Material scientist, Sudipta Seal, computer scientist, Ozlem Garibay, and PhD student, Mehdi Yazdani-Jahromi, all at the University of Central Florida, talk about artificial intelligence, or AI, changing the way new drugs are discovered and develop.
Interview conducted by Ivanhoe Broadcast News in 2023.
Is AI is changing the medical world by storm?
Seal: Absolutely. And as an experiment list, you need so much resources and it takes a lot of time to do an experiment. So if we can get advantage of the artificial intelligence, and machine learning to get this data, to give us some predictive modality, then we can cut down the experiments as well as cut down the time. And that’s one of the seals see it. During the COVID pandemic the vaccine came out in less than a year’s time frame. And I’m sure the computational think played a huge role on this as well.
How much time is spent on failures of developing drugs?
Seal: Of course from failures, we always learn something. But the time it fails, it can move the clock backwards. At the same time for experimentalist, you need lot of resources and dollars and not all the time, you actually have this things. And that’s what the computational approaches to solving a problem sometimes is very crucial to really develop the products that you’d like to do.
And your part in this project was the experimental part, right?
Seal: That is correct.
Can you tell me about that?
Seal: So, what happened was, as I was explaining to other people, we’re trying to develop coatings where the virus will be captured and be killed. So the captioning part is little bit different. So in the coating I need something for the virus to be attracted. So there are a lot of drug molecules out there. But we don’t know which drug molecules are there. So we identify some of them. But these are very expensive for us. It takes long time to work on this experiment. So we approached Dr. Rosenly, various team to look into that aspects and tell us out of hundreds and hundreds of molecules and the data, can you give us some indication? Narrow down from a couple of hundreds to only four or five of them. And then we’ll try the experiments and see and bingo, it worked. A couple of them really worked.
How accurately do you remember?
Seal: I would send close to 98 percent with confidence limit. It very accurately. We only developed that these are the molecules if I put on a coating or a glass or somewhere else, they will now capture it, the virus cannot fly out anywhere else, and then coating will have their own way to destroy it.
How quickly did that happen compared to the traditional way of finding?
Seal: This happened, this was the internal grant from seed- seed grant from UCF. This happened in, I would say less than six months.
How long would it take to find out, traditionally, if a drug will work than with AI?
Seal: I think we’ve traditionally router couple of years, and it will take lots and lots of resources. This one was done less than six months time-frame, so it was very quick. I would say it’s.
Do you think it’s changing your life, your students lives?
Seal: I think so. I think this is also gives us a flavor of interdisciplinary research. We are not computational people, but working with computational people, and computational people working with us, as an engineer, we can define the problem, we can solve the problem, and we can speed the problem to solving will be much easier.
How does AI do change the medical field?
Garibay: So, for the data there are different ways that AI can help with that. The approach that we are working on basically we’re trying to model drug and the target protein interactions and find the predict their interactions and they fill in some high level of accuracy. This is the way that our approach works, but there are so many different ways that AI can help speed up drug discovery.
Would you say the protein interactions is one of the biggest hurdles for drugs to either succeed or fail?
Garibay: It’s definitely very important because our inner bodies are made up of proteins. And I’ve been out there for identifying the drugs that is currently interact well with the protein molecules. It’s going to help us not only identifying, find, and deliver the drugs that is needed but also it can help prevent some of those the- But we don’t want it to interact as well. So very targeted approach to delivering drugs to where it’s needed.
So, can you say through AI and through what you’re studying, this drug is going to work and it’s not going to have these bad side effects?
Garibay: So, the AI can help it. That- We can’t say that the AI is going to interact with this protein. And then we can also have different modeling approach that can also help it. What are the things that are less or more favorable with that interaction. Yes.
So, what made it now? Why now? Why can you do this now and you couldn’t do it four years ago?
Garibay: So, I think the AI methodologies has advanced through the years. And also of course, the amount of data that is required to do some of this research is growing tremendously. And the computational resources are growing tremendously as well. So as computetional resources and the data become more and more available, we have opportunities to make this print out models for more accurately. And I think also there is quite a bit of interests from both researchers side and from industry and from government to make this work. And there’s incentives- financial incentives. And people- So I think all these things can come together, make it more available for researchers to do better things or more better models.
Can you explain how your model works?
Garibay: Our model, so we doubled up the pipeline, the experimental and computational pipeline where we extract the information from the proteins, and from the drug and then we observe their interaction. But the reason that our model building that is works for values that we draw insights from another field. An AI, an artificial intelligence, which is called natural language processing or NLP in short. Where the sentences, the model to extract information about the relationship between the meaning and the sequences of the sentence. So using this knowledge, we thought that we could do a similar thing. We could apply NLP methodologies basically treating the protein, and drug complexes as a sentence sequence, and with the relational information between these two pairs. And to see if we can- if we can identify what might affect the drug interacts with both part of the protein complex because protein is a large molecule and it has so many different sites that it can- it can interact with drug. And drug technically is very small but then we can also identify not only the site, but also we can identify how that relates and how- how they interact. And it can later on actually allow the chemists to understand how this relationship or how these interactions happen. Because we are seeing in more molecular level, this interaction.
How many things are you testing through AI?
Garibay: So, many different things that actually the drug coming to the model itself, extract that. What we do is we enter the drug sequence and then we also enter the protein structure. And the model itself discovers the properties and what properties are more salient and draw their graphical representation and then observe their interaction. But it’s an end-to-end model basically to try to predict the features, extract those features only by itself automatically. So more of the details probably they can explain a little bit more how to model works.
Is there any reason why anybody would start a testing a drug without running it through AI first now?
Garibay: I would say probably there will be more and more leaning for the AI for this discoveries because it can very quickly detect the drug datasets are large and you can very quickly identify what drugs are best suited. Now you can start looking at so many different drugs. You’re looking at their various small subset. And then experiments you can do them in a more a bit of accuracy. So I think that there’ll be less of just doing the experimental versus having some guidance from AI models and predictions.
So, you guys took like the features of the protein and the drug and then made analogues to sentences the drug, changed it to languages, so then they process at the words that we’re used to describe it, we’re going to change it to a language code from a protein drug code, right?
Garibay: Yes. It’s like the locus the sentence-like structures. So we basically took a protein and identify the binding sites, we call them. Those are the pheno-active active sites of a protein that interacts with the drugs and we identify those and then we create the graph representations because these are two dimensional molecule slide. So the three-dimensionality helps to be more precise. We know how this molecules interact. And then- then we create a sentence-like structure with it. Just like in sentence, the words would be context saying I’ll give you a better meaning to be using the self attention mechanism to identify either for where the drug fits everything that proteins sentence structure that are binding sites represented. With that we were able to improve the accuracy of the model because clearly there are some relational information in that biochemical properties of drug target pair. So basically the model is able to capture using this natural language processing. AI technique allows us to capture those context dependent, I would say, in language terms, properties we- we were able to capture those insights from that representation.
Is there anything I’m missing?
Garibay: No I think that capture- Well, if- if- I mean the other couple of things that I think this model is capable of is that it’s not only highly accurate, but it’s also highly generalizable, which means that sometimes when you create these models, it did separate only protino, certain protein drug pairs. But our model is very generalizable, meaning that it can maintain its high accuracy across different types of protein drug pairs, which is very valuable because you don’t need to customize your model for different direct protein pairs, but it can be applied to any interaction.
How long until we can see this used in pharmaceutical companies?
Garibay: I think our model can be used right away. If any pharmaceutical company wants to take and use it, they can use it right away.
For $120 or $50 million, right?
Garibay: I think we donate it to their very level for a fee, so they can actually realize it.
So, what was your role in this?
Yazdani-Jahromi: I was in charge of developing the model, but computational model and running the experiments for the computational side of this stuff, so that was my main role.
So, how does it work? What makes this different than anything else?
Yazdani-Jahromi: This work is actually based on natural language processing. As you see today in these days, the ChatGPT is very popular that there are very large language models that can answer to your questions and something like that. So this area of study was developed in past 10 years. And we wanted to leverage this area of the study to bring something new to the drug target interactions of that discovery field.
So, is ChatGPT the same type of natural language processing?
Yazdani-Jahromi: So there are very similar in the building block of them, but they are not the same. The building block of all of these models are transformers and attention mechanism. And we wanted to leverage those type of architectures to improve the accuracy of models in drug target interaction or drug discovery.
What did you do? Tell me the process.
Yazdani-Jahromi: The proteins are very, very big molecules. For using these molecules, we wanted to chop them actually or divide them into the pieces that are most important in that protein. The proteins most important parts for drug discovery are the binding sites of those proteins. The binding sites are essentially the docks. It’s like a ship that has a dock and if you want to attach something to this protein, you must attach it to this binding sites.
How many binding sites are there?
Yazdani-Jahromi: There are many binding sites for one protein. For example, one protein can have between 10 binding site to 200 binding sites. But this is not experimentally designed or approach to this. We actually estimated those binding sites in our method. So the thing about this method is that we separated this binding sites from the protein and treat those binding sites as a word in a sentence. So essentially it becomes a natural language. And then not exactly but ChatGPT, you can prompt with the natural binding sites of the protein and the drug and it will tell you if this protein binds to this drug or not. That’s the whole approach that is novel about this work.
What would be an example of the language you want it to turn into?
Yazdani-Jahromi: No, actually, the thing about this is that in ChatGPT also, you have the venue typing something that votes are then turned into some vector of numbers or some collection of numbers that represents that vote. So it’s like that in the sense that now we can put them together, now beside each other, like a sentence, but you cannot see the sentence. You just see the embedding of the words that you have.
If you take a protein and you want to see if it’s going to work, how quickly can you see if this drug is going to be?
Yazdani-Jahromi: So there’s two parts to the AI models. You have a training part and you have the testing part. For the training part, it takes a lot of information, a lot of data to train your model. And for the testing part, it’s just inference is just very quick. So when you train your model, you train it on millions of data points, millions of drugs, millions of proteins. And then you test it on one protein. And that’s very quick. So you can just give it, for example, the COVID protein and test it against all the FDA approved drugs and see whether or not they bind or not. That’s the beauty of this work.
So, is there a really cool part about all of this?
Yazdani-Jahromi: Yes. So this part was the first part of our study. The second part of the study was actually we were trying to see how strong is this binding between drugs or protein. The first part of this study was about whether or not it’s going to bind or not. But the second part of study is how strong it’s going to bind. So we developed an algorithm. Essentially we develop the model that can implement with other computational model and increase their accuracy. So the cool part about this model is that it can get augmented in essentially every, not every, but it can be implemented with other. The cool part of this is that you can implement our model with other computational models. So essentially, if you have another model developed, you can use our model alongside those model and improve their accuracy.
And you said by a lot, but maybe from 40 to 70 percent, right?
Yazdani-Jahromi: Exactly. So in some cases we see up to 20 percent boost in accuracy and that was a very substantial increase in these fields.
What’s the next for this?
Yazdani-Jahromi: The next step actually is that we are going to, hopefully we are going to create a website or create something like ChatGPT that people can put their protein and the drug and see whether or not they’re binding or not. That’s the end goal of this research that everybody can use this model that we developed, that everybody in the industry or in this research field can use this for research.
END OF INTERVIEW
This information is intended for additional research purposes only. It is not to be used as a prescription or advice from Ivanhoe Broadcast News, Inc. or any medical professional interviewed. Ivanhoe Broadcast News, Inc. assumes no responsibility for the depth or accuracy of physician statements. Procedures or medicines apply to different people and medical factors; always consult your physician on medical matters.
If you would like more information, please contact:
Robert Wells
Sign up for a free weekly e-mail on Medical Breakthroughs called First to Know by clicking here
