Sramana Mitra: What is the dataset? Across all these categories, how many drugs and treatments are we talking about that you have modeled and have data for?
Sungwon Lim: Our first service which we commercialized last year is for dog blood cancer. What we are trying to do is help doctors make decisions. The 16 drugs that are used for lymphoma and leukemia are being tested on our platform. We collect the real-world actual treatment cases and provide our reports today.
After three or four months, we approach the clinic and ask for their treatment decision and medical records. They provide us with the patient and medical charts which contains an extensive amount of information. We compare that. These are the real-world cases we are training our machine learning model with. So far, we have collected around 2,000 lymphoma cases. These are the training sets for us.
Sramana Mitra: How is the model set up?
Sungwon Lim: Drug testing data cannot be the input variable by itself. People actually did these kinds of things – tests outside of the body. If the cells die outside of the body, they just predict that this will work in the body. It was a very hot topic in the 1980s. It was a mixed result. This particular test is not being reimbursed by the insurance company.
What we need to do is take the systematic data like patient information and metrics which measure what kind of proteins or antigens are expressed on the cancer cell surface. We collect as much data as possible from the cancer cells and the patient. These are all the features of our machine learning model. For each drug, we develop our AI model. We have developed 13 AI models for 13 drugs. We start with correlation and then reduce to 10 to 15 features for each model. Each drug model has different input models.
Sramana Mitra: How is this going to play out as you get more drugs and patient data? What happens when two drugs interact? How are you going to scale this thing?
Sungwon Lim: People are very curious, especially pharmaceutical companies. They are interested in this synergistic effect. For lymphoma in both species, CHOP is one of the first lines of chemotherapy. For humans, there is rituximab. RCHOP is the gold standard, but it doesn’t work well for dogs. Dogs just use CHOP. These are the four-drug combinations. We built our model for this cocktail.
We first thought that we need to physically test four different drug mixtures on the plates. We probably need to test all the different permutations. It turns out that we didn’t need that. Each model for C, H, O, and P shows that the combination of all of these resulted in a good prediction model. We only need a single drug response model. The actual patient’s clinical outcome from this CHOP therapy is the most important part.
Sramana Mitra: The end game here is take the cells and use your models to figure out which combination of models is going to be the best?
Sungwon Lim: First of all, we cannot recommend the drugs. We recommend the drugs to the doctors, but the doctor should be the final decision-maker. We provide lab tests. What we are providing is the single-drug response. What is the likelihood of tumor size reduction when using this particular drug? We rank the drug in that order.
Sramana Mitra: Very cool. How far along are you? Do you have enough data to be able to predict?
Sungwon Lim: Yes, that’s why we started our commercial service in the veterinarian market for dog lymphoma and leukemia. We are rolling out our service for cats later this year.
This segment is part 2 in the series : Thought Leaders in Artificial Intelligence: ImpriMed CEO Sungwon Lim
1 2 3