Listen on
Overview
The chasm between preclinical models and clinical outcomes remains a persistent, costly challenge in gene and cell therapy development. When dealing with “living drugs” like engineered T-cells, the complexities multiply: each patient’s unique biology and prior treatments mean that even well-designed preclinical studies often fail to predict real-world efficacy.
This gap costs companies billions in R&D and delays life-saving treatments. Host Ross Katz speaks with Dipen Sangurdekar, VP of Data Sciences at KSQ Therapeutics, who discusses how his team bridges this divide using the CRISPR-Omics platform. His experience with highly personalized, autologous cell therapies provides a unique perspective on using data, even from “failed” clinical trials, to refine drug design and accelerate therapeutic development.
Dipen outlines how KSQ Therapeutics uses CRISPR-Omics for both target identification and direct therapeutic intervention, the immense data challenges of deeply profiling living drugs with limited sample sizes, and the critical role of a hypothesis-driven approach when traditional machine learning falls short. He also details the essential data engineering backbone required to integrate multimodal clinical data, enabling cross-functional teams to drive biological insights.
Key Takeaways
Clinical outcomes are the ultimate model validator, especially for living drugs.
Preclinical models, while essential, cannot fully recapitulate human physiology or the complex tumor microenvironment. Clinical data, even from trials that do not meet their primary endpoints, is invaluable. It reveals the gaps in our models, providing crucial feedback to refine hypotheses and drug designs for future iterations, significantly improving translation rates over time.
Deep profiling, not sheer scale, defines data challenges in cell and gene therapy.
Developing personalized, “living drugs” like engineered T-cells means extensive molecular profiling (single-cell RNA-seq, TCR sequencing, proteomics) on limited patient samples. This data depth, combined with low sample volume, requires a hypothesis-driven approach to analysis, rather than relying on large-scale machine learning models to discover patterns from insufficient data.
Data engineering is foundational for connecting preclinical and clinical insights.
A unified data backbone is critical for integrating multimodal datasets—from CRISPR screens and genomic assays to clinical trial records and patient biopsies. Reliable data ingestion, curation, inventorying, and tagging ensure that diverse data sources are connected. This enables seamless access for clinical and translational teams, fostering self-service analysis and reducing reliance on a single gatekeeper.
Prior knowledge and orthogonal evidence guide discovery with sparse data.
When sample sizes are small, data scientists must lean on strong biological hypotheses and Bayesian thinking. Instead of fishing for statistically significant p-values, the focus shifts to finding weak but suggestive trends that align with prior knowledge or are supported by orthogonal evidence from different assays. This targeted validation minimizes false positives and maximizes confidence in early-stage findings.
Related: CorrDyn helps biotech and life sciences companies realize data value through specialized data engineering and strategic AI guidance. We also offer data assessments to help companies understand their current capabilities and identify paths to accelerate innovation.
Full Transcript
Jason: Welcome to Data in Biotech, a podcast from CorrDyn where we explore how companies leverage data to drive innovation in life sciences. Every two weeks we sit down with an expert from the world of biotechnology to understand how they’re using data science to solve technical challenges, streamline operations, and further innovation in their business. Today we’re joined by Dipen Sangurdekar, VP of Data Sciences at KSQ Therapeutics. Dipen takes us into the world of cancer therapy, focusing on tumor-infiltrating lymphocytes and the groundbreaking applications of CRISPRomics technology. He and Ross discuss the challenges of translating preclinical discoveries into clinical success, the complexities of drug development, and the IND and clinical trial process. Dipen also shares why integrating data science with biology is crucial for driving innovation in cell and gene therapy. Here we go.
Ross Katz: Dipen Sangurdekar, welcome to the Data in Biotech podcast.
Dipen Sangurdekar: Hi Ross, good to see you.
Ross Katz: Awesome. So just to kick us off, could you give us a brief introduction to your background and what brought you here today?
Dipen Sangurdekar: Absolutely. My name is Dipen Sangurdekar. I’m currently a vice president, head of data sciences at KSQ Therapeutics. KSQ Therapeutics is an oncology biotech company based in Lexington, Massachusetts. Our mission is to solve or cure solid tumors with engineered cell therapies and what are known as cell therapies. If you’re familiar with cell therapies, they can be broadly of two kinds. One is autologous, where we take patient’s own cells, engineer them with either CARs or TCRs and put them back in. And the other broad category is allogeneic, where you take cells from healthy donors, potentially do some engineering to make sure that they evade host immune attack and put them back in after equipping them with CAR and TCR. So our cell therapies are called TIL therapies, which come in the autologous category, where we do all the engineering. However, instead of attaching a single CAR or a single TCR, we actually take the cells from patient’s tumors. These T cells have been trained by their own body itself to attack and fight those tumors. It’s just that they have gone through the process of attacking tumors and gone down the exhaustion path and perhaps need a little boost. So we take those cells out. They are called tumor-infiltrating lymphocytes, or TILs for short. We take them out, we expand them, recharge them and infuse them back. Our differentiation there is that using a platform called CRISPRomics, which is genome CRISPR screens, we have found gene targets which we believe by knocking them out really turbocharge these cells and help them overcome the difficult environment that these cells typically see, especially in solid tumors. So overcoming immunosuppressive environment, cytokine shortage and so on. We are using the CRISPR technology both as a discovery tool but as a therapeutic tool as well to modify our cells.
Ross Katz: Awesome. Well, that’s a fantastic introduction to KSQ Therapeutics and to TILs, tumor-infiltrating lymphocytes, and to your background. Before we jump into the fun data science work you’re doing at KSQ, can you just tell us, why did you get into this field in the first place? What led you down this path?
Dipen Sangurdekar: The first day of my graduate school, I picked up a pipette and started doing genomics experiments, which included doing genetic screens and a lot of transcriptional and microarray work. And having come from a quantitative engineering background, those early days in the lab really opened up the complex world of biology. Fundamental research, life sciences, and also what it could mean for health and understanding more complex systems such as human and disease systems. I really wanted to be on that interface of generating those hypotheses in a data-driven manner and having the ability to query those hypotheses in the lab. And doing that cycle-time which back then maybe it was not as obvious of a trajectory, but now, of course, you have an entire generation of scientists that are cross-trained between data science and experimental techniques. I think it was a matter of continuing to look for interesting problems to solve using my toolkit that was slowly developing over the years and an understanding of the platform, the genomic technologies, and one thing led to another and that’s where I am. Over the last few years, focused on the interface of how preclinical work, biology, what we call preclinical means when we use model systems either in vitro cell lines or in vivo data, how that understanding of mechanism of action of drug targets translates into the clinic and how we can learn from clinical data because we generate a ton of data even from a quote-unquote failed clinical trial. We can generate a lot of data that we can take back to the clinic either to diagnose or to improve drugs and being on that interface is just, I find it very fulfilling and thrilling and something that I feel I can contribute to in a very tangible way.
Ross Katz: Yeah, that’s awesome. And I really want to return to that question of how preclinical work and clinical work create feedback loops between each other and how you translate between the two. But before we do that, I want to return to TILs and the data-driven work you’re doing there. So you mentioned the CRISPRomics platform and that it has two applications that it’s used for. There’s the target identification aspect of it and then there’s the therapeutic intervention aspect of it. Can you just walk us through, maybe just a brief introduction to CRISPR-Cas9 and what it is and then how it’s used in these different contexts?
Dipen Sangurdekar: As perhaps many of your audience already know, the CRISPR-Cas9 system, which has won Nobel prizes in the field, is used to systematically create double-strand breaks in a mammalian system. So let’s say a human cell, you want to precisely at a particular genomic coordinate go in and create a double-standard break. Because of the DNA repair mechanisms, those double-standard breaks often get repaired in a way that causes what we call a short insertion or deletion. And if done right, if you find the right location in the gene, that can knock out protein function for that gene. Genetic screens, knockout screens have existed prior to CRISPR, but CRISPR has really enabled doing this at scale, especially in complex mammalian systems and human systems. What we have done and others have done as well is use libraries of these synthetic guide RNAs, which is what directs the CRISPR-Cas9 enzymes to particular location in the genome to cause those edits. We can create libraries of those and there are publicly available libraries as well as, it’s fairly straightforward these days to create libraries computationally. You can multiplex those and create screens where you take cells of interest — whether it’s a cancer cell line, whether it’s an immune cell, either in an in vitro setting such as in a dish or in a mouse model — and you conduct an experiment, what we call a forward genetic screen. Essentially you’re letting a bag of cells, each of which hopefully has one edit. And you can control your experimental parameters to make sure it’s only one edit per cell. You let them compete. What we’re looking for is a functional readout such as an expression of a surface protein or how well those cells kill a particular cancer cell line or how well those cells migrate to a particular location in a tissue. Once we do those experiments, you can sort either using flow cytometry or you can do extraction of a tissue and you can do NGS to find out which sgRNAs are enriched. That’s a proxy for which edits are winning in this almost like a survival of the fittest situation. And once you figure out which genes are knocked out, you can’t just take the top gene and take it to a target validation. You have to then layer in all of the biology and all the prior knowledge that we know about those genes. You can do pathway analysis, you can compare it to other screen data. And you have to be very judicious in down-selecting which of those genes we want to take to a target validation funnel. That’s how once it’s in the funnel, that’s your standard drug discovery — you can either design a therapeutic modality, such as a small molecule to target those genes or a biologic. Or in our case, design an ex vivo CRISPR editing system as a therapeutic by itself.
Ross Katz: My understanding of this is that the CRISPRomics-based system allows you to look at the entire genome and go gene by gene and do turn on and off decisions at the cell level and then see which cells perform in the way that you want them to perform or demonstrate the characteristics that you want them to generate or don’t generate the characteristics that you don’t want them to generate, so that you can identify the types of genetic modifications that you could make therapeutically that lead to the cells you’re taking out of people and putting back in to perform inside of the tumor in the way that you want them to perform. Am I thinking about that right?
Dipen Sangurdekar: That’s absolutely right. And in the case of you mentioned one application which is we want to potentiate the cells in our case, so that they have increased functionality — and whatever that functionality is, you can define it in multiple ways or whatever attributes you want to design into the system — and they go in and do a job better than if they were not edited. But the other way is you can actually just find targets. You can do those screens in cancer cell lines, which is something that we have also done, and find targets in cancer cells that under the right context kill those cells. So we have identified one such target, USP1, which comes into the class of synthetic lethality targets because in the right genetic context of homologous recombination deficiency, knocking out USP1 or inhibiting USP1 leads to cell death. Since that discovery, KSQ Therapeutics developed a first-in-class small molecule inhibitor called KSQ-4279 and now that program is being led by Roche further into clinical development. The discovery can downstream be taken into a variety of applications and that’s where the therapeutic drug discovery part comes in. And in the context of our gene edited TILs, they’re basically using the CRISPR modality itself where we design a therapeutic grade synthetic guide to create those gene edits in patient’s TILs themselves. Ex vivo means outside the body at a very high level of quality. Since this is a drug, we are held to a much higher standard of quality and when I say quality, in this case we want to make sure that we are editing where we intend to edit and we are not editing elsewhere in the genome and we are editing to a very high level. So it’s not like we are only achieving 5% of the cells being edited. We want to have a very high fraction of cells to be edited for it to be a therapeutic. So that’s where it comes into the CMC — this suddenly becomes a manufacturing thing that we have to pay a lot of attention to in terms of design of the therapeutic guide, controlling the parameters of doing the CRISPR-Cas intervention for each and every patient that we treat.
Ross Katz: Yeah, that makes sense. As I’m hearing you describe this process, what I’m really realizing is that we use this term personalized medicine but there are multiple levels to the personalization here. There’s the ways that the genetic modifications that you make will have dependent outcomes for the patient depending on the nature of the rest of their genome, the nature of their biology, the nature of the way that their immune system is responding to the cancer and the nature of the cancer itself, and then there’s also the personalized nature of the medicine because you’re doing an autologous therapy where you’re taking the patient’s own cells and then modifying their own cells and then sending them back to get them reinfused. Am I thinking about that right?
Dipen Sangurdekar: Yes. There are two things here that we have to think about. Not talking about our product specifically, but in general, whenever you go down that therapeutic discovery process, the edits that you find or the targets that you think are supported by the screen or they fall out from the screen, you have to repeat those experiments in a variety of different contexts. If you’re doing those in mouse models, you have to make sure that you’re testing this in mouse models with different genetic backgrounds. If you’re doing them in cancer cell lines, you want to make sure that you’re doing the screen in multiple cancer cell lines because you can’t just say that in a particular model, you see a target being prioritized, that target is necessarily going to be equally efficacious in other genetic backgrounds, whether it’s in human cell lines or in mouse models. Us and others who have done this ensure that this target is a valid target in a variety of those different contexts. And this is before we think about the translation from preclinical to clinical. Because in the real world, there are multiple effects that come in. Obviously everyone has a different genetic background and so this is where we want to make sure that our drug will be effective for everyone who has cancer and not just a particular subpopulation. They have their own medical history. Typically these therapies come in later in sequencing, so they have gone through multiple rounds of treatments which can affect the quality of their immune cells, but also their tumors are much more refractory. They have a lot of genomic enrichment done, their tumor microenvironment could be very immunosuppressive. So there are a lot of variables that come into play even in the clinic. And that’s why the translational hurdle from preclinical to clinic — one of the things that I really wish for is that using all the data analytics and AI, we could make that hurdle more of a step or a gap rather than the chasm that it is right now. Because if we can make predictions preclinically that will translate into the clinic, even if we improve our success rate by 10 to 20%, that will be a significant benefit. It will reduce development cost, it will reduce failure rates, it will benefit patients, it will be a win-win. But that’s where the gap is.
Ross Katz: Yeah, I was going to wait to go there, but let’s just go there now. What does that look like? How do you think about improving your ability to translate from the preclinical work that you’re doing in finding the target and validating the drug to the clinic, and then how do you use the data that’s coming out of the clinic to improve the drugs that you’re developing?
Dipen Sangurdekar: One thing I already touched on is making sure that any funnel that you’re constructing for target discovery includes looking at multiple model systems with different genetic backgrounds. So that still does not reconstitute human biology per se, but at least you’re assured that this is not just an artifact of looking in one cell line or one mouse model. You need to have multiple ones. The second piece which is definitely the harder thing to do is trying to reconstitute what the human physiology looks like. In the case of oncology, for example, trying to recapitulate tumor microenvironment in model systems — and people are making progress with that with organoid systems that try to mimic what the immune system, the microenvironment looks like — and there are multiple cells involved that are pro or anti-tumorigenic or pro and anti-inflammatory. We are making progress, but that’s the key. A model is by definition something that captures all the parameters that matter. And that’s going to be key. These systems are coming online, people have been using them to make predictions. Leveraging them, understanding them, generating single-cell data and multi-omics data and then characterizing what happens will go a long way in de-risking the translation across that gap.
Ross Katz: Yeah, it strikes me that the clinic is where your model of the world meets the real world. And so the clinical data is teaching you what the gaps are between the real world and your model. And so the feedback loop looks like identifying the questions you didn’t know to ask when you were building the model and the parameters you didn’t know when you were building the model so that you can iteratively refine that model based on the gaps in your understanding that led to unexpected effects once the therapeutic got in the world.
Dipen Sangurdekar: Yeah, and I think you’ve hit on a very important point. We don’t know what to build into the preclinical models if we don’t study what the clinical data is telling us. And for that, we need to have an expansive approach. Obviously within the budgetary constraints of running a clinical study, but we need to understand at a very detailed level once our drug goes into a patient — and remember, patients are really doing us a huge favor by participating in those studies because honestly a lot of the time these studies fail. But what we learn and what we owe to the patients is leveraging all the data that we can so that in the future we have a better shot on goal. I’ve been part of programs, not just here at my current company but previously as well, where we have derived not just the clinical data captured in CRFs and clinical databases, but also patient sample data, with consent, to do exploratory research. We often extract tissue — cancer tissue data from biopsies, serum or blood, radiological imaging data or other imaging data. These are non-invasive techniques that are part of the standard of care, so it’s not like we are asking patients to do anything out of the ordinary. But as part of the standard of care these samples are collected. We then generate multimodal data sets from those and really that’s when we start to learn. Even if a drug did not work in the sense that it did not produce a top-line benefit, can we see something in the PKPD data? Can we see some pharmacodynamics? Can we see some target engagement? And then can we build machine learning models that include the other data that we’re collecting that explains that outcome. Because obviously correlation is not causation, but when we know what we know about the mechanism of action, what we know about prior biology, we can really come up with new hypotheses that we take back across the hurdle to the bench, build them into a model system, tweak and do the loop as you mentioned.
Ross Katz: I love where we left off because it brings me back to where I was looking to go earlier. If we can zoom out a little bit, that’s an excellent example of where the data science team that you oversee engages in connecting what you’re seeing in the clinic back into what you’re doing to develop the therapeutics in the first place. Can you just walk us through broadly how the data science team — the combination of computational people, bioinformatics people and machine learning or AI people — work together to support the creation and then the optimization of the cell-based therapies that you develop?
Dipen Sangurdekar: If you look at it from a very high, 30,000-foot level, I think the mental model that most people use is still very valid. You have a pipeline with an early stage where you have target identification, target validation. Once a target goes through the stage gate of becoming a development candidate, the next part of the pipeline is IND-enabling studies — IND is investigational new drug, which is basically an application to the agency or FDA to say that we want to go into the clinic with this drug and here’s all the safety and efficacy data and some other background about our drug. Those studies need to be done, they need to be submitted. And then once a drug has moved into the clinic we have the Phase 1, Phase 2, Phase 3 clinical studies. That linear model is quite representative of how we think about it. We have folks — and I’m generalizing what our team does but also my previous teams and other teams that I’ve observed — who are focused on particular stages of the pipeline. For example someone is very focused on discovery, which is target ID, and often that’s associated with some form of platform development whether it’s an AI platform or a screening platform. Others who either travel with the project down that pipeline or focus on translational work are really skilled in understanding not just preclinical biology but also the data science aspects of what clinical data looks like, what are the different formats, what are the constraints and quality issues and very importantly how do you treat clinical data with care. Obviously apart from HIPAA concerns, we have to make sure that the data is handled with appropriate integrity, security and governance. Folks who are well-versed in that can operate at that interface very effectively. I wouldn’t call it background because I think having a learning mindset is more important. People who are willing to learn about preclinical biology, what do model systems look like, what does the data really mean, and also what the clinical data looks like, what are the different terminologies, how do clinical sciences and development and translational research think about those problems. Folks who are willing to learn those two and bridge or make that bridge and use data as that connective tissue or data systems, those will thrive at that interface. And then there is obviously an essential part of this that I want to make sure I’m covering, which is the data engineering and the technologies aspect of it. We want to make sure that across the pipeline we have the right systems, the right technology stacks to make sure that the data is ingested, it is curated, it is inventoried, it is tagged and it is housed in a way such that people who need to access the data have immediate access to it. Whether you’re using a scientific data management system or a LIMS system, whether you build bespoke databases, whether you’re using cloud technologies and AWS services to access it. We need to have the strategy in place and there are folks who own that piece as well. We consider that platform work, foundational work which supports everything but no less important than any particular project.
Ross Katz: I think you did. You’ve given it to us at a high level. Are you able to share any examples of what a project looks like in the target ID or platform development part of it or in the translational work — what data you bring together, what methodologies you apply, that sort of thing?
Dipen Sangurdekar: Maybe I’ll give a couple of examples. In one of my previous roles when I led a computational biology team, one example was working on a particular hypothesis — the hypothesis was we want to go in and find targets for a particular target cell type. And that target can translate if we have sufficient hypothesis about the target, then we can modulate it and modify the behavior of the cell type. We didn’t really have any internal data on that cell type at the single-cell level so we ramped up a campaign of sourcing tissues and doing data generation on the cell types whether it’s IHC data, single-cell RNA-seq data and so on. Members of my team focused on bringing in large public data sets. This includes but is not limited to TCGA data and other single-cell data sets that were published. We built a large catalog of single-cell data sets that existed out there and had brought in-house. We worked with a couple of companies in the space that did data curation and inventorying for us. Once we have the data, you have to ask a very specific question. What are the targets that are expressed in that particular cell type that are unique to that cell type and are not expressed elsewhere? For there to be a therapeutics index, that target has to be expressed preferably on the surface in that target cell type but not elsewhere. And that leads you down a bioins- that’s a very linear concept to follow and work through. Once we have a list of targets, you bring in other data modalities like IHC data, literature, the huge amount of knowledge that scientists themselves carry from their training and then it all comes together. The other example I would give is the screening data that we do in-house, which is exclusively in-house data that is our bread and butter in the sense that we generate large screens. We use libraries that we have synthesized to knock out genes at scale but one gene per cell. We do these enrichment studies where by sequencing methods, at a particular endpoint, we are seeing which of those sgRNAs are enriched or depleted. That’s a very linear concept but it has a lot of nuance. Someone who is leading that analysis project would have a very deep understanding of not just what the experimental design is but also informing the experimental design by saying what’s the number of replicates do we need, what kind of sequencing depth do we need, what’s the power do we need, and do a lot of simulations to set up the experimental parameters. But then also have the bioinformatics skill set to do the next-generation sequencing analysis, do a very high level of quality control and then finally at the end of that pipeline when the data comes out come up with the statistical methods, rigorous methods to prioritize and come up with the volcano plots and the ranked list that then go into downstream interpretation. Depending on the project, we could use internal data or we could use external data. Even for external data it’s not like it’s clean. External data comes with a lot of metadata and anyone who has downloaded data from GEO before knows there’s a lot of work that needs to go in to curate, to clean up data, to do normalization, to merge the data together — it’s not just that we start with a clean table and apply statistical models. A lot of thought goes into getting that data into a shape that can be analyzed downstream.
Ross Katz: What I’m hearing is that it’s highly hypothesis-driven and that the data sets that you bring together and the methodology that you apply are very much dependent on the hypothesis that you’re chasing down. How do you keep track of all the hypotheses that you could be chasing down and prioritize the work of the data science organization to make sure that you’re driving the insights that the organization needs to move along the most strategically important paths?
Dipen Sangurdekar: I’ve been fortunate enough to work in organizations or departments where data science teams or computational biology teams or bioinformatics teams that I’ve led have been highly integrated and embedded with the biology teams whether they’re discovery or translational research. To answer your question, we don’t do this in isolation. Every member of the team is highly integrated into project teams and the project teams are the ones that come up with the questions, the hypotheses. And then if you’re a data science member part of that team, your job is not just to react and do the analysis, but to bring your own perspective. Within the data sciences team or bioinformatics team what we try to do is make sure that everyone knows what other people are working on in their projects because that’s where a lot of the cross-pollination of data-driven hypotheses happen — an individual member can go back to their project team and say, someone else is working on their project and this is the idea that they applied or hypothesis they generated, maybe we could try a variant of that here. It’s important not just to be reactive to the questions that are generated by the biology team but also to bring unique ideas, innovative ideas and hypotheses. That’s how you grow as a scientist, as an expert, and that’s sort of brings the rises all tides or — rises all tides I would say — across the board. Bringing your expertise also means raising your hand if you believe that the questions that are being asked are not the right questions, or if the hypotheses being selected to work on are maybe not the most important hypotheses. But that’s something that needs to be decided as an entire project team for moving the project forward rather than something that you unilaterally do in the data science org.
Ross Katz: I’m interested in how working on cell therapies — these living drugs — makes these machine learning-oriented or statistical analysis-oriented data projects hard. What are the biggest challenges that you face in that environment?
Dipen Sangurdekar: I’ve been working in the cell therapy space, the gene and cell therapy space for a few years now and one of the reasons I’m excited and I feel like as a data science practitioner I have something to contribute here more than other modalities is because as you said, our drug product is highly personalized and it’s a living drug. And what that means is that unlike a small molecule or biologic, there are a few axes or dimensions in which you’re optimizing the drug and defining a target product profile — whether it’s chemical matter, chemical properties, ADME properties and so on. In a living drug, there could be many different dimensions and we may not even know what those dimensions should be or are when we start the study. As an example, in the field of CAR-T, it’s been well-established that you need healthy cells, proliferating cells that also have the capability of going to the target and doing the effector function and then living for a long time in the system. So initial efficacy but also duration and persistence of those cells is important. If you just double-click on those dimensions, that could mean many different things at a biological level. And the only way to interrogate that is to do very deep profiling of our drug products before they go into the patients, and if you can find them once they’re in the patients — extracting them either from tumor biopsies or from the blood, if you have ways to pull them out and do characterization there — that’s important. Because then you know what are the properties that you built into the system and what does the body do to the cells. So you have a plan and then when the plan meets the real world, what happens to that plan? Understanding both is important. And the only way we can do that is very deep molecular profiling using multimodal analysis, single-cell RNA-seq, in our case TCR sequencing, but also proteomics, CyTOF, etc. Now you have a large data cube, a very complex cube. Unfortunately, when we talk about AI we talk about scaling laws — we’re not talking about a lot of scale. We are talking about a lot of depth but not a lot of samples, just because of the space that we are in. So then you have to be very creative in asking those questions where you can’t just throw machine learning models at it or build foundational models and say that patterns will emerge. You have to ask some very guided questions because we don’t have enough data to escape the multiple testing problem. If you just do a fishing experiment you’ll get thousands of P-values and some of them will be significant just by chance. But if you go in with guided questions that are driven by prior hypothesis, you might get a signal that is weak but suggestive of a trend, backed by a very strong hypothesis, that gives you a lot more confidence that maybe this is real and now let’s go back and test this preclinically or nonclinically. That’s how that cycle works. Sometimes over time you build a large enough data set and then you can do exploratory data analysis. You can then see if broader patterns emerge that you were not looking for and that correlate with a particular trait or a downstream outcome. That can happen, but in the early days when the data is just evolving, you’re nowhere close to that so you have to be smarter about what kind of questions you ask.
Ross Katz: Rather than letting the data tell you the story, you go in with the hypotheses that you believe might explain what you’re seeing — for example, in the difference between the cells that you’ve edited and deeply profiled before they go into the patient and the ones that you’ve deeply profiled after they go into the patient — you use your hypothesis to shine a light between them and then see whether there’s any degree of signal there. If there’s enough signal to support the hypothesis even if it’s not rejecting the null hypothesis, you can still move on to some sort of experimental validation that is also a more targeted approach and a more rigorous approach but obviously more resource intensive.
Dipen Sangurdekar: That’s absolutely the right thing to do. There are statistical purists out there who might say this is not — you are doing things — I think experimental validation, the ability to go in and to quickly validate something frees us from these constraints of getting a very strong signal or a very low P-value and adjusting for multiple testing and so on and so forth. Equally important is that hypothesis that you have. I don’t want to say that the data can’t tell us a story. Even with a low amount of data if you do pattern recognition, EDA, you might see a trend. But then the key part is — does that trend go against the grain? What’s the level of surprise associated with that trend and, more importantly, somewhere in the data is there another trend or another weak trend that backs up this trend but is sufficiently orthogonal. So if you’re going through dozens and dozens of box plots and you see one box plot out of all of them — do you believe that? I wouldn’t. But if you see another association that pops up, which is weak, completely orthogonal, preferably from two different assays so that the noise is not correlated, but it points to the same biology, now that’s something that I would believe — maybe the data is telling us something that we should follow up on. It can work both ways even in the early stages, but you have to be very careful what you’re doing.
Ross Katz: And what you were saying earlier about having a sense of what your prior expectations are helps to guide you for knowing when you’re going to let the weight of the evidence push you in a given direction — and not making any major decisions in the absence of experimental validation is a critical check on whatever logical fallacies you might fall into in interpreting the data.
Dipen Sangurdekar: Right. It’s very traditional Bayesian thinking. If you have a prior then your data should tell you that the evidence to overcome those priors has to be very strong. And you have to think critically about the evidence. Obviously if you have a very weak prior or something that you have not thought about, then you have to follow the data a little bit more. And the other challenge of what I just described is more of the infrastructure piece, getting that data from the different vendors and from the different assays and from the different domains and connecting it to a unified data backbone. That unified data model or common data model or backbone is very critical. Every study, we have to build up that data architecture, the way you’re going to integrate the data, the way you’re going to connect the data. And I’m a big believer in doing that connection and then sharing the data with the clinical study teams or the translational teams so that they can go in and do their own investigations rather than us being the gatekeepers of all analysis. That’s something that we’ve done very effectively in the past and we continue to do so — we are all in this together. Our specialization is perhaps to ask some of the more complicated analysis questions like multivariate analysis, machine learning, but everybody who has access to a data dashboard that we built for them should feel enabled to go in and ask those questions or test a hypothesis quickly themselves without coming to us for a particular plot.
Ross Katz: Awesome. As we come toward the end here, I just want to end it with a couple of quicker questions. The field of data science and the field of biology are both moving very quickly. What are the innovations that you’re most excited about in your work in the next year or three years?
Dipen Sangurdekar: The evolution of foundational models is something that is absolutely going to change the way we do some research in the future. We are leveraging some protein language models to embed our high-dimensional data. Maybe not generate new modalities per se because we are not in that space, but just developing those encodings and doing data integration using those models is something that very exciting. The productivity enhancement from using coding tools like CoPilot and others and integrating them in your systems such that you are protecting your data privacy but also enabling the gains is something that I’m very excited about. But the biggest thing is on the biology. Using these tools, people in the research community and even in the non-profit sector are on the frontiers of biology. Just keeping up with all the papers that are coming out in the space, what we even know about the immune system, what we know about single-cell biology, how tumors escape IO therapy — the pace of that is just amazing. And keeping up with that is one of my keen interests and something that I try to do. So technologically and scientifically, it’s all going in the right direction. It’s very exciting.
Ross Katz: That’s all very interesting. And then, obviously having come up in this field and running multiple data science teams, I’m just interested in what advice would you have for people who want to contribute to therapeutic design and development using data-driven approaches?
Dipen Sangurdekar: I look to my own path as well and other folks who have been data science leaders that I’ve worked with or I’ve observed. Data is ultimately — the machine learning, everything — a tool. Just like doing certain experiments is a tool. What you do with the tools is more important than the tool itself, at least in our area of work. So I would say take a problem that you believe in, take a problem that interests you, and just go deep, understand the biology, understand whether it’s clinical sciences, whether it’s basic biology, whether it’s something else, go deep with that and then think about how you can leverage those tools that you have at your disposal in solving that problem. That might sound scary for folks wondering what do I pick? One thing I would say is don’t worry about it. Pick one problem for now, spend a couple of years, you might go deeper, you might say this is not for you or you feel like I’ve done this now I want to do something new. Options are always open if you’re very good at what you do, those doors will always be open. But just pick something that excites you, that interests you and become an expert, become differentiated and then you’ll be just fine.
Ross Katz: Awesome. And where can people go to learn more about you and your work?
Dipen Sangurdekar: Our company’s website is ksqtx.com and I can be found on LinkedIn.
Ross Katz: Awesome. Well, Dipen, it’s been a pleasure to have you on. Really appreciate the time. I look forward to connecting down the line.
Dipen Sangurdekar: Thanks for the invitation, Ross, and thanks for the conversation as well.
Jason: And that’s it for this episode of Data in Biotech. If you enjoyed the episode, please subscribe, rate, or leave a review in your podcast platform of choice. See you next time.
