[Sales] How AI is disrupting sales and drug discovery with Jason Steiner

There's clearly tools like Harrison you mentioned that are built into HubSpot and Salesforce and other tools that will be useful, you know, potentially for things like insight generation. There's, there's obviously tools that will accelerate things like prospecting. I mean, all of the language generation stuff. There's just going to be a lot of set of tools that will accelerate the ability to be better and perhaps just more effective in the way that you communicate with customers. In terms of, you know, how it will affect a salesperson's communication with customers. I think that just like all of the other trends, it will just enable customers to be more savvy and salespeople need to become more savvy as well. But also, I mean, the general trend of, of, of customers just becoming more informed has been a trend that's happened over a long time. Like, customers will be able to get more information faster to come to sales interactions. And this is just going to continue to augment that. They'll be able to do more on their own.

All right, Nick, we got a really good, fun one for everybody here today. We're talking about a really hot topic, which is AI and drug discovery. And we're going to be joined by Jason Steiner. And he knows a lot about this. This is the space that he is currently working in. He was also the first sales hire at Synthego and then worked his way all the way up to become the chief Strategy officer as Citigo as well, and now working heavily in the AI and drug discovery space. I know we talked a lot about a lot of good stuff in this episode, but what was some of the stuff that stood out to you?

Yeah, it was a good episode. I always enjoy talking to Jason in this episode. Something that was particularly interesting to me and something I've tried to piece together a lot is there's a lot of AI based drug discovery companies, and it's a really big hot topic. But what does that actually mean and how do you leverage, knowing they're an AI company, to have good sales conversations with them and dig into what their problems are and how you can work with them? And I think Jason gave us some good information in there how to frame each of those companies.

Yeah, lots of good stuff around AI and drug discovery. We talk about AI in the sales and commercial world. And to top it off, Jason tells us about his recent trip to Mount Kilimanjaro where he raised a bunch of money for cancer research. So with all that, let's dive on in. Hey, Jason, welcome to the podcast.

Happy to Be here. Thanks for inviting me. Good to see you.

Yeah, it's really good to see you again.

Yeah, it's been a while.

Yes, that is true. But no, it's great to see you. Great to have you here. And as you know, we prep, prep you ahead of time. We ask everybody the same question. What's a crazy or funny story that's happened to you in your career?

Yeah, I thought about this a lot and I think one of the things that people ask me quite a bit is how I got started in kind of commercial, the commercial side of biotech. Having come from an academic background and one of the things that I find perhaps most interesting about my journey is basically how I got started and how serendipitous it was and where I ended up. And so I think the, the funny thing that I, I keep coming back to is I had done my, my academic training in San Diego. I spent about 10 years at the university doing mostly engineering degrees, very, very technically oriented. And I had done a bunch of other things when I was in graduate school as well. I had been part of starting an organization that was part, part and parcel of, of training student entrepreneurs. I had done a lot of traveling in my 20s. And when I, when I was in graduate school I decided that I, I wasn't likely to be a scientist or professional academic for my career. And so I started looking at the different things that I was interested in or might go into. And as things had happened I was actually dabbling with the idea of going into venture or going to start and going to join a biotech type company, like a startup company. And I had been discussing with a number of different folks and actually I, I had ended up accepting a role at a venture capital firm in New York and it was for kind of an internship type role which, which actually might have been a very interesting, a very interesting career path. And at the time I was still interviewing with some companies in San Francisco, some biotech companies and. But I actually had driven across the country in, in an effort to, you know, go to move to New York. And by the time I had driven all the way across the country, I actually got a phone call from one of the biotech companies out here in San Francisco and they wanted to interview me. So after driving across the country, they flew me back to San Francisco.

What? At least you flew back and didn't have to turn around and drive all the way back.

I flew back. It was about a two or three day time that I was back in San Francisco. They interviewed me, they gave Me a job offer. So then I flew back to the east coast, and then I drove back across the country back to San Francisco, because I accepted that offer. And in a period of about a week or two, I had found a place and started. And interestingly, the role that I started in was in international sales role. So it was at a company called Natera, and they were just starting or building up an international business development team. And so we kind of. I was one of three at the time, and we split up the world, and I had about a third of the world. And in my first week I started there. And on Friday I was on a flight to Brazil. And so my experience of entering biotech was essentially driving across the country to join a VC firm in New York, flying across the country, back, having an interview, flying back, driving back, finding a place, and then being on a flight to Brazil in my first week.

And that.

That kind of started my. My commercial experience doing market development in, particularly in South America, but in other countries as well for the next couple years. And I was a sole operator doing that for a while. And it was just kind of an interesting way of, I suppose woefully underqualified for the. For the role at the time, which surprised me, but kind of consolidated the mix of my experiences of traveling a lot around the world and being more commercially oriented and technically minded. And that's kind of how I got started. And it seems pretty atypical, but perhaps.

An interesting story for any Europeans out there, though, right? The east to west coast drive is quite short, right? Just a little drive.

Super short.

Yeah.

It's only a few days.

Yeah.

It's a long, long haul.

Yeah, for sure.

And to do it. Do it twice in. In about a week, actually it takes about a week and a half if you just do it straight. But yeah, anyway, so that. That's kind of how I got started, and then. And then the rest is a bit history. I spent a couple years in international market business development and then proceeded to do a number of other things, but kind of cut my teeth in that area.

What do you think helped you be successful knowing that you were, as you say, underqualified to. To be in the position?

You know, it was just a. It. It was a unique combination of things that they were. That. That was basically specific for my profile. You know, I had spent a lot of time traveling internationally. I was. I was comfortable doing that in. In. In my 20s, I had done that. It was an interesting time because at Natera and then also at Synthego, one of the Real interesting times about these two companies is that they were both at the beginning of very big technology introductions into the commercial space. So at Natera, it was basically clinical genomics. So the first population based application of sequencing for, in this case, prenatal screening, where essentially now like any, any pregnant woman could have. Could have a sequencing based test that was, that was quite good. So there was a tremendous amount to pull, but it was also a very technical sale. So it's introducing a new technology. And because I had a technical background from, from my time at the university that was actually very, very important in communicating the, the nature of a new technology. And so that was kind of in my favor. And then I had spent a lot of time doing other types of things that were related to the commercialization of science while I was in graduate school. And so I was able to transition to that, that more kind of commercial perspective. So it was just the confluence of a lot of things that were very specific for my profile. So I feel quite fortunate to have, you know, fallen into that space and time and place. And it was a great experience to kind of start my, my career off after grad school.

Not, not to do a too crazy of a segue here, but speaking of international travel, you recently just returned from a big international trip. You want to tell us a bit about that journey that you recently went on?

Yeah, so that was super interesting. It was so for, for the audience. I was in Tanzania for about 12 days climbing Mount Kilimanjaro with a, with a group of biotech industry execs. And it was a very unique experience. There is. It's part of a series of campaigns that have been running for the past few years that are organized by a journalist by the name of Luke Timmerman, who writes an industry newsletter. And kind of the intent of these campaigns are to gather a group of biotech and pharmaceutical executives from across the industry. So lots of different types of disciplines represented to put together a group and to go raise funds for nonprofits. In this case, we were raising funds for cancer research, supporting a foundation called the Damon Runyon Cancer Research Foundation. Past campaigns have supported groups like Fred Hutch or Life Science Cares. And so to gather this group together to go raise some funds for a worthy nonprofit and then to kind of consolidate it with, with one of these group kind of expeditions. And so the, the particular Kilimanjaro trip, this is, I believe, the third year that it's been running. So this was kind of the third campaign. There were a group of 20 of us, about five or six months ago, we started, we started doing this in terms of raising funds. We, our goal was to raise a million dollars for the foundation. We actually beat that goal, which was awesome. And then we, we went together and did something kind of, kind of crazy for, for a lot of people, myself included, which was go fly to Tanzania and go climb the highest peak in Africa. Uh, so it was.

Did you train for it?

Super awesome experience. Uh, yeah, I think I did. Everyone did. Gotta get the, gotta get the cardiovascular system in place. It's, it's not an easy thing. It goes up to above 19,300ft or so and it's pretty quick. It's the tallest freestanding mountain in the world. So from the base to the top, the elevation gain is actually the highest in the world. It's not the highest mountain, but the elevation increase is the highest.

So it's the highest climb, basically.

Highest climb, yeah. And it's, it's a walking, it's a walking route so you can walk up the whole thing. You don't need ropes or crampons or. There's no, there's no glacier type work, so it's accessible in that respect. But it is a difficult thing. The altitude is a challenging thing. So very good bonding experience for the team. Personally challenging, professionally challenging and something. Now there's an increasing kind of alumni group of people that have done these types of campaigns which I think is a growing asset for the biotech community and feel a privilege to be a part of it. So that's what we just thought.

Really cool experience with all of those people now and even raising all of that money that you did is fantastic.

Yeah, really, really good causes, worthwhile causes. One that I resonated with because I did my PhD at a cancer hospital. Um, like I, I understood a lot of the dynamics around the type of, the type of funding that the Damon Runyon foundation was supporting. Um, and it just resonated with me personally. So definitely a worthwhile thing.

Thanks for listening to the podcast. It would mean the world to us if you could help spread the word by sharing it with someone you know. Succession is a life science sales membership that includes on demand courses to uplevel your sales skills, resources to plan to execute your sales process, community for peer to peer networking and learning, and live workshops delivered twice a month. You can get access to all of this as an individual for just $79 a month or up. Level your entire team with customized learning paths built for each individual rep. Come join us at Succession Bio.

I think we're Going to change gears a little bit and ask a different question. I know we've spoken before and you're really into the AI space in various guises. And one area I think is really interesting, particularly for reps listening is. And something I've struggled with a little bit is understanding the intersection between AI and drug discovery and what that really means. And I think if you could, if you've got any wisdom on that, I think that'd be really valuable.

Yeah. So I mean, drug discovery and development is a big, it's a big category of things. I think it's important. There's a couple important things to point out when talking about AI or machine learning. There's a lot of discussion around AI, artificial intelligence at the moment. But the reality is basics of machine learning, like how you learn to model things from data, that's been around for a long time. So a lot of the discussion around AI is kind of focused on this concept of generative AI, which is how you create, like how you create new information or how you model things that are very, very complex. So I want to just kind of separate this idea of machine learning of like linear regression, which has been used for obviously a very long time, and the current discussion around AI, which is probably more related to things like deep learning, large language models or the generative AI. So much more, much more complex type. But machine learning and quote, AI has, has been around for, for quite some time in a lot of areas of research. If you look at kind of this, these deep learning models, these very complex big models that are, their goal is to be able to extract patterns from lots and lots and lots of data, to be able to extract kind of the, the key features of them at different levels that human beings couldn't see and then to be able to potentially generate new content in the case of language, or to be able to make better predictions for defining hypotheses or generating molecules or things like that. So that kind of scaling side of AI is a new thing. And there's kind of three, I think of it as, from a technical perspective, as kind of three buckets where this type of very complex feature extraction and, or generation is being used in drug discovery. There's kind of the biological side of things where you're dealing with this is kind of your target discovery or your understanding of how cells work. So this is where you're taking massive amounts of things like genomic data, proteomic data, imaging data, and you're trying to develop models that will extract key features from that data and then Allow you to make predictions about how things might behave in the future. That's a very kind of complex system analysis and that's kind of core, kind of core target discovery or biological understanding. So a lot of the big databases that have been built over the past maybe 20 years, like the ENCODE database, things like the human cell atlas, some of these large, the links database, some of these large kind of public repositories, a lot of the data from that is being used to feed into these models to try to predict how things might behave for, you know, better target identification, better toxicity predictions, things like that. So that's kind of on the complex bio side. The middle side is kind of the generative molecular side. So this is all of your alpha folds, your protein engineering, your small molecule engineering. A lot of the models here are going to be built on essentially built on physics, and physics that we can't necessarily discreetly calculate, but that we can build a model to figure out how to approximate. So we can't calculate all of the interactions of how a protein will fold. But you could take a bunch of protein examples and build a model like AlphaFold did to actually approximate and be able to predict how a protein will fold. And there's a whole bunch of stuff in the generative space for protein engineering, small molecule engineering, where you're trying to both create or dream up new types of molecules and then be able to predict the properties of those molecules, how they might behave, and so on. And so that's kind of like the whole bucket of how molecules might be better designed to be more likely to have certain types of properties or functions. And a lot of that's kind of based in physics.

So there's an idea there then that you can create. Essentially you'll be able to create new molecules that may not have previously existed based off of these models. Right? Okay, exactly. Like you take this library, like old school screening, right? You take this library of known molecules and that's what you're using against your target versus this might be something completely different where it's like you actually are trying to find potentially new novel molecules that don't exist. And then you would go and then create those.

Okay, yeah. So you take a bunch of the information that you have around what molecules look like and then what they do, and you essentially train a model that tries to identify particular features in these molecules that are responsible for certain outcomes. And when you identify these features, they're kind of abstract features. You don't necessarily. Humans don't know what they are. The idea is that these models will learn some understanding of what these features might be and then you can use that basic understanding to then create new molecules that might have different compositions of these features that are likely to have similar types of properties or properties that you want. Those new molecules are not ones that exist in the real world yet. Perhaps this is the generative side of it. You're generating these new ideas based on an understanding of what features might be doing in molecules.

So could a parallel be there? You take, I don't know, let's take like Stephen King as an author and you train the model on all of the Stephen King books and his writing style. You could then use AI to generate what might be considered a Stephen King like book. So because it's trained on that data, it can then create new things with that information?

Yeah, exactly. It'll have, you know, the model will identify some features that are unique to Stephen King's writing style and then we'll create new content that has similar feature kind of feature characteristics, but it will be new content, but it might read like a Stephen King novel because the features are the same. So that's kind of in the, the molecular side and then the third bucket is really on the, the kind of knowledge management or the, the clinical side. And a lot of that is based on just consuming large amounts of data from things like EHRs, electronic health records, clinical trial data, very much like textual data. So this data isn't as much, it's not as much like the physical stuff like that you would generate in a wet lab, but just all of the data that exists in the healthcare system and then trying to understand information from that data. This, this is also like publication data. So building large language models around, you know, all of bioarchive, for example, and then trying to learn or query or get like fine things in these large data troves that you, that humans couldn't, couldn't do themselves. So this aspect of kind of generative AI or just knowledge management around all of the digital data that exists is also like a huge area for drug discovery and also in the clinical trial space. So these are kind of the three buckets, the bio bucket, the physics bucket, and then kind of the kind of like the more tech bucket or the more like data oriented bucket that, that's how I look at AI right now in, in the drug discovery path.

I think the data one's going to be quite interesting. There's a story I heard and this was, this story is probably about a decade old. So the, the kind of technology is being used now. But for those who aren't British, Tesco is equivalent to Walmart.

Right.

That kind of thing. And they were, they have a club card and this club card is every time you shop, you use it, you get points. But equally they get the data on what you're buying. And they, they found out that they could actually link, potentially use this data to infer health records. And I don't just mean like you're eating high fats, you will get X. They thought that they could predict diabetes based on one particular behavioral change. And it turned out to be quite accurate, apparently. And what it was is people are creatures of habit. You buy the same washing powder, the same detergent for years. But one of the symptoms of diabetes is you get particularly itchy, undiagnosed. And people would change their washing powder after years of using the same one because they think it's linked. And they said that this small amount of data, and that's very BAS level data. So if you think that's the kind of thing you can discover when you start to do it with new technology and new data, it's going to be incredible.

It's not just about data mining. It's about being able to ask questions of the data in an, kind of an arbitrary way. And that's, that's actually really, it's really fascinating. Just as an example, one of the things that was interesting about that I heard about OpenAI's relationship with Microsoft, for example, when they were putting GPT4 and integrating it into Microsoft Office products. One of the interesting things about like a program like Excel is Excel is a super powerful program, but most people don't know how to access all of the menus and all of the functionality because it's just a very complex program. And so most people use kind of the top level, the top level functionality of Excel, but being able to have a language model that allows people to query the thing that they want, that can then access the appropriate tools essentially opens up the entire power of the software to an average user without having to, without having them to be required to learn how to use each of those tools. And I think that a lot of the kind of the generative power and things like language models and is doing that which allows like a lot of the, the functionality of data that might be locked away or even just unknowable to, to general audiences, allowing them to be able to interact with that data without having to know that it even exists in advance. And I think that that's a huge step forward for Knowledge management. And it's just a big lever on. On how people can be productive.

I mean, from like a sales perspective. HubSpot has a lot of this AI stuff already built into their CRM. So you can just query it and ask it questions about contacts, about reports, whatever it might be, and it'll just like give you that information so you don't have to become like a HubSpot admin to learn how to create reports, learn how to build lists and go through all that process. You have the ability now to just ask a general question and get the answer that you need.

Yeah.

I think there's a researcher named Andre Karpathy who used to run Autopilot at Tesla and at OpenAI, and he has a tweet which is quite prescient, which he says, the hottest programming language is English. I like this quote. It kind of went viral, but it's true. The ability to be able to query things in natural language is an extremely powerful tool in all of knowledge management as well in commercial organizations.

I mean, even creating. Right. I mean, if you wanted to create a script to handle something on your desktop right now, you could just go to ChatGPT, tell it what you want, it'll give you the script, and if you don't know how to run it, you can ask it, how would I deploy this? And it'll walk you through.

It's incredibly.

And then if you run into an issue. Yeah, you run into an issue, you just ask it to debug it for.

You and it'll do it all for you. It's incredibly useful.

It even debugs itself, like on a very base level. The other day I was asking it to create an image and I was playing around with it with my daughter to show her how to use some tools in AI, and it said, we can't do this due to copyright. And I said, rewrite my script to get around the copyright and produce the image I want.

And it just did it.

Well, that's the other interesting piece is like when you ask, when you ask ChatGPT or some of these systems say, hey, what questions should I be asking? Then you actually get a lot of really great questions that then you can answer those questions and then it can generate the output that you're really looking for because it's going to ask you questions that you may not have thought of on how to prompt it.

Yeah.

Which is interesting. When you talked about you have to ask the right questions of the data, you have to ask the right questions of the models in order to get the right response. Well, what if you could all of a sudden start asking the AI model to give you the right questions to be asking? I think that comes down to an interesting world that we're in at that point.

Yeah, I think people, there's an overestimation and an underestimation of some of these models in what they can do. They're extremely surprising and powerful because of the amount of data that has gone into them that they can, they can generate things that are, that are quite strikingly accurate and interesting at the same time. They can entirely break because they are based on rules that are hackable or like breakable. Like there are little things in the technical nuance of how these models work where they can, you can just completely break them by doing simple things. And so the kind of brittleness coupled to the impression of sophistication is this, it's kind of a weird, a weird dichotomy. And so you don't want to kind of like overestimate what they can do. But at the same time they're oftentimes quite surprising in terms of how good they are.

One thing that's interesting, and I'm curious to get your take on this because there's, I think there's two trains of thought, which is one is you have these like big massive data models that are like, take ChatGPT as an example or GPT4 as an example. It's got everything right. It's just like it's got all of the information in it. So a lot of the responses then are the average of what exists out there. Yeah, but then there's also this other concept which is what if you train like a very niche model around like one very specific thing, say protein folding as an example, like now you're getting better results from these smaller, more niche models. So do you actually see a world where there's like hundreds of different models that exist that and then you pick and choose which one you want to use for different purposes or what are your thoughts on that?

So, so the way, the way that I think about this, and it's kind of like a visual way, there's this idea of these foundation models and then there's idea of like fine tuning these foundation models. So if you think about like what the purpose of one of these deep learning models is to do, it's to create, it's essentially to, to. They're, they're function approximators. So they're trying to create a, a mathematical function or a mathematical set of equations. That approximates some underlying reality. And the way that I kind of visualize this is just imagine like you have like a fishing net and you kind of throw a fishing net over like a space and like you have all the little knots in the net, right? And those knots are all of your data points. And the goal of the model is to have the fishing net adopt the shape of whatever you're throwing it over. And then you're, you're trying to make it so that you can basically predict the, the distance between the data points, like the, the gaps and the nets, where you can kind of interpolate in those areas and generate new things. But your data points are kind of like the points on the net. And I think about like a foundation model as kind of being like a uniform net. So like every, like every square in the net is the same and it just kind of like drapes over this, this, this, this distribution. When you are fine tuning like this kind of model, you're essentially like pulling parts of the net together to have them be more concentrated in certain areas, which may also stretch out other areas. So you may get better at doing the specific region that you're actually trying to model. You get kind of denser data points there, but you kind of stretch out other regions. So you may actually get worse in other areas.

And so because it's probably a balancing act there too, right? Which is if you're getting a lot of data in this one area, there might be other functions or other things that are out here that now are not as sure that might be useful in a broader net, but not so.

Much a hundred percent. And so this is like the difference between like, like what, what would be considered these big, kind of trained on the Internet foundation models, which are kind of this blanket. And then when people take their specific data, they're basically trying to like, like bunch up the net in certain areas and get like denser, denser representation. So they're better there, but it may get worse in other areas. And so you'll end up like, it really depends on what your, what your goals are. You know, a large language model like GPT4 is a, is a great foundation model, but if you're a company that has, you know, very specific data, you might want to train on that. And this is where things like retrieval augmented generators or rags come in, where you're, you have like a base model and then you're pulling in data from specific data sources to augment that model. It's not necessarily training or fine tuning the model, but you're prompt. Engineering the model to be better at certain areas. So yeah, there's going to be lots of specific use cases and use cases for big foundation models as well.

So I have. This is all really interesting. I can carry on with this stuff forever. But I have one other question. Before Harrison asked the last question we normally ask because I know we're coming up on time. Um, but talking a lot about AI and drug discovery is crazy interesting and I really enjoyed the explanations. Do you have a viewpoint on how it's going to impact sales and marketing in that kind of area as well? The AI that you've seen.

Are you talking about how AI will affect the commercial process or how AI will affect the types of products that are sold?

Probably the prior.

Yeah, the priority.

The commercial aspect.

Yeah, yeah.

Well, I mean there's clearly the AI. There's clearly tools like Harrison you mentioned that are built into HubSpot and Salesforce and other tools that will be, that will be useful, you know, potentially for things like insight generation. There's obviously tools that will accelerate things like prospecting, I mean, all of the language generation stuff. So there's just going to be a lot of set of tools that will accelerate the ability to be better and perhaps just more effective in the way that you communicate with customers. Probably tools that will be better at designing things like territories or forecasting. A lot of these things have existed in some form. I don't know if deep learning is going to impact them as much. You know, if deep learning specifically, although the implementation of large language models, I think it will. It'll just be like a tool set in terms of, you know, how it will affect a salesperson's communication with customers. I think that just like all of the other trends, it will just enable customers to be more savvy and salespeople need to become more savvy as well.

Yeah, I was going to say I think we're going to end up in a host pathogen war. You know, like when you see the, the immune system trying to evade the virus and vice versa with not to call salespeople viruses or customers. But you know, this interaction of us trying to get our message out and inboxes and LinkedIn trying to protect people from.

Well, so yeah, there's going to be, there's going to be a lot of filtering of AI generated emails because spam will probably be augmented. But also, I mean, the general trend of customers just becoming more informed has been a trend that's happened over a long time. Like customers will be able to get more information faster to come to sales interactions. And this is just going to continue to augment that. They'll be able to do more on their own.

You gotta be less transactional, right? It's like if, if. If you're thinking about things from a transactional perspective, that can be probably a lot of. It's automated or, or the customer's gonna have so much information that they'll be able to make these decisions on their own. You have to be able to. To connect the dots that maybe connect the people within an organization, right? So if one set of people have a particular problem and another set of people have a particular problem, you have to be able to connect these things under, like, a unifying problem statement to bring people together to ultimately make, like, a group decision. And I think that is probably more so where we're headed from a sales perspective is how do you. How do you navigate the entire, like, buying committee in an effective way? Rather than thinking transactionally of like, okay, I need to know these features and benefits around my product because all of that is going to be easily accessible. Well, Jason, it's been a pleasure. We could talk about AI all night, but we got to get going. It's great to have you on the show, Sam.