Listen on
Overview
Drug development faces immense costs, with each new drug costing between $350 million and $2.5 billion, coupled with a high failure rate. This financial burden and slow pace directly impact market competitiveness and patient access to new therapies. This episode reveals how applying AI and machine learning across discovery, manufacturing, and clinical development can reduce these costs by 25% to 50% and significantly accelerate timelines.
Mitra Mosharraf, Chief Scientific Officer at HTD Biosystems, brings two decades of pharmaceutical industry experience—including a PhD in pharmaceutical sciences and an MBA from MIT—to explain how AI is transforming nanomedicine. Her insights span early-stage molecular discovery to optimizing complex manufacturing processes and improving clinical trial efficiency. She offers a unique perspective on the intersection of deep scientific knowledge and practical business application.
Ross Katz and Mitra discuss concrete applications, such as using AI to identify novel drug candidates, dramatically speed up formulation development, and enhance patient recruitment for clinical trials. The conversation also addresses the critical need for high-quality data and rigorous model validation to mitigate risks and ensure that AI’s promise translates into tangible, safe, and cost-effective outcomes in nanomedicine.
Key Takeaways
AI/ML can reduce drug development costs by 25-50% and accelerate market entry for nanomedicines.
The substantial investment in new drugs can be significantly cut. Specific applications, such as HTD Biosystems’ iFormulate platform, compress formulation development from six months to just one week by intelligently optimizing critical parameters. In clinical trials, AI-driven site selection has proven capable of halving patient recruitment time, turning an 18-month process into nine months and bringing therapies to patients faster.
Rigorous data quality and model validation are paramount for effective AI applications in drug development.
The true benefit of AI in nanomedicine hinges on the integrity of its training data and the validity of its models. Misapplied models or those trained on insufficient, biased data can lead to costly false predictions and undermine the value AI promises. Leaders must prioritize rigorous data provenance, diverse training sets, and strict validation protocols across all development stages to prevent significant financial setbacks.
AI optimizes core operational processes in manufacturing and clinical trials, not just discovery.
Beyond identifying new drug candidates, AI improves the efficiency of established processes. It helps pinpoint manufacturing bottlenecks, refines process parameters to ensure consistent drug quality batch-to-batch, and allocates resources more effectively. In clinical operations, AI speeds patient recruitment by targeting disease hotspots and automates data review, shortening data analysis timelines from a month to hours.
Regulatory frameworks for AI/ML in drug development are evolving, requiring industry collaboration for effective standards.
The FDA is actively updating its guidance for AI and ML, acknowledging the rapid pace of technological change. While progress is being made, clearer, standardized protocols for data diversity, model validation, and specific application contexts are needed across discovery, preclinical, and clinical phases. Collaboration between data scientists, regulators, and industry—particularly through data-sharing consortiums—will be essential to shape effective standards that balance innovation with patient safety.
Related: CorrDyn helps biotech firms realize data value through data engineering and AI strategy. We also specialize in data cost optimization to maximize ROI.
Full Transcript
Mitra Mosharraf: It cost $350 million to $2.5 billion to develop drugs because if we utilize AI and ML across each stage, so in discovery, pre-clinical development, and clinical development…
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. Here we go.
Ross Katz: Mitra Mosharraf, welcome to the Data in Biotech podcast.
Mitra Mosharraf: Thank you so much.
Ross Katz: Awesome. Well, just to kick us off, would you mind giving us an introduction to you and your background and what brings you to AAPS PharmSci 360?
Mitra Mosharraf: Sure, I’ll be happy to. My name is Mitra Mosharraf and I work in drug development. I have more than 20 years of experience from the industry. I did my PhD in pharmaceutical sciences in Uppsala University in Sweden. I also did my MBA at MIT. I worked with nanomedicine drug development and also protein formulation development and development of biologics. I worked for companies like Pharmacia and Pfizer in Sweden. Now I have my company, HTD Biosystems, where we work with high throughput drug development. We are a CTMO in San Francisco Bay Area. Also co-founded a company called Engimata for development of lipid-based nano delivery systems for vaccines and immunotherapy. I’m happy to be talking to you today.
Ross Katz: Awesome. Well, you have an extensive background, so very exciting to talk to you. One of the things that you’ve published research on is all the different applications of AI and ML to nanomedicines. But before we dive into applications of AI and ML, I would love to just hear from you for the audience who aren’t as closely affiliated with nanomedicines as you are, could you just give an outline of what nanomedicines are and how they’re used in the pharmaceutical space?
Mitra Mosharraf: Yeah, absolutely. Nanomedicines are drugs that are in the size range of one nanometer to 100 nanometers, usually. That’s at least how Britannica has them defined. FDA and ISO, which is standardization organization, has defined them as any drug delivery system-based medicine that uses a component that is in that size range, one to 100 nanometers. So they’re very, very small. They’re very exciting because they can interact with cells in a specific way which larger components cannot do. So that’s why we’re very interested. The smaller the particles, they have better retention time in the tissue and they can interact with our immune cells in a different way. In terms of what we see with nano lipid particles, which are a type of nanomedicine, they can transfect cells, like when we put the payload inside them, like the COVID-19 vaccine by Pfizer and Moderna, we saw the mRNA that was delivered to immune cells that way, by using lipid nanoparticles because they could transfect and deliver.
Ross Katz: Right. So we’ve got this relatively fast-emerging mode of delivering medicines to the body. It’s already very big as a result of the COVID-19 vaccine, but also other therapeutic approaches. Can you maybe outline why nanomedicine is so fast-growing or where the fast-growing areas are coming from?
Mitra Mosharraf: Yeah, I think it’s very important to see that in connection to other revolutions that are ongoing, like we’ve seen in gene editing and gene therapy, and also AI and ML. These are all reasons that we can develop nanomedicines faster now and come up with nanomedicines that don’t have the hurdles to get to market that we had previously because we can now develop safer and less toxic nanomedicines. With the lipid nanoparticles that are using ionizable lipids, that are less toxic, we can see that now with mRNA-based vaccines for COVID, and also before that, the siRNA product that Alnylam called Onpattro. That was another example that came actually out in 2018, the first mRNA-based vaccine. We need a delivery system for bringing in products that use gene editing as a tool is very important, especially with the viral delivery system, which is for AAVs, which also we saw in J&J and AstraZeneca’s vaccines. We saw how that could be a useful gene delivery using viral delivery system, but I think the US government recently gave a directive that it’s good to focus on non-viral delivery systems also, and do a lot of research for that because they have less immunogenicity and you can give them multiple times rather than just once.
Ross Katz: Yeah, it sounds like toxicity and the safety were some of the old challenges with nanomedicine, are there other challenges that you would point to that have stood in the way of the growth of nanomedicines over time that—and I’m assuming that AI and ML are helping to overcome some of those challenges in terms of enabling nanomedicines?
Mitra Mosharraf: Yes, so with liposomes and positively charged cationic lipids, that was one of the problems, right? Because they were more toxic, since they could interact with immune cells in a different way, and we were afraid of cytokine storm. Using now AI and ML, we have been able to detect better ionizable lipids that can also for cancer therapy be used to target the nanolipid particles to other organs than lymph nodes, where they normally might go to, or if they are given IV, they might go to liver also. A lot of research using AI and ML is being done to detect ionizable lipids that can direct the lipid nanoparticle to other organs, which is very, very important. I think Agile’s platform is one of the platforms that I can mention that researchers are using to be able to do this kind of predictions.
Ross Katz: Yeah, I think that would be great. You have this pharmaceutical research paper that outlines all the different applications of AI and ML to help bring nanomedicine to market and also ensure that they’re successful. I’m wondering if you could walk us through the different stages where AI and ML are being applied in nanomedicine, and then we’ll dive in a little bit deeper to each stage afterwards.
Mitra Mosharraf: It cost $350 million to $2.5 billion to develop drugs. You can just see how AI and ML can help us save costs because there is a report by Wellcome Trust and BCG that came out, I think in 2023, which points out that we can save 25% to 50% of that cost if we utilize AI and ML across each stage, so in discovery, pre-clinical development, and clinical development.
Ross Katz: Yeah, that makes a lot of sense. AI and ML are being used for the purposes of de-risking, of reducing the cost, of allowing these drugs to come to market faster and with lower resources utilized. If it’s okay with you, I’d love to talk about a little bit of the how underneath, how AI and ML is being applied at each of these phases. Starting with discovery and early development in nanomedicine, what are some of the use cases or some of the methods that are being applied to reduce the cost or reduce the time that it takes to get drugs through that phase?
Mitra Mosharraf: Yeah, there are many companies, especially I know that specifically large pharma, they are actually establishing their own data libraries, to use as source of data to be able to evaluate all their historical data and also create experimental data to use in building predictive models to be able to discover novel molecules that can target different disease strategies, for cancer, for example, if they want to identify which molecule would bind to the receptors on the surface of cancer cells or for the delivery systems, which lipids as I mentioned earlier is more useful or which polymer for polymeric nanoparticles is more useful for targeting. They look at the interaction between the molecules and cells from the molecular fingerprints that they identify first and also for the delivery systems from the delivery particle’s fingerprints, which is what for example, size, particle size, particle charge and PDI, which is polydispersity index, because we want to make sure the critical quality attributes are identified and discovered and put them in that discovery model so that we can see there what we need. How changes in size might change which cells they’re interacting with and how changes in a composition can impact where they go in the body, how should we give them, should we give it intravenously or intramuscular? Oral is more difficult now for biologics, but there is a lot of research being done in that field as well.
Ross Katz: Right. On the discovery side, what makes it the determination of whether a nanomedicine is needed is dependent on where, when, and how the drug is going to reach the particular tissues and the different molecules that it’s going to interact with. If I understand you correctly, AI and ML is helping to characterize where those interactions are going to occur, how strongly they’re going to occur. Am I thinking about that right?
Mitra Mosharraf: Yeah, I mean, it speeds up that and also helps us from what we already know to detect, for example, which type of molecule might have better success rate and also for repurposing of drugs. Sometimes you might have a lot of, if we have a lot of historical data for drugs that already exist, if we want to repurpose them. It can scan what is needed for a disease strategy and then see, okay, what drugs are already available and how we can use them to address another disease now. That’s the drug repurposing. With AlphaFold, for example, the designers of AlphaFold got the Nobel Prize, right, a few years ago? That is a good example of how they developed this ML model that could predict protein structures from all the proteins that existed on the planet and could come up with novel protein structures now and look at the interaction. That has been a great example of how we can use AI and ML in discovery.
Ross Katz: Yeah, I’m interested in the nanomedicine field, are there any examples of places where you can point to AI and ML really speeding up the discovery process, or—I know that a lot of these companies are utilizing AI and ML to, as you put it, narrow the universe of candidates that they’re selecting from and also test them, they’re testing them in silico, but I don’t know that I have any examples to hand. Would love to hear if you have any.
Mitra Mosharraf: Yes, an example that comes to my mind is the Agile platform, which is a platform that has been developed by one of the companies, to find ionizable lipids that are novel, so it’s an ML-based platform and it speeds up that in discovery. We can use that, and the other example is for mRNA, I think, finding which mRNA would be more effective in a genome, to be able to find, or gene sequencing, that comes to my mind.
Ross Katz: Do you have a sense, especially in the discovery domain, of meta-principles that people are applying that have them to ensure that their models are less biased or that they’re covering more of the space that they need to cover in order to effectively characterize the phenomenon that they’re trying to characterize?
Mitra Mosharraf: I don’t know if they cover more space than they should, but I think they are covering basically what they want to look at. They cannot just apply to another, if they look at a specific molecule, they cannot just say that would also apply to another molecule. That’s very important to have diverse and large amount of data to prevent bias, and not to just apply your model to any other model. If you look at proteins, we cannot just apply that, can we apply that to antibody drug conjugates? Yes, to certain extent, but not generalize it, right? Because then we need to validate the model again for, even if there are small changes, so that we have been careful with that validation part before we draw a lot of conclusions, otherwise we might have the wrong predictions and that would cost a lot of money.
Ross Katz: That makes a lot of sense, and my understanding is that’s one of the limitations of AlphaFold 2 and to a certain extent AlphaFold 3 is its ability to characterize things like ADCs or any proteins that have multiple proteins interacting simultaneously. Because that’s not necessarily in the protein database that was used to predict it. Understanding the provenance of the data is really important for understanding how well the model’s going to do it predicting it.
Mitra Mosharraf: Exactly, you’re absolutely right. There is a lot of resistance, I must tell you. I was just presenting about AI and ML in a conference and a lot of people were resisting and they were objecting. I felt for a second I was a representative for ChatGPT there because they were saying ChatGPT lied and there were so many things they were talking about that they were thinking AI is misleading them sometimes with the outcome. I thought, but you have to be attention to the model. How is it trained and how did you validate it? You have to check, to turn off your data, you have to use for training, and then what is the quality of that data that you use in the training? That’s so important. That’s the problem we saw with COVID-19 in the beginning for clinical trials that the data maybe was not as diverse and did not reflect the minorities because, now I jump from discovery to clinical trials, but I thought that’s good to bring up that example because, for example, they did not include minorities in the design initially because the historical data that was used maybe did not reflect that part of the population as well and then in practice they noticed that their simulations were missing that part and then they adjusted it. That shows the importance of diversity and to include minorities so that we don’t have the overfitting problem.
Ross Katz: Yeah, for sure. It’s interesting to hear the story of you being this advocate for AI and ML and having pushback from the academic ecosystem about—but that there’s this nuance there of, all models lie, it’s just a matter of understanding where they’re likely to lie to you and based on how the models are constructed and your ability to interact effectively with the model is directly proportional to your ability to understand the limitations of the model that you’re working with based on the data that’s used to train it.
Mitra Mosharraf: Yeah, absolutely.
Ross Katz: We’ve spent a lot of time in discovery. I would love to go on to manufacturing and development. Can you give me some sense of where in the manufacturing and development side, AI and ML is being applied in nanomedicines?
Mitra Mosharraf: AI and ML in manufacturing are able to screen this large amount of data that is produced and find patterns that we couldn’t maybe find ourselves with human eye, basically, and much faster. And help us with optimization of process parameters, which are so important because those impact the drug quality. Always we have to make sure that the drug quality is maintained throughout different stages of manufacturing.
Ross Katz: Yeah, that makes a lot of sense, and this is actually related to some of the work that we’ve done as well, implementing systems that capture all of the sensor data that’s coming off of the machines and then use that to predict the quality attributes that are measured after the fact. Are there any organizations or any methodologies or any examples that you’ve found where companies are doing a really good job of this and some of the attributes that are coming off of the machines that are useful for this kind of prediction?
Mitra Mosharraf: At HTD Biosystems, we have developed iFormulate, which is for formulation development. It’s a platform that we have, using DOE, to predict and optimize the lead formulation for development of biologics and nanomedicines. Also, we can speed up formulation development from six months to one week, so it’s very good and we have seen a lot of good traction from the industry. There are also others, I think Formulation AI is one of these platforms that can be used to predict the best formulation.
Ross Katz: I would love to hear more about, what are some of the parameters that are in the optimization problem in DOE, how that goes? How do you reduce defining a formulation like that from months to a single week?
Mitra Mosharraf: Yeah, we have identified a few variables that are important, like pH, ionic strength, and the concentration and type of sugar and buffer, we have designed a DOE design that changes these variables in formulations and then we measure the response by our high throughput technologies for protein formulation and those are like looking at denaturation temperature and the biophysical stability, like aggregation. Then, we do rank them. When we do that, then we have 25 trials, only, and a few of them are replicates, and then we rank them and we look at the surface response, quadratic surface response, and based on that, we understand the design space and where in that design space we can find the optimized formulation. Then we select three lead formulations based on that, and we do accelerated stability study and stress the molecule a bit more. Based on those data then, we select a lead that we put on longer stability. Yeah, so it has really helped us help our clients a lot with accelerating their drug formulation development from months to weeks. They use our data because we create all these very interesting graphs and DOE results for them in their IND application. FDA usually loves that companies look at the design space because then you can find the edge of failure and also know how if you’re a little bit shifting in your specifications, how that in that design space might impact the quality of your product.
Ross Katz: Yeah, that’s really interesting. A company comes to you with a new compound that they believe is appropriate for nanomedicines, and you’re doing this design space analysis, doing the entire DOE that looks at all of these different variables and characterizes the terrain of the factors that are most important things like the effectiveness of the drug and then also the stability of the drug. I’m assuming some other things as well, solubility or something along those lines. In doing that, you can help them to select the optimal formulation that gives them those parameters for their drug, but then also you get to draw a circle around that and say, if you go down this area in the design space, then you’re going to fail from an effectiveness perspective. If you go down this way, then you’re going to fail from a solubility perspective. If you go down this way, you’re going to fail from a toxicity perspective or whatever. Am I thinking about that right?
Mitra Mosharraf: Yeah, exactly. You’re thinking right about that. This is very important to do in drug development, especially, I think nowadays it’s even applied in manufacturing in lyophilization development. You can just change the process variables to find the optimized and robust models that can be applied across this.
Ross Katz: Yeah, it is. Is there anything else on the manufacturing side that you would talk about from an AI or ML in nanomedicines perspective?
Mitra Mosharraf: Yeah, I think AI and ML also can find bottlenecks in the operation and help us allocate resources and that is very important from cost and time perspective and how much resources you need to allocate to the process, manufacturing process.
Ross Katz: In terms of finding bottlenecks and resources, should I think of that as understanding where time is being spent in the process or where resources are being invested in the process or, how do you set up the problem in terms of optimizing resource allocation?
Mitra Mosharraf: Yeah, well, you have to of course, we have an engineering run first, right, before even doing any Phase I or any type of manufacturing. During that engineering run, we look at different process parameters and we look at how many FTEs we might need in different stages, what resources we need to have. We take time for how long it took. Did we assign the right time? Did we assign the—and then, if we are at the same time having ML and AI look into this to see where we could have improved processes, I think that would be very helpful for us, and that’s what we do. We look at the bottlenecks are where the time was too long, right, and slowed down the process. Maybe there were hold times. Maybe something happened during manufacturing. Then you have to see when you look at batch to batch consistency, how we can improve batch to batch consistency also, right? It’s not just based on one run, but engineering run was an example for one run, but you also need to look at the historical data to help you with that.
Ross Katz: One of the things that we’ve done before is create simulations of the manufacturing process in order to understand which process steps are blocking the throughput. Is that what you’re describing?
Mitra Mosharraf: Yes, for that, actually, I have done those simulations also when I was doing my executive MBA at MIT. We did a simulation for operation management, which was very, very useful to understand how many machines we need to have at each step depending on the workflow that is coming. The more workflow would, because sometimes the workflow is more, sometimes it’s less. Now this, in this example. How to understand how many FTEs you need to assign to each stage of that workflow. This can also be applied in addition to manufacturing to hospital operation management, as you probably know. In any kind of operation management, you have to see this different stages, how much equipment are you utilizing and are they using all the time, are they being used? The bottom line is that 68% efficiency, they have to be used 60% of time, at least, but not 100%, so that they are not overwhelmed. They’re always used so that you can control the workflow. When it’s more or when it’s less, that’s the best approach. The same with clinical supply management.
Ross Katz: Yeah, interesting. Since we’ve stepped into the clinical domain, would you mind talking a little bit about applications of AI and ML in the clinical trials or the clinical development phase, and how the different applications, in particular for nanomedicines?
Mitra Mosharraf: Yes, I’d be happy to. For clinical trials, I think applications have helped a lot by patient recruitment, identification of clinical sites. I think during the COVID-19 trials, I heard that they could really find the disease hotspots and based on those, understand where there is a higher probability of success for patient recruitment. Some companies that didn’t do that, they had much more difficulty finding and recruiting patients. I worked for a company once that couldn’t find enough patients because they had, before pandemic, because at that time nobody used AI and ML. People used to go with historically which sites they had used, which doctors they knew, and surveys that they would do with clinical sites. I heard that later, for example, a company called Amgen, used Atomic, a clinical site selection software to be able to speed up patient recruitment, from 18 months to half. It’s sped up that much. There is more than that, if you want me to expand on it.
Ross Katz: Please.
Mitra Mosharraf: Yeah, for clinical trial design, it’s been used and for simulations of clinical trials to see the outcome. We discussed a little bit about this before. In general, for data review and data management during COVID trials, I read somewhere that Pfizer used data query software which saved data review time, and a process that would have taken one month could be done in 22 hours. I think I read it on their website.
Ross Katz: That’s reviewing the data that’s coming out of the clinical trials for accuracy, to make sure that the trial is progressing the way it needs to be progressing.
Mitra Mosharraf: Definitely. Yes. And data processing for bioanalysis, yeah, and the data review of the results.
Ross Katz: Yeah, that makes a lot of sense. In terms of recruitment, my basic understanding of the recruitment technology is that you’ve got all of these electronic health records out there. You’ve got data about the patient population that you can understand about people who are likely to have particular diseases or problems that can be addressed by a clinical trial. You’re searching the patient population for where patients are likely to be clustered and then recruiting sites in that area. Am I thinking about that right?
Mitra Mosharraf: Yes, exactly. You’re exactly thinking right. Also, ClinicalTrials.gov is a very good resource to be able to see what historical data exists, the health records, patient health records, if we have access to them, a lot of hospitals have access, but you have to think also about ethical issues and make sure you have the patient’s approval to use those, and also the wearables that they use for like the Apple Watch and other wearables devices have helped us a lot with gathering data from patients, live data, and be able to use those data in simulation.
Ross Katz: What are some of the exciting applications of wearables that you’ve seen for nanomedicines? I’m assuming inside of clinical trials, what kind of information are you getting and then how is the industry able to use that?
Mitra Mosharraf: For wearable devices, I think it’s more for compliance, to see if patients are really using, for example with diabetes, now that’s not exactly an example of nanomedicine maybe, but the doctors can follow very well and see how the drug intake is impacting the health of the patients. If the patient’s not really taking the drug, you can also see that if you review those data. In general, I know that hospitals are really using this to evaluate patient’s health and for clinical trials also for compliance I think it’s been used a lot. Also monitoring what are the biomarkers, is the heart rate goes up, or vitals of the patients, they can see all those together at the same time and this is very important and then being able to use those data in later modeling.
Ross Katz: The entire data ecosystem is very interesting from the perspective of, we’re always measuring proxies or proxies of proxies for something like compliance. You can’t really ask a patient whether they took the drug and trust that information. But something like heart rate or temperature or things like that, might not tell the full story of compliance. We have bioanalysis shops that are doing much more intricate experiments and measurements to try to understand what’s happening. I’m just curious, where do you see all of this converging at some point where things like compliance or things like bioanalysis can happen in closer to real time? Or do you think there’s always going to be this wall between what’s happening with the patient while they’re out in the world and then the samples that need to be shipped around and understood?
Mitra Mosharraf: Yeah, I think that if it will change to better because the way we make drugs will change too. As we go through this revolution, I can see that in the future, probably we don’t even need to take the drugs, maybe they would just be released inside our body when there is a need by measuring with sensors exactly what’s the concentration at each time inside our body and when a release needs to happen, that can be completely automated. It’s a long-term strategy, but I think we will get there. There are now a lot of companies that are working in implants, to put in your body for slow release of drugs when it’s needed. Of course, depends on the strategy for each disease. It might work for some diseases and not for the other ones.
Ross Katz: Yeah, that’s really interesting, and it just highlights for me the way that all of these problems are being attacked from a variety of different angles.
Mitra Mosharraf: There are so many different approaches, with personalized medicine, there is a huge opportunity for AI and ML to be able to scan each patient’s biology and need. We see that now in CAR-T cell therapy or other aspects of cell therapy, and we will see more of that as we evolve more in gene therapy. But we have to be very careful with the risks also, and we have seen that some clinical trials, there were some withdrawals recently with gene therapy. I think companies are understanding that they have to really carefully investigate and watch for pitfalls. Those were not because of AI and ML, it was just because of other reasons for viral delivery. That’s why some non-viral delivery is important.
Ross Katz: Right. If I understand you correctly, nanomedicines, one of the benefits is the alternative to viral delivery, right?
Mitra Mosharraf: The nanolipid particle parts of it or polymeric nanoparticles. Yes, they are different types of nanoparticles that don’t have the same risks maybe and can reduce our risks, and then combining those with AI and ML and being able to learn what are the problems and give those to the models, right? So that we can de-risk.
Ross Katz: You’ve mentioned risk now in terms of medical risk, but I’m interested in, across the entire landscape of AI and ML applications that we’ve talked about today, what do you view as some of the biggest risks to the application of AI and ML or its misuse in the discovery, the development, the clinical application of AI and ML?
Mitra Mosharraf: I think the major risk is if we don’t evaluate the underlying data that is being used in the model. That I think is very important to put a lot of focus on that upfront before and validate the model very well so that we can trust the data, right? Because we can have false positives, false negatives, and we don’t want that. The problem with bias as we discussed earlier, distribution shifts, which would mean for example, if we are using AI for one sort of data like for cells, can we now apply that to animals? No, we have to now develop a model for animals or expand our model so that it also has different species and before we go to humans. These are very important to consider in development, to be able to trust the data, I think.
Ross Katz: There’s the tantalizing possibility of AI and ML, reducing the cost or speeding up the timelines of bringing drug to market, especially in a landscape where, as you put it at the beginning of the conversation, they take a very long time and they’re very expensive and the failure rate is very high. If you can increase the failure rate—reduce the time, reduce the cost, then the economics of the entire industry improve, but then also being held back by the generalizability of the models and the quality of the data that’s available and making sure that we’re careful about how we apply these models and the assumptions that are built into them. You can see how there’s a tug of war between those two effects and how it manifests.
Mitra Mosharraf: Oh yeah, if our model is the wrong model, we lose all that money that we have invested because we predicted wrong. It’s very important that the predictive models are predicting as correct as possible, and that is just done with that validation of the data and the model.
Ross Katz: Obviously regulators have a say in how AI and ML is applied, and I know that the FDA is continuing to update its guidance on AI in drug development. What is your sense of the direction that that regulation is going and then, if you have opinions, the direction that that regulation should go in order to enable the industry to be as impactful as it can be and while maintaining, obviously, patient safety?
Mitra Mosharraf: Yeah, I think FDA is doing its best to catch up with this revolution of AI and ML. It does take time because they need to also—I read that they were applying AI and ML inside FDA also and they had created a committee to oversee that work just for their own work also. In 2025, they gave new additional guidelines, now we have I think five guidelines that I have seen for AI and ML in manufacturing and also in development of biologics and nanomedicine as part of that. These are very good, but there is of course much more needed because there is no clear guideline like how do we exactly, the standardization of protocols for using AI and ML, I think they need also more feedback from industry and this collaboration between data scientist, FDA and the industry as a feedback loop will be very important for advancing those guidelines. I hope that, because this is quite a new field for everyone, I think that’s why it’s taking a little bit time to bring those standardized protocols out, but there are consortiums that have been formed and for data sharing because we don’t have a lot of data also, especially in nanomedicine, I think that’s one of the limiting hurdles, the shortage of data. Now those platforms that have been created to share data, I hope they can help also FDA and industry to come up with better guidelines.
Ross Katz: Do you have any hopes for the protocols or the guidelines you would like to see?
Mitra Mosharraf: What I would like to see, I mean a lot of it also is like, how do you judge what is good? They need to tell the industry exactly what are they we need to look for in a model exactly, what should be the amount of data, what should be, when we say diversity in a specific model, what do we need for clinical trials, for discovery, these are I think very important to include in guidelines.
Ross Katz: Right, that I can understand from their perspective. Even experts in the field have difficulty defining what good looks like because it’s always relative. You can understand what’s better, but understanding what is good enough, especially from a regulatory perspective is a hard problem to solve.
Mitra Mosharraf: It is, we will get there though. I have no doubt that we will get there and because that’s the way it is, we do have to develop standard operating procedure for everything, and AI and ML’s part of that.
Ross Katz: Yeah, absolutely. You could say the same thing about health outcomes. What is risk? What do we consider to be safe and what do we consider to be risky? The ambiguity is across the entire ecosystem and that’s kind of their job is to help cut through that ambiguity. As we head toward the end of our conversation, I’m just interested in, what in the next two to three years is most exciting to you about the nanomedicine field, about applications for data science, AI and ML in that?
Mitra Mosharraf: Yeah, I think that we are—I hope that we can improve our understanding of cancer and how to address cancer. That would be very important, even now if we look at the market for nanolipid particles, we see that we have now shifted from, even if initially it was for rare disease and infectious disease like as vaccines, COVID-19 vaccines, now we see 33% of drugs that are in clinical trials for using mRNA LNP are for cancer therapy, which is very important. That tells us that in the future we will have cancer therapy products that would use this kind of system. In order to have more success and reduce their risk, of course they would use AI and ML too, I think, for both identifying the targets and how to make those molecules and delivery systems and automation across manufacturing. Would be also much more, have you heard about those manufacturing processes in China I heard they are making cars in dark, without any people. I think maybe things like that will come to our field in pharmaceutical sciences and pharmaceutical industry as well. It’s not good for the people so I hope not everyone gets replaced, but I think the future labs will also not only be very automated, but you could sit in other countries and just access it virtually to do your experiments using automation, that would be very neat.
Ross Katz: Yeah, interestingly, it takes away jobs, but it also increases the creative capacity for the human mind to consider what experiments to run and what things to build. As all things, there’s benefits and drawbacks to the whole technological change. As you were talking, if it’s okay to ask you one more question, we’re hearing about all of the benefits of AI and ML to the biotech ecosystem at large and to nanomedicines in particular and the economy as a whole, but measuring the effectiveness of AI and ML in realizing the value across the entire economy is really hard to do. The question that I have is what do you say to people who are skeptical of the value of applying AI and ML, in particular, inside of the nanomedicine field?
Mitra Mosharraf: Well, reports like the one I mentioned in the beginning by Wellcome Trust and BCG, they have looked at the impact that it would have on drug development and nanomedicine development as part of that would not be different. They are telling us 25% to 50% cost saving. That’s important. Increasing the success rate, we already see that, there are some clinical trials that are actually using AI and ML now, I heard. That’s important. Increasing success rate reduces risk of failure and also cost saving. Those are very important impacts that AI and ML will have apart from being able to discover new molecules to address diseases we couldn’t address before.
Ross Katz: Well, Mitra, it’s been a pleasure having you on today. I really appreciate the time. Where can listeners go to follow your work and learn more about you?
Mitra Mosharraf: Well, I am in LinkedIn. I hope that they can follow me on LinkedIn. We have also a company website, it’s h2dcorp.com. They can follow us on my company HTD Biosystems on LinkedIn as well. I am really happy to answer any questions if they ever reach me in any way.
Ross Katz: Well, thank you again for joining. Really appreciate it.
Mitra Mosharraf: Thank you so much for this opportunity. I enjoyed talking to you.
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 player of choice. See you next time.
