones?

Jo (29:20.638)
Clear ROI by using artificial intelligence is what makes an AI product great.

Prateek Joshi (29:30.186)
As a founder, what have you changed your mind on recently?

Jo (29:36.978)
Building public credibility is more important than I really ever emphasized or thought about.

Prateek Joshi (29:45.767)
That's a good one. All right, next question. What's your biggest AI prediction for the next 12 months?

Jo (29:55.714)
I want to go back to my answer that having FDA being more keen and more lenient in understanding AI applications is going to be really big in the next year.

Prateek Joshi (30:09.31)
All right, final question. What's your number one advice to founders who are starting out today?

Jo (30:16.402)
Hmm. Get comfortable with disappointment and failures. It's all part of the building the company and be bold. It's a full combat sport. There's nowhere to hide once you start a company.

Prateek Joshi (30:32.231)
That's amazing. Love that advice. And yeah, I think I agree. I think just being mentally and physically tough because it's going to take a lot. So Joe, this has been such a brilliant episode. Love your depth of knowledge in this field and loved how concise and sharp the answers were. So thank you for coming on to the show and sharing your insights.

Jo (30:48.286)
Thank you.

Jo (30:54.946)
Yeah, no, really excited about when the episode comes out. I hope this will help folks who are taking interest in biology, which I think every machine learning engineer in today's time should understand, you know, different challenges and databases. And of course, please feel free to reach out to us if you're interested in learning about any specific areas. And we do a lot of webinars, which are free to attend. So please follow us, verisimlife.com.

Looking forward to staying in touch. Thank you so much, Pratik. Have a great day.

Prateek Joshi (31:27.773)
Perfect.