Kalim Saliba | The Loop is the Future: Scaling AI in Drug Discovery

HUMAN & AI ·

Show notes

In this episode, we sit down with Kalim Saliba, Chief Product Officer at Dotmatics, to explore how AI and autonomous workflows are transforming drug discovery R&D.

Kalim brings 17 years of software orchestration experience and a "Ted Lasso" coaching philosophy to the challenge of moving beyond isolated protein structure predictions into scalable, lab-in-the-loop pipelines.

Why are life sciences teams racing to eliminate software silos? How do AI models move beyond structure prediction to simulate binding affinity and off-target liabilities? And what happens when an AI agent attempts to design a protein starting from a massive 16,000-word prompt?

Kalim unpacks why generative inference must be anchored within governable platform harnesses, how computational startups and Big Pharma are pushing AI-designed candidates into clinical trials, and why autonomous "labs-in-the-loop" will become the central engine of discovery in the future.

Transcript

I think one of the biggest data quality problems the industry has right now is one of context and missing contexts, right? Because even if you have data and you're sure it's the right data and it's represented correctly, if it's not contextualized within the appropriate surrounding data, it's very hard to really understand what you're supposed to do with it. Welcome to a brand new episode of our podcast, Human and AI Mind Machines in a crate in the sand. Thanks for tuning in to our geeky podcast where we discuss the fascinating field of AI and machine learning, corporate craziness, passion for technology and the role of humans. We're your hosts, Oli and Abuy, and it means the world to us that you tune in to listen, learn and grow with us. And today we're heading into the lab. AI has already changed how we predict what a protein looks like. Now. The races on to design and make molecules that actually work as drugs with AI agents writing their own scripts and running their own analysis along the way. And that creates quite an interesting tension. So in software, you can copy a project with one command and get the same results over and over again. And in a lab, that's still kind of a dream. Recently, AI agent designed a protein starting from a 16,000 word prompt, writing its own script. As it went, impressive, but nearly impossible to repeat exactly. The 1st run can be creative. By the 100th run, it has to be a reliable product. How do you let AI build freely while scientists can still inspect it, trust it, and sign off on it? To help us detangle that, we have someone who sits right at the intersection. Uh, he has spent 17 years growing product in engineering teams and scaling SUSS businesses. Today he's the chief. Product officer, that's a cool, cool name anyway, at Dot Mattox, leading the Duma platform. He's also a writer, a public speaker, Hauldas, a photographer, I love this, and a self-declared follower of the Ted Lasso School of Coaching. Please welcome Kalim Saliba. Thank you guys. Really great to be here. Appreciate your time and the opportunity to talk about these really important and exciting subjects that are impacting our industry and our customers. Yeah, we can't wait to dive in. And before we do so, we have a little Challenge for you, Kalim. Since the first run can be creative, as we heard in the intro, why don't you ask your first run answer? So maybe you could introduce yourself real quick. Who is Kalim? What is it that really drives you? Why don't you go ahead? So, Kaliem Saliba, I'm the chief product officer at Dotmatt, as you mentioned, I'm leading product and engineering, heavily focused on our lab informatics platform called Luma. I've been with the company for about 4 years and prior to that, I worked at a company called uh Peas Systems for about 15 years, which is a leader in low code data management and, uh, workflow orchestration. A lot of important principles and and concepts I've kind of been a part of over my career that have a lot of relevance to life sciences, and I think what drives me and what I'm the most passionate about is kind of using the knowledge that I've learned, the leadership skills that I've gained to work with teams and work with software and technology to really do that, which has been very difficult. You know, in some cases seemingly impossible, and especially now with the advent of AI and all of these amazing technologies, like the ability to unlock value if approached in just the right way is incredible, and I think that's very motivating to me and the, the teams that I work with, just really trying to do whatever we can to bring together software for the benefit of our customers and patients and others that are really stand to benefit significantly from the work that we are doing. Yeah, that was almost 60 seconds, I would say, right? Hey Kalim, so you, you know, you, you mentioned something on Ted Lasso school of coaching, right? So obviously maybe for the audience worldwide, you know, give us, give us a bit of your perspective. We haven't seen maybe also the show, not yet, right? So what actually means for you, Ted Lasso's school of coaching? Why this matters for this conversation here with us? Yeah, that's a great question, and it's an interesting question because I am not at all a sports fan. I, you know, I think, um, you know, there's, there's always a running joke with my friends that when an important game was on and they were all, you know, we're all talking about which team do you prefer, who do you like, my go to was just, yeah, go sports, like very generic, like, I don't even know who's playing in this game, but happy to be a part of it with you guys. I think for me, Ted Lasso and really like the coaching style that's embodied by the main character. Ted, it's an exaggeration of what good coaching looks like, I think. And by being an exaggeration, it highlights the things that I think a lot of leaders try and strive to be really empowering, being able to connect with people at a very, you know, personal level to understand their specific struggles and then their broader struggles within the team, but also knowing when to be decisive, right? Knowing how to balance distinction between decisive and empowering leadership is something that I'm not perfect at it. I think it's, it's always going to be a lifelong, uh, piece of learning for me, and I think Ted Lasso represents that objective for me, and I believe others, which is why it's such a popular show. I will say though, I've watched it way too many times. I think, I think I'm on like, I think I'm on like the 5th run of the, of the earlier seasons, and now they, of course, they just released a new, a new season that's about, um, the women's team, so we're watching that with my wife as well. So yeah, it's just been a nice kind of thing. It's been in the background that's just fun to watch, but also represents, I think, you know, what I strive towards from a leadership perspective. Yeah, we can see that you're really radiating that. And I also, I I'm such a big fan of that show. So we touched upon Ted Lasso now, but um actually your your day to day role um is like the CPO at Dotmatt for the Luma platform. Maybe for those in the audience who don't really know what Dotmatt does and what Luma is about, can you maybe share a bit? About like, how did you get in that space and what do you guys actually even do? Sure. So Dotmattox is essentially a life sciences and material sciences software company. We have about over 800 people, many of them with advanced degrees and PhDs. Over 2 million scientists use our software and we have, you know, around 10,000 customers and what we do is really bring together a combination of highly specific, precise scientific software within specific Verticals and domains such as proteomics, chem informatics, biologics, cloning, but we also are working on Dotmattic Luma, which is this cohesive lab informatics platform that drives precision in how materials are represented and in how processes are orchestrated. Because if you think about the scientific landscape, there's so much software, right? There's so much software, and it's so precise in what it does, and the number of software is growing, of course, with the advent of AI, lots of our customers are writing their own software. And so how do you actually stitch together work that's being done in these different pieces of software across scientific functions in a way that allows our customers to extract higher order insights around what it all means, how is it all connected together. So, that's really what we're driving forward as a company is like excellence in science, working with our champion and challenger brands that we've acquired over the history of the company, plus. Excellence in scientific orchestration and lab informatics, really driving autonomy and effective use of AI in end to end capability sets that really bring the context cohesion that's needed to apply AI effectively. How did I get into this space specifically? I, you know, I was always interested in life sciences. After my engineering degree, I actually almost went into med school until my father, who was a physician, lovingly talked me out of it. Let's just say. And encouraged me to go use my my degree in in in other pursuits. But, you know, hearing about the opportunity at Dot Mattox to work on something as really groundbreaking and comprehensive as what we're doing here was an amazing opportunity that I just wanted to jump out. And I know a lot of people on the team are, are there for a lot of the same reasons. So it's a really good group of people working on a really important mission. Hi, Colleen, thanks so much. So, drug discovery in an area and then protein, right, and documenting about the lab and moving from lab. To let's say the more the manufacturing space, but if you engage with the community on on drug discovery from early innovation to scientific aspects, right? One of the key themes over the last, let's say, a couple of 4 or 5 years is now radiate around alpha fold, right? When somehow the jet GPT momentum was right in the consumer world, we have the feeling the alpha-fold did a similar kind of aspect somehow, right? It showed us what proteins maybe looked like, it earned, you know, creators a Nobel Prize. And then independent analysts found out that researchers using it submitted over 40% more new protein structure, but it hasn't led yet to a new wave of new drugs, has it, right? Knowing the shape is only half of the story probably or how is AI now helping teams getting from predicted structure, probabilistic stuff, right, to molecule that, you know, works as a drug and brings this new innovation, or are we at the pivotal stage here? What's your view on this kind of, let's say momentum? I think there's a lot that has to come together in order to harness the advances in AI in an operational way, in a drug discovery pipeline. I think typically these days, the advances in AI and the application of AI within the pharma space, it exists, right? It exists in spades, but it's not necessarily as connected to the operational way that scientists and different functions in the company do work, and I think that has to really change in order for AI to be effective. But I think there's Combination of things. So, let's start with the AI models. There've been massive, massive advancements. You you mentioned alpha fold. There's also bolts, RF diffusion. There, there's a growing number of increasingly capable models. And I think the first thing that has to happen is the accuracy of these models and the capability of these models has to improve. And I think as you were alluding to, and this is already starting to happen, moving beyondtru prediction into things that are more tantamount to simulating aspects of the biological system, things like binding affinity. selectivity, really understanding like where does the liability, which would be the thing within a protein or therapeutic that might cause off-target effects or not be as selective as as you want, how do you actually have a model predict what that looks like? How do you have a model predict the ability of a of a molecule to bind to your target of interest? How do you use models to look for targets and expand on our understanding of biological pathways, which is a fundamental part of Really any drug discovery pipeline. So I think that's number one, like continuous advancement of the models, and I do truly believe we're going to see advancements as a result of a lot of the, you know, change and rapid improvement in these frontier models that are going to continue to surprise us in terms of how capable they are. But then I think the other big thing that's needed and that frankly is, is kind of missing and that we're trying to resolve is that the software systems have to support the follow through, because if you think about how AI is used in these pipelines, it's at discrete points in the process. You're not really sending AI a prompt and saying, hey, go discover, make tests, analyze, and do a tech transfer and then manufacture the drug. That'd be lovely if you could have one prompt that did all of that, but that's just not how it works. So really, I think that the market and our customers are thinking about how do we insert AI. Within software systems that allow us to orchestrate its use in a way where we can harness what it's good at, but make more predictable that which it's not. And I think all of those things really have to come to play in order for the space to see some more of the transformative. I'll say transformative end to end. Capabilities that we're looking for. There are plenty of transformative examples of AI being used to predict structure better or with significantly higher performance or to do binding infinity calculations or to even do de novo protein design like Bolt'sGen, for example, uses generative model to actually predict the protein. That will uh feasibly be able to address a particular target, but again, that's just one part of many, many, many, many steps that have to happen to validate and understand whether or not that thing is therapeutically significant and viable or not. Very well put and also interesting how you described like that end to end process and how AI can can accelerate that and where it really can unfold a lot of value. Do you feel like there's certain organizational structure? um or setups that are have like greater benefits in that space like to win that game and I'm thinking about, I don't know, we hear a lot about like the AI native companies, a lot of startups out there, as opposed to like maybe big pharma tech companies. Do you have a feel on like who will be able to actually translate that AI accelerated end to end pipeline in drug discovery quicker or better? Yeah, it's a good question, and I think um it's hard to say for sure which style of company is going to lead to the most disruption, but I believe that for sure there will be disruption, right? It's a question of where is it going to come from, and it may not come from one or the other. I think if you look at a lot of the AI native or AI focused startups in the space, there's typically heavy involvement from pharma, right? Either the pharma is providing. information about the targets and pathways or providing funding to receive understanding or maps that increase their understanding of the biological system or in some cases, you know, out licensing some of the molecules. So there is a very strong relationship between pharma and some of these AI native startups that we're seeing. Like for example, recursion does a lot of work on disease biology, doing experiments on cells and then imaging them to help. Increase the understanding of how the biology works, and they have tons and tons of data on that. They're working heavily with Roche, Genentech, and Sanofi. Isomorphic is interesting because you mentioned Alphafold at the beginning of the call. It's a DeepMind spin-off from Google, which is essentially they took Alpha Fold 3 and improved upon it significantly. So they are the most computationally focused, but they're relying on partnerships with Lilly, Novartis, and most recently J and J, you know. One of their models is to source targets from these customers to work with them to help build out using their computational pipeline, potential candidate molecules. The most interesting one is probably in silica. They're they're the furthest along. You may have heard about them as, as I think in early September, they just entered phase 3 clinical trials for one of their pulmonary fibrosis drugs. So this would be, if it's successful, the first end to end AI discovered. Drug that's brought to market. So lots of exciting stuff is happening, right? And like different companies have different approaches, but in all cases, there appears to be heavy involvement with pharma, and that's not all that surprising because these pharma companies are the ones that have all of the data. They have all of the relationships cataloged. They have like decades of experience in understanding disease biology and information about targets and what has worked and what hasn't worked. And so, I don't know exactly which style is best for the time being, I believe it's more about making sure folks work together to bring the best of what each one is able to deliver, to bring something to market. Do you think they speak the same language? So usually, right, traditionally you have a little bit of the designs back, you have the task, you have model, these are different personas, these are different scientific backgrounds or educational backgrounds even, right? So discovery has lifted, you know, the chemist. Design the components, the people running the lessons and screens, and then you have the modelers who are rather computational work, let's say, and then, you know, you have this interface between prediction and the lab result, how to compare it and how to speak actually the language. Do you think AI will, you know, enable this kinds of or is everyone moving now into computational design, computational modeling, computational this, so AI is the world, Colleen, in these kinds of worlds. I think we're probably a ways away from AI doing everything, but I believe we're closer than people realize to the market trying to address higher degrees of autonomy in the way agentic systems work, right? And I think that's in part due to the massive advancements in all of these models and the increasing focus on orchestration and making sure like the Discovery pipelines are well engineered systems that are able to kind of employ. AI and humans across the design make test to decide cycle at the right opportune points. In terms of, you know, is it the same language? Like there couldn't be more variability in how science is, is discussed and how it's interpreted across these different phases of research, development, and manufacturing, and it very much is a different language, and I'll, I'll give you a small personal anecdote, you know, as, as an engineer by training, I joined the company. And I think naively thought that learning the science, working with my colleagues to understand where are some of the greatest problems would be like applying, you know, an engineering mindset to understanding the concepts, the language, the relationship between those concepts. But there are things that are very fundamental things in the space that don't even called the same thing by the market or our customers. And so one of the things that emerges as, as a really important. Characteristics of these systems and in driving, maybe not necessarily commonality, but interpretability across the full life cycle are ontologies, right, that allow our customers to specialize the science to the way that they think about, talk about, rationalize, conceptualize the work. And I think without that, it's very hard because without that, sure, you have data, you can put it all in one place, but if you can't understand, you know, how does design A actually relate to Protein panel B relate to assay results C and then the decisions that derive from that and all the things that have to be described in order to analyze and understand that data. It's very hard for people, even harder for AI. And it's interesting, right, because you still have to put in so much groundwork to actually establish that and make those connections and then lay that foundation before you can actually start to like reap the benefits that AI can enable. And I guess It's like another aspect that um always comes up is like the question around like data quality. I assume that there's lots of context and like data being generated um throughout like all the testing that you have to do, a lot of the screening. …

Full transcript (WebVTT)