Aug 20, 2024
Where are the biggest opportunities to leverage AI in cancer
diagnosis and treatment? What are the biggest barriers remaining to
move away from a one-treatment-fits-all approach to treating
cancer? And how are AI, radiomics, machine learning and deep
learning helping to understand which patients will respond best to
which treatments?
We will learn all that and more in this episode of Research in
Action with Otavio Clark, M.D. Ph.D. and Principal Research
Consultant at Oracle Life Sciences.
---------------------------------------------------------
Episode Transcript:
00;00;00;00 - 00;00;26;16
Where are the biggest opportunities to leverage AI in cancer diagnosis and treatment? What are the biggest barriers remaining to move away from a one treatment fits all approach to treating cancer? And how are ready omics, machine learning and deep learning. Figuring out which patients will respond best to which treatments will learn. All that and more on research and action in the lead in the world.
00;00;26;19 - 00;00;48;02
Hello and welcome to Research and Action, brought to you by Oracle Lifesci Answers. I'm Mike Stiles, and today our guest is Ottavio Clark and Oncology and Specialty Therapeutics executive at Oracle Life Sciences. Now that's a field he's been in his entire career. He has his Ph.D. in oncology and specializes in all things evidence based research, real world data, real world evidence.
00;00;48;02 - 00;01;13;25
And what we're going to be talking about today, AI and the critical field of cancer research. Octavio, thanks a lot for being with us today. I might tinker for the end of the invite. It's a pleasure to be here. And we are really discussing a fascinating issue. That is how the ACA is changing the healthcare landscape. But before we start, I'd like to make a disclaimer.
00;01;13;28 - 00;01;45;27
We would discuss a lot about the study's findings, but we have to to have in our minds that these results that you discuss, they are still early. These findings. We have yet to be validated in prospective longer term studies, but we will discuss the only things that we have a clear direction of the trend. You added that things are going, so it's important for everybody to to think about this product by the cancer, something introductory.
00;01;45;29 - 00;02;10;17
So I think that's pointing towards trends but not about something definitive when you see something moving on in this direction. Okay. Okay. Yeah, that's totally understood and understandable that that would be the case. I do really want to dive right into this so we can make good use of our time. So what are some of the more impressive advancements that we've made in cancer treatment lately?
00;02;10;17 - 00;02;40;18
And does that mean success rates are satisfactory? Has personalized medicine helped to that? Where are those most promising opportunities to improve personalized medicine where cancer is concerned? It's a revolution in personalized medicine. It changes everything in oncology. And honestly, when I was in the medical residence in 1996, 1998, I did not think that we could see these during my lifetime spent This person.
00;02;40;25 - 00;03;15;05
The medicine has changed the way that we practice quality because it's today for many different types of tumors. We can pick treatments that are tailored to read their genetic profiles, and it enhances the precision and the effectiveness of the therapies. We left our scenario before Where do we use the same drug for everything? And now we can get the genetic profile of the patient of the tumor and try to find a targeted therapy that is limited to any specific type of cell.
00;03;15;06 - 00;03;42;03
Sometimes growth genes. This is wonderful. It has improved a lot. The outcomes of the patients have been becoming better and better in the last years, but we still have challenges here. The first one is that we don't have this kind of personalized medicine for all types of tumors, and one very important things. Not all patients respond to the personalized medicine as we would expect.
00;03;42;05 - 00;04;13;05
What it means. We still have patients that do very well, but we still have patients that don't do so well as we would want to to to have it. So the overall success rate in treating cancer with this personalized medicine approach have improved, but they are not yet 653 across all cancer types in demographics. We are still trying to see some improvements in upfront patients elections.
00;04;13;08 - 00;04;39;27
That is, how can I making this personalization even better by selecting out the fraud patients that have a similar genetic profile, but that they can I can identify those that. Do you have a good response to the therapy and those that will not get a good response to the therapy? If we could do this separation based split, we would have a much more effective treatment.
00;04;39;27 - 00;05;10;15
Of course, what are the opportunities and being able to select those patients who are most likely to respond to a particular treatment and identify those who aren't likely to respond? I mean, how might those kind of better patient classifications affect the current staging systems and the epidemiology of cancer? That's a long history. But let's start. If you if you can select patients, we will, of course, be able to do two things.
00;05;10;15 - 00;05;32;29
The first one is offering the patients that whom you will you expect to have a good response to the treatment, to give an effective treatment, and you split the basis that we expect that you not respond to that kind of therapy. To me, you try to offer them some sort of therapy or to select a clinical trial for these patients.
00;05;33;01 - 00;06;06;08
Well, how are we dealing with this? First, there is are there is an artificial intelligence to that we call radio omics today. These are the army is is is a technique that can extract huge quantities of information from medical imaging like key MRI scans and so on. And these really omics can analyze very complex patterns that we human beings can not see and it can give us an additional classification.
00;06;06;08 - 00;06;41;20
And this is something that will help us in dividing this patient, possible responders and possible night responders when we integrated these Arabian Sea tourists in deep learning machine learning technologies, we can identify the subgroups of patients that will really be more beneficial. There is a very interesting study that was recently published this year to the European Studies. This patient included 1300 patients with no small cell lung cancer without early stage disease.
00;06;41;20 - 00;07;17;16
You let these early stage stage one station through this model was able to predict three, six, seven, 6% accuracy. The patients that would be old in not have a nearly relapse just after the treatment. So they analyzed the data from 3000 patients they put inside of these machine learning system. And in this system the tools could be told that around 40% of the patients could have avoided treatment that was not effective for them.
00;07;17;19 - 00;07;44;25
40%. This number is huge and it reflects what we see in practice. Even in this personalized medicine, we still have 46% of patients that would not respond adequately. The problem is we don't know how how to split the patients to be, how to they try to station. So they and these new tools, these artificial intelligence tools, the omics machine learning, deep learning, they are offering the opportunity for this better selection.
00;07;44;28 - 00;08;17;05
And of course it opens huge opportunities for research and development because, okay, we have now these subset of patients that we respond, what do you do with those that don't respond? So it's brought to the need for developing new drugs and new tools that when you get to these subset of patients that are not responding to current treatment into new developments and new new forms of treatment, well, but it is complex and it is still in its infancy.
00;08;17;05 - 00;08;40;27
Everyone's still trying to figure out what it can and can't do best, what the best applications are, What are the complexities of bringing a high end to cancer diagnosis and treatment? And, you know, in what ways do we need to kind of be careful as we start incorporating it? Yeah, we need to be very, very careful with this because we still don't know everything about even the specialists.
00;08;40;28 - 00;09;12;16
They they really don't understand how these tools fully functions. Well, we can really spend a day discussing this topic, but I'd like to call attention to three important feature is here. The first line is we have to care about data, privacy and security because these systems, they use patient data to be treatment. You know, you have to teach the machine about what to do, about what to do, analyze, and we have to have data from real patient.
00;09;12;18 - 00;10;03;21
And often these training data sets that people are using in different approach. So we have to be sure that they have privacy of the data. The security of the data is is assuring and that we have a legal standards like HIPA and that can maintain the confidentiality and the trust of the patient in the system. The second and very important one is the bias in many of these A.I. systems that you see that we have today, because they way that they are trained and again, the machine is learning what we want them to learn and they can sometimes perpetuate or amplify biases if they are trained in data that is not representative of the food
00;10;03;24 - 00;10;32;17
of the food population. One One very good example of these is that the accuracy of some A.I. tools into the noses of a melanoma. Melanoma is that I see that has a black sheet, a black color, and it is very common in people. It can occur in white people and in black people, but they must the A.I. tools, they have a bias for the white people.
00;10;32;20 - 00;10;58;21
They are very accurate in the most melanoma, in white people. But the loss to some of these melanomas in black people because of the way they treated, because it's more common in white people. So we have to find ways to avoid it is the third point is the clinical validation. We have seen an evolution of these A.I. tools during the the last decade.
00;10;58;21 - 00;11;34;29
Don't say we. Until five years ago, we saw some publications with a few small datasets of patient and very strict validation. In the last five years, we started to see a directions towards a better validation in bigger groups of patients, but we still don't have prospective validation for most of these tools going on in our prospect. You a in large group of patients, this is still something that we will have to deal with for most of the cases in the next year.
00;11;35;02 - 00;12;11;03
And importantly, we have to evolve. If the regulation the FDA has has been trying to regulate some of these Albury's machine learning tools that we have there, there are some specific regulations already in place, but we still have to advance a lot here. And of course, and the most important one is the ethical considerations for assessing how much decision making multipolarity should we give to machine is in making decisions about patients that think about the cars, the autonomous cars that we have.
00;12;11;03 - 00;12;35;25
We basically in San Francisco, you can carry a driverless car, just enter and say, I'm going to this place and the car will make all of the decisions for you. How is it going to happen in our health care environment? So it's challenging, but it's evolving and this is working really fast. That's right. I mean, innovation only comes at us faster and these things are only going to get better, we assume.
00;12;35;25 - 00;12;59;23
But there are the remaining challenges when you think about what AI is setting out to accomplish or what you're setting out to accomplish with it, what's the difference between overall survival and progression free survival? Because those are what we want AI to predict, Right, right, right. And these are two types of message remains that you re using in oncology.
00;12;59;23 - 00;13;36;09
Mostly we using not have some specialties also, but in oncology and progression free survival is the time that we have between the date of diagnosis until the disease progresses. It means until the disease grows or the patient dies. And overall survival is the time from the date of diagnosis until the patient dies. Well, it's a it's a small difference, but the PFC, the progression free survival is mostly about the disease evolution itself and the overall survival is about the patient evolution itself.
00;13;36;12 - 00;14;10;28
So the progression free survival, disease, free survival are very treatable to be used in in the types of cancer that have a long evolution time like stage two melanoma, we are talking about melanoma, stage two, stage three melanomas that have an evolution of years, sometimes decades. If you wait until measuring the efficacy of a drug of therapy in overall survival, it would take ten, 15 years because of the time of the evolution of death of the disease.
00;14;11;00 - 00;14;43;17
So we tend to use this progression free survival. That is the time until the disease grows again, relapse or something like this. And overall survival is the same because it is the lifetime of survival of the of the patient. Sure. Well, not all cancers are the same. Of course not all people are the same. So is there any AI driven methodology that can not necessarily personalize, but just segment Ty's patient populations and kind of point them to various treatments that are going to work best for them?
00;14;43;17 - 00;15;16;09
And and if so, how does that dream call for real world data? How is real world data applied in that process? Well, we classify today we classify cancers based on the region of the body or the organ that eat it appear like lung cancer, breast cancer, liver cancer and so on. We have in the last, I would say, true 2 to 3 decades evolving towards a more specific classification.
00;15;16;11 - 00;15;46;28
That is some types of cancer, for instance, that we have in breast cancer. We have today many different types based on the genetic profile of the tumors, like if the tumor has specific times of a symptoms like estrogen progesterone or one that we call how true and why they do, we evolved towards these subclasses situation. We perceive that that inside of breast cancer or lung cancer.
00;15;46;28 - 00;16;19;27
But let's talk about breast cancer. We had patients we have a very different prognosis that would respond very differently to treat to different treatment. And as the knowledge evolved, we were able to classify, let's take of both these three subtypes HER2 positive or negative estrogen receptor positive or negative progesterone receptor positive or negative. So by doing this some classification, we could offer better treatments for each of these subpopulation.
00;16;20;00 - 00;17;05;01
And then this is where the awarded data entries in our in our discussion, because today we have the new tools in a we can get information from huge datasets, from websites that have millions of patients, he said. Like electronic I have to records claims from hospitals, insurance claims, death certificates and everything that we were not able to do before in way because these official tally systems are able to enter in electronic health records like for this has been one that Oracle had.
00;17;05;04 - 00;17;55;25
We can use some tools that we call metrology natural language processor that can read the records that were imputed by the doctor. And it came extracted the patients that we want. We can give a common to the in AP to and say look please select from these 100 million records only patients with stage three breast cancer had two positive had estrogen receptor negative and it can go inside of the electronic I have to and from the 100 million records that you have that he can come back with 500,000 patients and then again using AI tools like deep learning, machine learning, many different tools, we can get to this 500,000 patients in the models, how they
00;17;55;25 - 00;18;27;06
were treated, how they did evolve in the real world, not only in the control edit environment of clinical studies. So we can evaluate how the drugs that are approved out of performance in field would. That is a completely different environment from the, from the clinical studies. We can try to identify subsets of patients that do have a better or worse response to a given treatment in.
00;18;27;08 - 00;18;55;21
We can also use these tools to make very sophisticated integrations. Often the patient profile will be if the genomics of the patients and this is where we are seeing a lot of development in the last year, is this attempt off of trying to identify how this genomics influences the treatment of of the patients. There is a good example for that.
00;18;55;23 - 00;19;28;04
We have we did a study using our electronic health records that was presented at ASCO last year and that we use in the Metro language processing tools to identify among millions of records patients that had a very specific, very rare mutation that is called entity AKI or anthrax. People, as people call it. These mutations very rare. It occurs in 0.22% of the tumors at most.
00;19;28;06 - 00;19;57;14
But the importance of these is that today we have three drugs in the market that very design needed to act against any specific mutation. But this is very rare. These thirds that way, Don, it included 50 patients, seven patients, and that was it. So by looking at our electronic I have to heck with those in these tools, we could identify 200 patients with these.
00;19;57;14 - 00;20;34;06
An entire chemo patient that looks very, very ridiculous is small number, but it's not at the time it was the largest cohort of patients with these mutations that very study and the data were published looking for patterns. How did they respond to treatment? They they said in all of these using here word data. So if we can leverage it, I wish that here were the evidence, we can add this in much better health treatment performances in very diverse populations, and we can adjust the strategies to improve the patient's outcomes effectively.
00;20;34;06 - 00;20;56;10
And this is how I am here with this data interconnect. Well, I think about rank and file health care providers. It feels like we're still pretty far away from it being used by health care providers to definitively make treatment decisions. And I think about how busy they are. They don't have time to make themselves experts in AI technologies.
00;20;56;10 - 00;21;24;03
So what kind of partnerships or collaborations have to happen in order to make A.I. analysis usable by your average doctor? Well, the first thing it has to be easy. It has to be easier than what we have today because today the doctors already spend a lot of time in administrative tasks like the entry date, and then they let them go head to head with the feeling forms.
00;21;24;05 - 00;22;02;06
And the first thing that we have to to keep in mind is that our objective today is about it's talking about the impact of action that is in the health care clinical setting itself decisions, how doctors decide how drugs are delivered. But hey, we also and these are really being integrated in the clinic in the administrative part of the medical practice by selecting codes, building codes by there are some systems that are able to field a little bit for the patient.
00;22;02;08 - 00;22;28;28
So these will make the lives of the doctor easier. And I think that these administrative tools will be very well received by the doctors. That's my impression and good design impression. But we we also have a not I but a part of that is what we're discussing here, is how these tools will be integrated in the decision making process of the doctor and in the practice.
00;22;29;00 - 00;22;59;10
And to do this, we have to find ways to make it easy for the doctor to use. How can we do it? The first thing is we have to put it developers in health care professionals together so those that are developing the system, they can understand the needs, they can understand the difficulties, and they can understand where the improvements are necessary, how it will work to help them better the patient.
00;22;59;13 - 00;23;28;29
The second, of course, academic institutions tend to be early adopters of these of these tools, and they will be very important in not only developing but conducting independent validation of studies that like the ones that we cited here, that that will predict the outcomes of patients. So do the patient selection. Are we talking about to be FDA and the regulatory board?
00;23;28;29 - 00;24;05;21
They are essential for for these to guide us towards the right direction about what we can, what we can do. And this will be a huge impact in everything that we do if or when one of these eight companies or researches is able to develop an interface that will make the life of the doctor easier with information that is very trustable, the adoption will be very quick, I think.
00;24;05;23 - 00;24;35;09
Well, I've heard about a I don't know if it's a tool or a program called Deep Profiler. What is that? I mean, it's, you know, studies undertaken around AI's use in cancer diagnosis and treatment. What is it? What can it do and what are we learning from it? The profile can refer to too many different things. It usually used it to describe different tools that are integrated in a system that can use deep learning techniques.
00;24;35;09 - 00;25;03;10
And they came in. My eyes are very complex data. They can go to a huge data set of genomic information for the users, try to cross the information we should often these genomic database will be information from, let's say, medical images and try to find out to see how they how they are combining together, what they can offer to the medical decision process.
00;25;03;12 - 00;25;50;24
Usually the aim of these deep profilers, these things are to to give me a more accurate biological profile of the tumor of the patient and say, look, this patient with this kind of profile is will respond better if these treatment, the patient with this profile will have a better outcome or a worse outcome or something like this, that these these the profile is they are using it to do, as I said, genomic and molecular characterization predict response to drugs, and they will try to to point towards the gaps in the development and identify where reception developing it is necessary.
00;25;50;26 - 00;26;17;25
This is this is how these how these deep profiler systems works based integrating huge amounts of data from genomic is from molecular, is from saw, from images, and try to make different profiles to help in the decision making process. Well, we all heard about, you know, people who there was a suspicious suspicion of cancer and a biopsy was done.
00;26;17;25 - 00;26;47;06
So that involves examining tissue samples. But are there any other parameters other than the biomarkers found in tissue only samples that might lead to better predicting durable clinical benefit? Yeah, this is what the Radiometer is trying to do. They are trying to identify based on the medical images, not only biopsy, not antibiotics, they are trying to say, okay, this is the image of noise, muscle, lung cancer in a patient.
00;26;47;09 - 00;27;20;21
This has this type of profile and we don't need the complex, time consuming biopsy tests anymore to see that this patient is from A, category A, B, C, or D, and should be treated with the drug H, Z, or the. This is basically what you are doing because this process of the biopsy and sending this sample to to a lab to do the pathology, to do the genetic test, it's it's time consuming.
00;27;20;21 - 00;27;47;28
It's very expensive. So if we can find a way that basically the only on the medical images, PET scan, an MRI could already see what is the profile of the tumor. We would experiment, we would spend time, we could study the treatment of the patient very early so that they are damaged. That's one of the attempts that misread the cell, trying to do well.
00;27;47;28 - 00;28;15;18
We've all known someone with a weird looking mole who went to the dermatologist, got checked for melanoma. Are you kind of saying that A I driven approaches are appearing to outperform, just getting looked at by a dermatologist for spotting things like melanoma early? That's one of the things that we already have data that really points towards to an almost conclusive thing.
00;28;15;20 - 00;28;42;07
But the studies that have been published using these tools to analyze skin spots on their early diagnose of melanoma, they are performing better than dermatologists. They are already 30% better than the dermatologists that they point out. What what they like to call the attention here is that it's not only about the eight to being better than the dermatologists.
00;28;42;07 - 00;29;09;08
So this is this is true, but it's not only based can you imagine about someone in a very remote area that see a sporty skin. He could just take a picture, send the speaker through to a central hospital or to a clinic yet to be analyzed. He does not need it to get a car to go there and then bathe in the only on the photo that he took.
00;29;09;10 - 00;29;29;26
We can see if this is malignant or not. If the patient needs to travel for miles and miles and make an appointment. So it's not only about the air itself, it's about the consequences of the good use of this tools that we have to think about. Well, I want to wrap up with a few relatable kind of point blank questions.
00;29;29;29 - 00;29;56;29
If I'm a patient and it's determined, a typical course of treatment is unlikely to be effective for me. What happens then? I go to a plan B, or do I get referred to a clinical research program or and how does I help make those kinds of determinations? Yeah, that's a very good question, because it will happen as soon as these systems are in place to separate the patients upfront.
00;29;57;01 - 00;30;34;06
When you have a huge number of patients that you needed this kind of case and then it comes to two or three things. The first one is we can also use, as we spoke before, these eight tools in huge electronic I have to here could he word data to try to make the characteristics of this patient that he not responded to a given therapy with other patient that are inside of this 100 million headquarters in an electronic health record that had the same problem, had the same profile.
00;30;34;08 - 00;31;07;23
And then he by matching the patient, we were better treated, which he had better outcomes inside of these he already data electronic to her take on this took it to a practical was this the first one the second one is it you create the need for new research in development of new drugs, develop the development of new clinical trials and we have a systems that they we call it matching.
00;31;07;25 - 00;31;31;05
They match the patient. We have clinical trial. How does it work today? The process of sending a patient for a clinical trial is very time consuming and is really ineffective because I have to to get the data from the patient. Let's get the patient is no is my cell lung cancer EGFR positive? I have to say, okay, where do I have a clinical trial for this patient?
00;31;31;05 - 00;31;50;19
And then I have to look around and see if there is a clinical trial for this patient. So there are already today. Some day I may systems that you research the data of the patient in the system and then you'd say, okay, in this patient feet, this clinical trial that is been doing in this place and this recruiting patient.
00;31;50;19 - 00;32;40;19
So that's the second way that these tools can help these patients that will be, let's see, unselected for the treatment. It's really how we can help this patient. And I think also we have seen it's not directly related to the daily care, but we have seen some new drugs that are designed by a they get the researchers inserts the configuration of a protein from the tumor cell in in one of these machine learning deep learning system and say please design one molecule that can bind here and that could inhibit the proliferation of the tumor.
00;32;40;25 - 00;33;07;14
And there has been some success, not yet specifically in quality, but we have seen some success in antibiotic use and in inflammatory diseases by designing molecules that will be tested in human beings. You know, all this is absolutely amazing and exciting to hear and think about, but the question from the public always comes back when all this is still being tested and vetted.
00;33;07;14 - 00;33;33;11
When can we expect to see the benefits of these capabilities show up in the field? Is that one year, five year, ten years? That's a good question. And of course, if you be headed to and so we would not be here talking about it. But I would like to to give my impression and again, this is my impression we to talk about short term, let's say up to three years.
00;33;33;14 - 00;34;00;28
What I see happening makes it three years is a better integration of some of these supports diagnostic tools that we some of them that you talk about early in their clinical practice that will improve. They did some of the decisions not not largely by but in some very specific cases like melanoma and mammography for breast cancer. Some of these radiometer from a my cell lung cancer.
00;34;01;00 - 00;34;28;12
And it's going to be of course, in mainly in the major health care centers. Then we can talk about, let's say, medium term, 3 to 5, six, six years. And then in this length of time, what I see is that they studies, I will matter and they will gain regulatory approval. They will be validated in the clinical settings and they use you.
00;34;28;12 - 00;35;02;10
You will expand maybe exposure in this period of time because it will be you have it in my view, already the system, the regulator approval and the validation in in prospective studies. So it you basically make the adoption of these tools not only desirable but almost mandatory. It would be like if I knew I new drugs that is is these is efficient to treat the patient comes to the market.
00;35;02;15 - 00;35;32;06
I think that this is how it will be seen and in the long term, let's say ten years these tools will be widespread everywhere and it's not going to be restricted to more sophisticated places. It will be I see it being taken the word everywhere because we have to remember a good part of these decisions, tools. They not only help, but they make the system more efficient and less costly.
00;35;32;08 - 00;36;00;25
So to me, you likely take the word fast. What I'd like to to call that definition is that they speed of the adoption of these tools with you depend on many different factors, but you defend in the one thing also that I think is very important is how how much do we tolerate mistakes made by machines? Because one things about a human being make a mistake.
00;36;00;27 - 00;36;32;10
A person can give a different perception and in a medical MRI, for instance. But what happens if a machine does it? Is it going to be tolerated or not? And I like to think about it when I think about the autonomous cars every single day we have it is maybe thousands of patients being heard in car accidents, then somehow reported in the TV or in the newspapers or something like this.
00;36;32;10 - 00;37;09;24
But if we have one accident, we should one of these autonomous cars that does not injury anybody, It is reported everywhere. So I think about how we make these how we. So yeah how is our our room in this because objectively the accident that I caused and I checked it this is that this is the accidents that are caused by autonomous car they are much less severe than those that are caused by human beings in a million driven basis analysis.
00;37;09;27 - 00;37;34;28
But you have to really press them if the accidents that are caused by autonomous car, how are you going to act? If there is things that to happen? It's unavoidable. When one of these machine tools make a mistake that this even if we prove mathematically that it's better, it makes less mistakes than if a doctor see an MRI or a mammogram.
00;37;34;28 - 00;37;58;07
So basically, our our ability to tolerate errors that will have a huge role in the adoption of A.I.. Yeah, we're we're very forgiving of humans and not forgiving at all of technology and and machines. So, yeah, we it's on us to be open to adoption. All of that sounds great. And thanks again so much for being our guest today.
00;37;58;12 - 00;38;22;17
Otavio Again, cancer is something that's touched almost everyone's life in some way and we're excited for any advancements we can expect. And it's diagnosis and treatment. If our listeners want to learn more about Oracle's initiative or to get back in touch with you, is there a way for them to do that? Yeah, they can reach us out at Oracle dot com and Oracle has a lot of initiatives in.
00;38;22;20 - 00;39;10;11
Yeah, in health care we have a huge team here studying, especially the integration of Iris here. We do have this data. Great. Got it. Well thanks again, Octavio. And to our listeners, we don't want you to miss any episodes of research and action, so please subscribe and if you want to learn more about how Oracle can accelerate your own life sciences research, you can just go to Oracle dot com slash life dash sciences and we'll see you next time.
Where are the biggest opportunities to leverage AI in cancer diagnosis and treatment? What are the biggest barriers remaining to move away from a one treatment fits all approach to treating cancer? And how are ready omics, machine learning and deep learning. Figuring out which patients will respond best to which treatments will learn. All that and more on research and action in the lead in the world.
00;00;26;19 - 00;00;48;02
Hello and welcome to Research and Action, brought to you by Oracle Lifesci Answers. I'm Mike Stiles, and today our guest is Ottavio Clark and Oncology and Specialty Therapeutics executive at Oracle Life Sciences. Now that's a field he's been in his entire career. He has his Ph.D. in oncology and specializes in all things evidence based research, real world data, real world evidence.
00;00;48;02 - 00;01;13;25
And what we're going to be talking about today, AI and the critical field of cancer research. Octavio, thanks a lot for being with us today. I might tinker for the end of the invite. It's a pleasure to be here. And we are really discussing a fascinating issue. That is how the ACA is changing the healthcare landscape. But before we start, I'd like to make a disclaimer.
00;01;13;28 - 00;01;45;27
We would discuss a lot about the study's findings, but we have to to have in our minds that these results that you discuss, they are still early. These findings. We have yet to be validated in prospective longer term studies, but we will discuss the only things that we have a clear direction of the trend. You added that things are going, so it's important for everybody to to think about this product by the cancer, something introductory.
00;01;45;29 - 00;02;10;17
So I think that's pointing towards trends but not about something definitive when you see something moving on in this direction. Okay. Okay. Yeah, that's totally understood and understandable that that would be the case. I do really want to dive right into this so we can make good use of our time. So what are some of the more impressive advancements that we've made in cancer treatment lately?
00;02;10;17 - 00;02;40;18
And does that mean success rates are satisfactory? Has personalized medicine helped to that? Where are those most promising opportunities to improve personalized medicine where cancer is concerned? It's a revolution in personalized medicine. It changes everything in oncology. And honestly, when I was in the medical residence in 1996, 1998, I did not think that we could see these during my lifetime spent This person.
00;02;40;25 - 00;03;15;05
The medicine has changed the way that we practice quality because it's today for many different types of tumors. We can pick treatments that are tailored to read their genetic profiles, and it enhances the precision and the effectiveness of the therapies. We left our scenario before Where do we use the same drug for everything? And now we can get the genetic profile of the patient of the tumor and try to find a targeted therapy that is limited to any specific type of cell.
00;03;15;06 - 00;03;42;03
Sometimes growth genes. This is wonderful. It has improved a lot. The outcomes of the patients have been becoming better and better in the last years, but we still have challenges here. The first one is that we don't have this kind of personalized medicine for all types of tumors, and one very important things. Not all patients respond to the personalized medicine as we would expect.
00;03;42;05 - 00;04;13;05
What it means. We still have patients that do very well, but we still have patients that don't do so well as we would want to to to have it. So the overall success rate in treating cancer with this personalized medicine approach have improved, but they are not yet 653 across all cancer types in demographics. We are still trying to see some improvements in upfront patients elections.
00;04;13;08 - 00;04;39;27
That is, how can I making this personalization even better by selecting out the fraud patients that have a similar genetic profile, but that they can I can identify those that. Do you have a good response to the therapy and those that will not get a good response to the therapy? If we could do this separation based split, we would have a much more effective treatment.
00;04;39;27 - 00;05;10;15
Of course, what are the opportunities and being able to select those patients who are most likely to respond to a particular treatment and identify those who aren't likely to respond? I mean, how might those kind of better patient classifications affect the current staging systems and the epidemiology of cancer? That's a long history. But let's start. If you if you can select patients, we will, of course, be able to do two things.
00;05;10;15 - 00;05;32;29
The first one is offering the patients that whom you will you expect to have a good response to the treatment, to give an effective treatment, and you split the basis that we expect that you not respond to that kind of therapy. To me, you try to offer them some sort of therapy or to select a clinical trial for these patients.
00;05;33;01 - 00;06;06;08
Well, how are we dealing with this? First, there is are there is an artificial intelligence to that we call radio omics today. These are the army is is is a technique that can extract huge quantities of information from medical imaging like key MRI scans and so on. And these really omics can analyze very complex patterns that we human beings can not see and it can give us an additional classification.
00;06;06;08 - 00;06;41;20
And this is something that will help us in dividing this patient, possible responders and possible night responders when we integrated these Arabian Sea tourists in deep learning machine learning technologies, we can identify the subgroups of patients that will really be more beneficial. There is a very interesting study that was recently published this year to the European Studies. This patient included 1300 patients with no small cell lung cancer without early stage disease.
00;06;41;20 - 00;07;17;16
You let these early stage stage one station through this model was able to predict three, six, seven, 6% accuracy. The patients that would be old in not have a nearly relapse just after the treatment. So they analyzed the data from 3000 patients they put inside of these machine learning system. And in this system the tools could be told that around 40% of the patients could have avoided treatment that was not effective for them.
00;07;17;19 - 00;07;44;25
40%. This number is huge and it reflects what we see in practice. Even in this personalized medicine, we still have 46% of patients that would not respond adequately. The problem is we don't know how how to split the patients to be, how to they try to station. So they and these new tools, these artificial intelligence tools, the omics machine learning, deep learning, they are offering the opportunity for this better selection.
00;07;44;28 - 00;08;17;05
And of course it opens huge opportunities for research and development because, okay, we have now these subset of patients that we respond, what do you do with those that don't respond? So it's brought to the need for developing new drugs and new tools that when you get to these subset of patients that are not responding to current treatment into new developments and new new forms of treatment, well, but it is complex and it is still in its infancy.
00;08;17;05 - 00;08;40;27
Everyone's still trying to figure out what it can and can't do best, what the best applications are, What are the complexities of bringing a high end to cancer diagnosis and treatment? And, you know, in what ways do we need to kind of be careful as we start incorporating it? Yeah, we need to be very, very careful with this because we still don't know everything about even the specialists.
00;08;40;28 - 00;09;12;16
They they really don't understand how these tools fully functions. Well, we can really spend a day discussing this topic, but I'd like to call attention to three important feature is here. The first line is we have to care about data, privacy and security because these systems, they use patient data to be treatment. You know, you have to teach the machine about what to do, about what to do, analyze, and we have to have data from real patient.
00;09;12;18 - 00;10;03;21
And often these training data sets that people are using in different approach. So we have to be sure that they have privacy of the data. The security of the data is is assuring and that we have a legal standards like HIPA and that can maintain the confidentiality and the trust of the patient in the system. The second and very important one is the bias in many of these A.I. systems that you see that we have today, because they way that they are trained and again, the machine is learning what we want them to learn and they can sometimes perpetuate or amplify biases if they are trained in data that is not representative of the food
00;10;03;24 - 00;10;32;17
of the food population. One One very good example of these is that the accuracy of some A.I. tools into the noses of a melanoma. Melanoma is that I see that has a black sheet, a black color, and it is very common in people. It can occur in white people and in black people, but they must the A.I. tools, they have a bias for the white people.
00;10;32;20 - 00;10;58;21
They are very accurate in the most melanoma, in white people. But the loss to some of these melanomas in black people because of the way they treated, because it's more common in white people. So we have to find ways to avoid it is the third point is the clinical validation. We have seen an evolution of these A.I. tools during the the last decade.
00;10;58;21 - 00;11;34;29
Don't say we. Until five years ago, we saw some publications with a few small datasets of patient and very strict validation. In the last five years, we started to see a directions towards a better validation in bigger groups of patients, but we still don't have prospective validation for most of these tools going on in our prospect. You a in large group of patients, this is still something that we will have to deal with for most of the cases in the next year.
00;11;35;02 - 00;12;11;03
And importantly, we have to evolve. If the regulation the FDA has has been trying to regulate some of these Albury's machine learning tools that we have there, there are some specific regulations already in place, but we still have to advance a lot here. And of course, and the most important one is the ethical considerations for assessing how much decision making multipolarity should we give to machine is in making decisions about patients that think about the cars, the autonomous cars that we have.
00;12;11;03 - 00;12;35;25
We basically in San Francisco, you can carry a driverless car, just enter and say, I'm going to this place and the car will make all of the decisions for you. How is it going to happen in our health care environment? So it's challenging, but it's evolving and this is working really fast. That's right. I mean, innovation only comes at us faster and these things are only going to get better, we assume.
00;12;35;25 - 00;12;59;23
But there are the remaining challenges when you think about what AI is setting out to accomplish or what you're setting out to accomplish with it, what's the difference between overall survival and progression free survival? Because those are what we want AI to predict, Right, right, right. And these are two types of message remains that you re using in oncology.
00;12;59;23 - 00;13;36;09
Mostly we using not have some specialties also, but in oncology and progression free survival is the time that we have between the date of diagnosis until the disease progresses. It means until the disease grows or the patient dies. And overall survival is the time from the date of diagnosis until the patient dies. Well, it's a it's a small difference, but the PFC, the progression free survival is mostly about the disease evolution itself and the overall survival is about the patient evolution itself.
00;13;36;12 - 00;14;10;28
So the progression free survival, disease, free survival are very treatable to be used in in the types of cancer that have a long evolution time like stage two melanoma, we are talking about melanoma, stage two, stage three melanomas that have an evolution of years, sometimes decades. If you wait until measuring the efficacy of a drug of therapy in overall survival, it would take ten, 15 years because of the time of the evolution of death of the disease.
00;14;11;00 - 00;14;43;17
So we tend to use this progression free survival. That is the time until the disease grows again, relapse or something like this. And overall survival is the same because it is the lifetime of survival of the of the patient. Sure. Well, not all cancers are the same. Of course not all people are the same. So is there any AI driven methodology that can not necessarily personalize, but just segment Ty's patient populations and kind of point them to various treatments that are going to work best for them?
00;14;43;17 - 00;15;16;09
And and if so, how does that dream call for real world data? How is real world data applied in that process? Well, we classify today we classify cancers based on the region of the body or the organ that eat it appear like lung cancer, breast cancer, liver cancer and so on. We have in the last, I would say, true 2 to 3 decades evolving towards a more specific classification.
00;15;16;11 - 00;15;46;28
That is some types of cancer, for instance, that we have in breast cancer. We have today many different types based on the genetic profile of the tumors, like if the tumor has specific times of a symptoms like estrogen progesterone or one that we call how true and why they do, we evolved towards these subclasses situation. We perceive that that inside of breast cancer or lung cancer.
00;15;46;28 - 00;16;19;27
But let's talk about breast cancer. We had patients we have a very different prognosis that would respond very differently to treat to different treatment. And as the knowledge evolved, we were able to classify, let's take of both these three subtypes HER2 positive or negative estrogen receptor positive or negative progesterone receptor positive or negative. So by doing this some classification, we could offer better treatments for each of these subpopulation.
00;16;20;00 - 00;17;05;01
And then this is where the awarded data entries in our in our discussion, because today we have the new tools in a we can get information from huge datasets, from websites that have millions of patients, he said. Like electronic I have to records claims from hospitals, insurance claims, death certificates and everything that we were not able to do before in way because these official tally systems are able to enter in electronic health records like for this has been one that Oracle had.
00;17;05;04 - 00;17;55;25
We can use some tools that we call metrology natural language processor that can read the records that were imputed by the doctor. And it came extracted the patients that we want. We can give a common to the in AP to and say look please select from these 100 million records only patients with stage three breast cancer had two positive had estrogen receptor negative and it can go inside of the electronic I have to and from the 100 million records that you have that he can come back with 500,000 patients and then again using AI tools like deep learning, machine learning, many different tools, we can get to this 500,000 patients in the models, how they
00;17;55;25 - 00;18;27;06
were treated, how they did evolve in the real world, not only in the control edit environment of clinical studies. So we can evaluate how the drugs that are approved out of performance in field would. That is a completely different environment from the, from the clinical studies. We can try to identify subsets of patients that do have a better or worse response to a given treatment in.
00;18;27;08 - 00;18;55;21
We can also use these tools to make very sophisticated integrations. Often the patient profile will be if the genomics of the patients and this is where we are seeing a lot of development in the last year, is this attempt off of trying to identify how this genomics influences the treatment of of the patients. There is a good example for that.
00;18;55;23 - 00;19;28;04
We have we did a study using our electronic health records that was presented at ASCO last year and that we use in the Metro language processing tools to identify among millions of records patients that had a very specific, very rare mutation that is called entity AKI or anthrax. People, as people call it. These mutations very rare. It occurs in 0.22% of the tumors at most.
00;19;28;06 - 00;19;57;14
But the importance of these is that today we have three drugs in the market that very design needed to act against any specific mutation. But this is very rare. These thirds that way, Don, it included 50 patients, seven patients, and that was it. So by looking at our electronic I have to heck with those in these tools, we could identify 200 patients with these.
00;19;57;14 - 00;20;34;06
An entire chemo patient that looks very, very ridiculous is small number, but it's not at the time it was the largest cohort of patients with these mutations that very study and the data were published looking for patterns. How did they respond to treatment? They they said in all of these using here word data. So if we can leverage it, I wish that here were the evidence, we can add this in much better health treatment performances in very diverse populations, and we can adjust the strategies to improve the patient's outcomes effectively.
00;20;34;06 - 00;20;56;10
And this is how I am here with this data interconnect. Well, I think about rank and file health care providers. It feels like we're still pretty far away from it being used by health care providers to definitively make treatment decisions. And I think about how busy they are. They don't have time to make themselves experts in AI technologies.
00;20;56;10 - 00;21;24;03
So what kind of partnerships or collaborations have to happen in order to make A.I. analysis usable by your average doctor? Well, the first thing it has to be easy. It has to be easier than what we have today because today the doctors already spend a lot of time in administrative tasks like the entry date, and then they let them go head to head with the feeling forms.
00;21;24;05 - 00;22;02;06
And the first thing that we have to to keep in mind is that our objective today is about it's talking about the impact of action that is in the health care clinical setting itself decisions, how doctors decide how drugs are delivered. But hey, we also and these are really being integrated in the clinic in the administrative part of the medical practice by selecting codes, building codes by there are some systems that are able to field a little bit for the patient.
00;22;02;08 - 00;22;28;28
So these will make the lives of the doctor easier. And I think that these administrative tools will be very well received by the doctors. That's my impression and good design impression. But we we also have a not I but a part of that is what we're discussing here, is how these tools will be integrated in the decision making process of the doctor and in the practice.
00;22;29;00 - 00;22;59;10
And to do this, we have to find ways to make it easy for the doctor to use. How can we do it? The first thing is we have to put it developers in health care professionals together so those that are developing the system, they can understand the needs, they can understand the difficulties, and they can understand where the improvements are necessary, how it will work to help them better the patient.
00;22;59;13 - 00;23;28;29
The second, of course, academic institutions tend to be early adopters of these of these tools, and they will be very important in not only developing but conducting independent validation of studies that like the ones that we cited here, that that will predict the outcomes of patients. So do the patient selection. Are we talking about to be FDA and the regulatory board?
00;23;28;29 - 00;24;05;21
They are essential for for these to guide us towards the right direction about what we can, what we can do. And this will be a huge impact in everything that we do if or when one of these eight companies or researches is able to develop an interface that will make the life of the doctor easier with information that is very trustable, the adoption will be very quick, I think.
00;24;05;23 - 00;24;35;09
Well, I've heard about a I don't know if it's a tool or a program called Deep Profiler. What is that? I mean, it's, you know, studies undertaken around AI's use in cancer diagnosis and treatment. What is it? What can it do and what are we learning from it? The profile can refer to too many different things. It usually used it to describe different tools that are integrated in a system that can use deep learning techniques.
00;24;35;09 - 00;25;03;10
And they came in. My eyes are very complex data. They can go to a huge data set of genomic information for the users, try to cross the information we should often these genomic database will be information from, let's say, medical images and try to find out to see how they how they are combining together, what they can offer to the medical decision process.
00;25;03;12 - 00;25;50;24
Usually the aim of these deep profilers, these things are to to give me a more accurate biological profile of the tumor of the patient and say, look, this patient with this kind of profile is will respond better if these treatment, the patient with this profile will have a better outcome or a worse outcome or something like this, that these these the profile is they are using it to do, as I said, genomic and molecular characterization predict response to drugs, and they will try to to point towards the gaps in the development and identify where reception developing it is necessary.
00;25;50;26 - 00;26;17;25
This is this is how these how these deep profiler systems works based integrating huge amounts of data from genomic is from molecular, is from saw, from images, and try to make different profiles to help in the decision making process. Well, we all heard about, you know, people who there was a suspicious suspicion of cancer and a biopsy was done.
00;26;17;25 - 00;26;47;06
So that involves examining tissue samples. But are there any other parameters other than the biomarkers found in tissue only samples that might lead to better predicting durable clinical benefit? Yeah, this is what the Radiometer is trying to do. They are trying to identify based on the medical images, not only biopsy, not antibiotics, they are trying to say, okay, this is the image of noise, muscle, lung cancer in a patient.
00;26;47;09 - 00;27;20;21
This has this type of profile and we don't need the complex, time consuming biopsy tests anymore to see that this patient is from A, category A, B, C, or D, and should be treated with the drug H, Z, or the. This is basically what you are doing because this process of the biopsy and sending this sample to to a lab to do the pathology, to do the genetic test, it's it's time consuming.
00;27;20;21 - 00;27;47;28
It's very expensive. So if we can find a way that basically the only on the medical images, PET scan, an MRI could already see what is the profile of the tumor. We would experiment, we would spend time, we could study the treatment of the patient very early so that they are damaged. That's one of the attempts that misread the cell, trying to do well.
00;27;47;28 - 00;28;15;18
We've all known someone with a weird looking mole who went to the dermatologist, got checked for melanoma. Are you kind of saying that A I driven approaches are appearing to outperform, just getting looked at by a dermatologist for spotting things like melanoma early? That's one of the things that we already have data that really points towards to an almost conclusive thing.
00;28;15;20 - 00;28;42;07
But the studies that have been published using these tools to analyze skin spots on their early diagnose of melanoma, they are performing better than dermatologists. They are already 30% better than the dermatologists that they point out. What what they like to call the attention here is that it's not only about the eight to being better than the dermatologists.
00;28;42;07 - 00;29;09;08
So this is this is true, but it's not only based can you imagine about someone in a very remote area that see a sporty skin. He could just take a picture, send the speaker through to a central hospital or to a clinic yet to be analyzed. He does not need it to get a car to go there and then bathe in the only on the photo that he took.
00;29;09;10 - 00;29;29;26
We can see if this is malignant or not. If the patient needs to travel for miles and miles and make an appointment. So it's not only about the air itself, it's about the consequences of the good use of this tools that we have to think about. Well, I want to wrap up with a few relatable kind of point blank questions.
00;29;29;29 - 00;29;56;29
If I'm a patient and it's determined, a typical course of treatment is unlikely to be effective for me. What happens then? I go to a plan B, or do I get referred to a clinical research program or and how does I help make those kinds of determinations? Yeah, that's a very good question, because it will happen as soon as these systems are in place to separate the patients upfront.
00;29;57;01 - 00;30;34;06
When you have a huge number of patients that you needed this kind of case and then it comes to two or three things. The first one is we can also use, as we spoke before, these eight tools in huge electronic I have to here could he word data to try to make the characteristics of this patient that he not responded to a given therapy with other patient that are inside of this 100 million headquarters in an electronic health record that had the same problem, had the same profile.
00;30;34;08 - 00;31;07;23
And then he by matching the patient, we were better treated, which he had better outcomes inside of these he already data electronic to her take on this took it to a practical was this the first one the second one is it you create the need for new research in development of new drugs, develop the development of new clinical trials and we have a systems that they we call it matching.
00;31;07;25 - 00;31;31;05
They match the patient. We have clinical trial. How does it work today? The process of sending a patient for a clinical trial is very time consuming and is really ineffective because I have to to get the data from the patient. Let's get the patient is no is my cell lung cancer EGFR positive? I have to say, okay, where do I have a clinical trial for this patient?
00;31;31;05 - 00;31;50;19
And then I have to look around and see if there is a clinical trial for this patient. So there are already today. Some day I may systems that you research the data of the patient in the system and then you'd say, okay, in this patient feet, this clinical trial that is been doing in this place and this recruiting patient.
00;31;50;19 - 00;32;40;19
So that's the second way that these tools can help these patients that will be, let's see, unselected for the treatment. It's really how we can help this patient. And I think also we have seen it's not directly related to the daily care, but we have seen some new drugs that are designed by a they get the researchers inserts the configuration of a protein from the tumor cell in in one of these machine learning deep learning system and say please design one molecule that can bind here and that could inhibit the proliferation of the tumor.
00;32;40;25 - 00;33;07;14
And there has been some success, not yet specifically in quality, but we have seen some success in antibiotic use and in inflammatory diseases by designing molecules that will be tested in human beings. You know, all this is absolutely amazing and exciting to hear and think about, but the question from the public always comes back when all this is still being tested and vetted.
00;33;07;14 - 00;33;33;11
When can we expect to see the benefits of these capabilities show up in the field? Is that one year, five year, ten years? That's a good question. And of course, if you be headed to and so we would not be here talking about it. But I would like to to give my impression and again, this is my impression we to talk about short term, let's say up to three years.
00;33;33;14 - 00;34;00;28
What I see happening makes it three years is a better integration of some of these supports diagnostic tools that we some of them that you talk about early in their clinical practice that will improve. They did some of the decisions not not largely by but in some very specific cases like melanoma and mammography for breast cancer. Some of these radiometer from a my cell lung cancer.
00;34;01;00 - 00;34;28;12
And it's going to be of course, in mainly in the major health care centers. Then we can talk about, let's say, medium term, 3 to 5, six, six years. And then in this length of time, what I see is that they studies, I will matter and they will gain regulatory approval. They will be validated in the clinical settings and they use you.
00;34;28;12 - 00;35;02;10
You will expand maybe exposure in this period of time because it will be you have it in my view, already the system, the regulator approval and the validation in in prospective studies. So it you basically make the adoption of these tools not only desirable but almost mandatory. It would be like if I knew I new drugs that is is these is efficient to treat the patient comes to the market.
00;35;02;15 - 00;35;32;06
I think that this is how it will be seen and in the long term, let's say ten years these tools will be widespread everywhere and it's not going to be restricted to more sophisticated places. It will be I see it being taken the word everywhere because we have to remember a good part of these decisions, tools. They not only help, but they make the system more efficient and less costly.
00;35;32;08 - 00;36;00;25
So to me, you likely take the word fast. What I'd like to to call that definition is that they speed of the adoption of these tools with you depend on many different factors, but you defend in the one thing also that I think is very important is how how much do we tolerate mistakes made by machines? Because one things about a human being make a mistake.
00;36;00;27 - 00;36;32;10
A person can give a different perception and in a medical MRI, for instance. But what happens if a machine does it? Is it going to be tolerated or not? And I like to think about it when I think about the autonomous cars every single day we have it is maybe thousands of patients being heard in car accidents, then somehow reported in the TV or in the newspapers or something like this.
00;36;32;10 - 00;37;09;24
But if we have one accident, we should one of these autonomous cars that does not injury anybody, It is reported everywhere. So I think about how we make these how we. So yeah how is our our room in this because objectively the accident that I caused and I checked it this is that this is the accidents that are caused by autonomous car they are much less severe than those that are caused by human beings in a million driven basis analysis.
00;37;09;27 - 00;37;34;28
But you have to really press them if the accidents that are caused by autonomous car, how are you going to act? If there is things that to happen? It's unavoidable. When one of these machine tools make a mistake that this even if we prove mathematically that it's better, it makes less mistakes than if a doctor see an MRI or a mammogram.
00;37;34;28 - 00;37;58;07
So basically, our our ability to tolerate errors that will have a huge role in the adoption of A.I.. Yeah, we're we're very forgiving of humans and not forgiving at all of technology and and machines. So, yeah, we it's on us to be open to adoption. All of that sounds great. And thanks again so much for being our guest today.
00;37;58;12 - 00;38;22;17
Otavio Again, cancer is something that's touched almost everyone's life in some way and we're excited for any advancements we can expect. And it's diagnosis and treatment. If our listeners want to learn more about Oracle's initiative or to get back in touch with you, is there a way for them to do that? Yeah, they can reach us out at Oracle dot com and Oracle has a lot of initiatives in.
00;38;22;20 - 00;39;10;11
Yeah, in health care we have a huge team here studying, especially the integration of Iris here. We do have this data. Great. Got it. Well thanks again, Octavio. And to our listeners, we don't want you to miss any episodes of research and action, so please subscribe and if you want to learn more about how Oracle can accelerate your own life sciences research, you can just go to Oracle dot com slash life dash sciences and we'll see you next time.