Scientist Stories: Charbel Eid, Enabling Faster Therapeutic Antibody Discovery through AI/ML

Hello and welcome to today's webinar on enabling faster therapeutic antibody discovery through an integrated artificial intelligence machine learning wet lab platform. I'm Tom Speedy, product manager for synthetic biology at Integrated DNA Technologies, and I will be serving as the moderator for today's presentation. We are joined today by our guest speaker, Dr. Shahbaud Idd. Dr. Eid is a leader in life sciences R&D with over 10 years experience blending engineering, science, and technology to solve big problems in the life sciences and biomedicine. Dr. Eid is currently serving as the head of platform development at Big Hat Biosciences, leading the growth and expansion of Big Hat's Milliner platform. Before joining Big Hat, Dr. Eid was manager of technology development at Gardant Health, and head of lab technology at Driver. Dr. Eid holds a PhD and a master's from Stanford University, where he was a Lieberman Fellow specializing in microfluidic assay development, and he has a bachelor's from Carnegie Mellon University. For today's talk, he will present an introduction to BigHat's integrated AI/ML wet lab platform and explain how BigHat's platform enables faster discovery for better antibody therapeutics. In recent years, therapeutic antibody market has experienced significant advances that have positioned the drug class as a promising tool in the treatment of a plethora of ailments. Yet therapeutic antibody discovery takes time from targeted identification to validation. Advances in synthetic biology are accelerating therapeutic antibody development, and in this webinar brought to you by Integrated DNA Technologies, Dr. Eid will discuss recent antibody discoveries enabled by synthetic biology. Before we begin, I will first provide a quick introduction to IDT and our capabilities and services before handing over the presentation to Dr. Eid. For those not familiar with IDT, we were founded in 1987 and have since assumed a market leadership role in DNA and RNA synthesis. As an operating company of Danaher Corporation, IDT serves more than 130,000 life science researchers in over 100 countries. An average of 98,000 highest quality custom oligos are produced at our manufacturing sites per day. These are delivered with a fast turnaround and backed by expert design and application support. Researchers trust the better sequence fidelity and higher yield of full-length product delivered by IDT, making our oligos the industry industry standard for quality and purity. As I mentioned earlier, synthetic biology is inspiring a new wave of innovation in society. In its simplest terms, synthetic biology is a multidisciplinary field that applies engineering principles to create recombinant products. Synthetic biologists study natural biochemical pathways and functions of biomolecules and through genetic engineering techniques improve existing processes or confer new functions to them. Products of synthetic biology have been applied across many industries, including biomaterials, biofuels, renewable chemicals, therapeutics, food and agriculture, environmental remediation, and many others. IDT supports the build segment of the design, build, test, learn cycle. Through our double-stranded gene fragments, cloned gene constructs, pooled DNA oligos, and a suite of genome editing solutions. IDT offers various high-quality synthetic DNA products with fast turnaround times and custom options. There are 3 types of double-stranded DNA fragment to suit the needs of every synthetic biology experiment. EBlox, gBlox, and gBlox HiFi are versatile double-stranded DNA fragments used in gene construction, gene assembly, and other genomic assays. Custom genes are NGS-verified sequences cloned into plasmids for protein or antibody expression or subcloning. You may already be aware of our other comprehensive genome editing research solutions We offer custom oligos including modifications, formulation, pooling, and duplexing services. We have probes, assays, and reagents for PCR and qPCR, nucleases and guides for CRISPR, constructs for RNAi, and sample prep products for next-gen sequencing and various other reagents. And with that, I'd like to hand the presentation over to Dr. Charbel Eid.

Thank you all for giving me the chance to give this presentation today and for attending this talk. So as my name is Charbel Eid, I'm the head of platform development at Big Hat Biosciences, and today I'm excited to tell you about how Big Hat is building better biologics faster through artificial intelligence guided design. So in this talk, I'll be introducing BigHat's technology and Millinera platform, highlight some of the unique characteristics that make our platform differentiated, and talk about how our platform is helping BigHat produce better therapeutics faster. So as I mentioned, our mission at BigHat revolves around creating safer and more effective antibody therapies for patients using a mixture of machine learning and synthetic biology. We are currently focusing on next-generation antibodies, though as you will see through the stock, our approach can easily be generalized and applied to many other types of formats and molecules. And we're targeting some of the hardest and most impactful diseases known today, you know, cancer, infectious diseases, and many others. As many of you may know, traditional monoclonal antibodies have had many successes over the past 30, 40 plus years and have made an impact across many areas. However, traditional antibodies have limitations. They can often only bind one target, have low specificity, and as a result, there are many targets that remain undruggable. So next-generation antibodies have gained a lot of excitement as a potential alternative to traditional antibodies, and they have the potential to overcome many of the challenges. And they offer the potential of being more specific, in many cases bi- or multi-specific, and also more potent. And in the last 10 years, there have been some initial FDA approvals across different formats of next-gen antibodies, which has very been very exciting for, for this field. However, despite their promise to overcome the limitations of traditional mAbs, they have their own challenges, and that can often make them difficult to move forward into clinical studies and have them actually be used for patients, to help patients. So our traditional monoclonal antibodies are mature, better understood, and have multiple developability pathways. Next-generation antibodies are more challenging and less understood. They often suffer from things like low yield, low manufacturability, inconsistent stability profiles, and all of that makes it harder to produce them at a clinical grade. This is widely acknowledged as a major hurdle for the widespread adoption of this otherwise very promising technology. And in the last 10 years, in addition to the notable successes that they've had, they've also established— the last 10 years have established that there are several challenges that next-gen antibodies carry with them that need to be overcome for broader utility. So at BigHat, our vision is to use our AI-guided antibody design platform, which marries high-speed wet lab with active learning design processes to broadly and holistically create better next-gen antibodies in a faster way. By integrating wet lab with AI/ML, we are able to iteratively refine what started out as rough initial designs and to optimize clinic-ready next-gen antibodies. Very critical to our success is the ability to propose mutations in a rapid succession, which enables a rapid feedback loop. And this enables us to quickly narrow down the parameter landscape and converge on optimized properties. This can be very hard for a person to conceive how all these different attributes that are fairly complex on their own and become even more so when taken in aggregate, how they're all related to each other. And this is kind of one of our advantages over traditional approaches. We are because of our ML-driven approach, able to take antibodies that have many biophysical warts in them and improve them and treat them holistically, where we are able to optimize multiple properties to make them easier to engineer. Our platform is fairly generalizable. We are, in a way, in a sense, able to optimize any property that can be quantitatively measured. This is very exciting because, you know, think about meat and potatoes common properties to more bespoke custom functional assays and properties. We're able to do it all at scale in an automated way, and this is what enables us to succeed over traditional approaches which are not usually able to do so. So our design cycle loop consists of two main phases. Designing a set of promising AI molecules with AI/ML and then building and testing them using the wet lab. We leverage AI/ML to design variants that are most likely to be high quality, run them through our laboratory workflows, generate and then QC data, which we then use to teach and improve AI/ML models. And this teaching and improvement is then translated to the next set of batches, which kicks off another loop of this design-build-test cycle. One of the key features of our approach is that this learning system gets better with more use, and this applies both within a campaign or a program. You can imagine getting better and better at identifying promising candidates for a given target, but also because of the overall generalizability of our approach, this also applies across campaigns. So by going through the cycle, we are able to contribute to the development of all our cycles on all our campaigns, in a sense, through this distributed advantage. Importantly for us, this design-build-test-model cycle involves models in every aspect of the process. We like to say that all roads lead to the models. Looping in ML allows us to start from a very small number of sequences, as little as one really, and build a dataset entirely in-house that gets us better models and better antibodies in very small set sizes. So where someone might think of big data, you know, a commonly used term, as consisting of millions or even billions of data points to start seeing predictive power, we're able to generate substantial improvement as we show on this figure here on the right in as little as a few hundred data points. This is very exciting because not only does it enable us to build better antibodies, but also build better processes. Making the most out of a relatively small data set means that we can spread our overall capacity as a platform and as a company onto multiple programs or projects, and in doing so, we're able to amplify our impact as a company. All right, so zooming in a little bit on the lab, our workflow starts with DNA synthesis, which is based on the aforementioned in silico produced designs. This is then followed by protein production and purification, which currently exists in multiple modalities in-house, but we primarily use cell-free protein synthesis for screening purposes. Batches are then run through a battery of assays, some that we call core assays, which for us means we run them in most, if not all, programs, and then others that are more specific to individual programs or targets. We're able to go from designs to data in a matter of a few days, and that's in part because of the efficiency of our lab operations and our wet lab platform. We typically process batches in microplate format. So you imagine 96, 384 wells, etc. But different assays have different configurations and capacities, which allows us to on the fly adjust the batching to fit the capacity of the different assays in the sequence. Throughout our platform, we invest in gold standard characterization instruments to ensure that the quality that comes out of these instruments is top-notch. But then we supplement that with a lot of investment in sophisticated custom-built QC pipelines, which ensures in conjunction with high-level instrumentation that the quality of the data that we're feeding back into our AI/ML models is of the highest quality to trigger the next round of designs. So all we need really to set the wheels in motion are starting antibody sequences, as I mentioned earlier, as little as one will do, along with target proteins and some established design goals. These can be set either internally or in partnership with a partner or collaborator. In this example here, we specified goals around three different parameters: affinity, stability, and solubility. We run through the BigHat platform, we iterate on designs, and then The outcome of all this is one in which we convert an initial sequence into what we call a Big Hat lead and then make substantial improvements along all the relevant axes. This is what our platform is capable of and kind of how it serves the overall greater Big Hat mission. So now that we've given a brief overview of the overall machinery, I'd like to talk a little bit more about some aspects of Big Hat's platform and what makes them unique. I'll start with REKI, which is Big Hat's bespoke cloud-based LIMS++ system, as we like to call it. Uh, REKI is where, in a central spot, all of our biological data and metadata, our AI/ML training datasets, experimental results, operational statuses, and logistics— they're all managed in a single API-driven core, and that's converted for the broader stakeholders at Big Hat through a web app for utility. Reki allows us to capture pretty much everything from sequence designs to order, ordering, tracking, aliquot generation. It allows us to streamline experimental data creation and ingestion, to onboard new experiments and instruments, design new experiments, and have reproducibility built into them. It also very importantly allows us to close the data ingestion to training to design loop. Using built-in automation. It's, in short, critical to everything that we do at Big Hat. I want to zoom in on one of those capabilities, what we call Reki Workflows. At a high level, Reki Workflows is a one-stop shop that allows us to dynamically track the journey of a batch of molecules all the way from conception through execution. It also allows us to track the status of multiple journeys at once. And capture critical information and metadata about each batch. So what you're seeing here is a snapshot which covers quite a bit of ground. In the top left panel, you can easily see an easily adjustable graphical representation of a given workflow. So what we're seeing here is a certain set of instructions that a batch has to go through from production through a sequence of assays. And on the top right, we capture various important metadata for each step of the workflow, things like which SOP it originates from, what samples are eligible for that step, etc. And on the bottom, we show this sequencing and status for many batches with colors indicating progress or state. So this allows us to solve many problems at once. One of these challenges is creating explicit and enforceable sequencing to our processes. This sounds simple in a way, but it can be very hard to scale across different programs because each one has their own set of assays and their own sequence of which assays should be done in what order. Workflows allows us to establish a very visible, scalable configuration for different programs without creating confusion in the process. Another challenge we are faced with is the need to track and broadcast what is happening in the lab and what is happening in our data pipelines. We want stakeholders to have easy visibility into progress states, and we want them to have that in a low-touch manner that doesn't require them to, you know, reach out to individual people or need to be on site to see what is happening. And then operationally, we want to take away as much of the uncertainty and on-the-spot decision-making as possible from lab processing. We don't want our operators to have to make on-the-spot decisions about what to do next.. And in a high-speed lab, which ours very much is, this means we need things like notifications. We want the interactions to be seamless. That means we want to notify people when they need to be involved and then keep them out of the loop when they don't need to be involved. And as we juggle our expanding platform with more programs, more steps, more assays, Workflows allows us to track all of the activities and all the data generation. And very nicely, it's just represented in a visual design that's pretty intuitive. It's stored securely in the backend and in a structured way that allows us to make sense of that data and use the data that's generated in each one of these assay steps for downstream analyses. Another critical aspect for our platform is lab automation. Different companies, they take different approaches to how to invest in automation. That can range anywhere from a lonely robot sitting in the corner all the way to more heavy upfront investments and large and integrated systems. Big Hat's automation strategy is built on flexibility and nimbleness, and I'm going to show what that looks like in the subsequent slides. This starts with our investment in automation systems. Our platform features a diverse array of instruments from automated pipettors to small limited function liquid handlers to more large-scale and highly powerful robots. They're all united by being controlled and communicating with our underlying software infrastructure. I like to say that our automation has grown with our company one robot at a time. And that means for us that we're very insightful and intentional about how we invest in adding more capacity and more growth into the platform to make sure that we're maintaining our high standards for the data that comes out of it. Okay, similar to other aspects of Big Hat, the automation suite makes heavy use and contributions to the cross-company advanced technology stack that we have. We have built and continue to build deep integrations between instruments, robots, and Recce. And this allows us to do a lot of important but also fun things like automate data ingestion, have easy progress tracking, mobile notifications, error handling and reporting, make all of these things more visible, more accessible, and more distributed in their, in their nature. But we also make heavy use of our cloud-based systems and processes that would otherwise be very operationally complex. One example of this is the process of autom— of handling, storing, and deploying automation scripts. And this traditionally is a very manual and laborious and sophisticated process that is very error-prone, where instruments are maintained at a local level in a more manual way. But our approach using our cloud-based general storage systems is much more robust. We're able to deploy changes and updates to our fleet of instruments in a single push implementation, which is one more robust because it protects against catastrophic local issues. It is distributed. The single push means we can, the click of a button, make sure all our instruments are consistent. And then this consistency is really critical because it ensures that our robots are proceeding with identical set of instructions. Which, given how much we care about reducing noise and variability and providing high-quality data, is itself very essential. So part of our platform's value proposition is that we can easily accommodate and scale programs and assays that are within these programs, and that allows us to go after the most promising targets and partnerships. And as you can see in the slide, program workflows— and you're showing 5 here— they will converge on some aspects like protein production. They might share a certain set of core assays, but then they will diverge for certain program-specific functional assays. Now, our automation suite reflects this. We invest in high throughput at these convergence junctions, then we invest in flexibility and the more bespoke functional assay suite. And so this allows us to pivot in whatever direction the company needs. We can scale a program that is selected to move forward to make it more high throughput. We can accommodate a new assay or new workflow via our lower throughput but still automated approaches, or we can even sunset the workflow that is no longer worth pursuing. So we're able to match Big Hat's overall ethos with the way we've created and grow our automation suite. And then this particular ethos is extended to the development process itself, where we also prioritize and care deeply about flexibility. It is very common to pursue a serial development approach when it comes to automation and scale. You can imagine something like showing the feasibility of a new assay developing it manually, and then later on developing an automated version of that process. But what often happens, the assay may not translate moving from manual to automation. You may have to go to the drawing board in some cases. In other cases, you might have to endure substantial friction losses in this transition. Performance may have to be sacrificed, and if not, then perhaps reoptimization is needed, which will sacrifice time. None of these things are abundant, but at Big Hat we're big proponents of a co-development model to assay development. For us, this means doing assay development on the automation itself, which allows us to iterate repeatedly and quickly. This is not trivial or easy. It requires substantial commitment and collaboration across all the development teams, but what this gives us allows us to do is have a much more frictionless development trajectory. Also allows us to deploy scaled and high-throughput implementations much earlier in a program's timeline, and thus that enables us to have more and faster screening cycles which improve our ML models and move our programs faster. So in that way, we are able to give our state-of-the-art machinery more chances to prove itself and to move antibodies further down the development pipeline. And I wanted to highlight also the rigor that we are committed to at BigHat. I hope that has come through in some of the previous slides, but I wanted to highlight it in a couple of different ways as well. So as you all would know, when it comes to models, the quality of predictions will only be as good as the training data you give these models. So we are consistently and very diligently thinking about how to improve the quality of our data, how to catch low-quality data, and how to maintain the health of our platform. And this manifests in several ways, but one of them is controls. Our approach to controls is very involved and multi-tiered at BigHat. We use controls of different types, end-to-end synthesis, assay-specific. We use them strategically and efficiently throughout our microplate format processing. The very process of introducing and implementing these controls requires substantial coordination across our personnel, lab staff, our automation and software to both introduce them, know where they are, track them, replace them as they go through the assay suite. But what this enables us to do is leverage our strong computational infrastructure to scalably and automated way track controls performance, which in turn means that we can keep a close eye on platform performance. We can catch drift and outliers, and very importantly for us, we can normalize sources of noise away to produce high-quality and reproducible data. And what I'm showing here on the right is an example of this in action. Where we have caught and acted on various trends and observations. And as an aggregate, we've also used this data to very effectively and powerfully normalize away batch noise to increase consistency of development and improve the quality of the data going into our ML models. And this rigor also extends to all aspects of lab operations. We very closely monitor instrument usage over time to get ahead of operational bottlenecks, to know where to smartly invest, to unlock capacity and achieve redundancy where it's needed. We also use state-of-the-art laboratory and facility state monitoring, both to catch any issues with things like temperatures, gas levels, et cetera, but also to have the ability to quantitatively correlate environmental factors with observed data. Very importantly also, and this is something I'm very proud of, we have designed an SOP and documentation system that allows us to put pen to paper all critical processes both inside and outside the lab to varying level of stringency as needed by the process itself. This is low touch where it needs to be but enables us to have thorough review of our documentation by different stakeholders. And this is both helpful to increase the quality of the documents and consistency of the processes and the, thus the data, but also it facilitates knowledge transfer and maintains operational excellence as we onboard new team members. And for a company like ours, which is growing rapidly, this is a very critical utility to get out of this process. And on the backend of all this is a traceable documentation system, which is easily accessible, ties directly into experimental workflows, and can be customized to the same level as all other aspects of our software infrastructure. This uses Reki as well, as another example of where Reki is being used at Big Hat. And as such, it is also state-of-the-art in terms of the traceability and security of the documents, as well as their visibility and accessibility for all stakeholders within BigHat. So now that I've described some aspects of what makes our platform unique, I wanted to end with a couple of quick examples of our machinery in action. So I'll talk about a couple of case studies in a couple slides. And this first example, we were faced with the challenge of needing to introduce a new assay modality and a new target. Along with a multi-objective goal of improving both affinity and stability. So on the left of the slide, we're seeing the journey that this process took, starting with a rapid onboarding process, took about 2 weeks, thanks in big part to our platform's existing infrastructure and ability to quickly accommodate new assays and new targets. We focused the first couple design rounds on improving stability while maintaining affinity, which you can see on the right, and we were able to achieve a substantial improvement in melting temperature. And then our last design rounds, they built upon the first two, where we took our stability-improving mutations and added on top of them affinity improvements that also maintained the stability gains achieved earlier, and you can see that in rounds 3 and 4 on the right. We then invested in replicating the best performers to validate the affinity and stability improvements and get more confidence in the output of our program. This entire process took about 9 weeks and 4 design batches of variants and is in large part due to the various machinery of the BigHat platform coming together to enable 1, quick onboarding, 2, quick iteration, and then 3, very high-quality data generation. And this, the second case study, uh, we wanted to induce pH sensitivity in an SCFV format antibody. It's worth highlighting that our platform is able, with minimal adjustment, to produce sufficient yields of different format antibodies. This further on its own opens up our ability to accommodate a broader set of targets, so we're not limited as much in in terms of easy-to-produce formats. So we developed the necessary automation with very quick turnaround, which enabled parallel processing of batches at multiple pHs, which was critical for a project like this. We then used ML and AI-based approaches to create a broadly diverse set of novel mutations, and we leveraged our controls approach to design a normalization scheme selectivity index, as you can see on the right there, that adjusted for all sorts of externalities and external sources of noise that were present in executing batches. But within 3 rounds, we were able to find highly pH-selective antibodies. It was progressive improvement, and then we were able to invest in replicating those. This itself was another 10-week or so process, and it's rather difficult assay configuration. This was another program success, and it's supported by our versatile Millner platform, the various pieces coming together to elevate and make sure that our programs are supported as needed. So in wrapping, I wanted to highlight our wonderful leadership team and call out our equally wonderful interdisciplinary talent and team across multiple functional areas and expertise areas. As you can imagine, what we're doing involves many pieces coming together in a fairly cohesive whole, and our approach is very much consistent with a need to be harmonized in how we operate across these functional areas. We are based in the San Francisco Bay Area, and we are growing rapidly and doing increasingly exciting things. So if you are interested in joining us or learning more, please visit bigheadbio.com, and we'd love to Love to chat with you. So with that, I'd like to thank you all.