Listen: Designing Anticancer Drugs with Reinforcement Learning
"Having a background in cognitive science and computational neuroscience and so I've been like focusing focusing on brain research for my pastor five years of education and now recently I've been doing more work on computational tation assistance biology and specifically on cancer and cancer trying to understand mechanisms of how cancer work and how we can find new treatments against cancer specifically quickly and in this work. I've been using mostly deep learning techniques and this will be part of like my presentation here at this conference. It's also and so yeah. So how do those things go together. So I think like many people think it's in a way weird if you come from brain scientists and then you go into machine running right and this is something I would say. It's like it's a very obvious thing to do in a way because if you look back into the history of machine she learning where it all came from like McCulloch and Pitts the first artificial neuron and then a few years later Frank Rosen ballot the perception. And so these were all computational neuroscientists and they were in the end really trying to understand how the brain works and they basically develop the The fundamental of the field of machine learning and so at some point this community and in a way it split up into groups and one group was more trying to and actually understanding the brain works and the other group was more interested in solving the problems. Right right and from this from this community. The machine learning learning community evolved into but whereas computation neuroscience right. Now it's it. It's still a field. It's still out there. It's has been separating more and more from the machine community what's there and originally it has been one like one big community. Yeah and so therefore I think it's quite natural to to have the process. Yeah Yeah you know I think Particularly here at Noor ups I have opportunity to speak with many folks that are kind of working on on that edge of cognitive sciences brain sciences and both using that to inform the way we think about machine learning using machine learning to validate you know some of the biological theories it was maybe more novel is coming from Cognitive Science and brain science and applying machine gene learning to developing cancer pharmaceuticals out in that. Come about yeah. How did that come about a good question? So like if you look at brain scientists this really this problem of seeing the brain which is arguably the most complex thing we have in the universe and and seeing like observing this brain and trying to understand his brain from at different scales at different spatial scales so to speak. So you can think about about the brain in the very abstract and cognitive ways thinking about cognitive phenomena like language and memory those things and you can think about it more from from neural perspective like how do act like what is the most fundamental unit of information processing. How do these units interact? How does information arise? And so like these two fundamentally different approaches and so I like in the first three years of my studies focused on cognitive science which has more top down approach unlike thinking from the big concepts and then down towards the implementation level whereas competition neuro science. They have more like the spot. Him Perspective They in the end and they're trying to solve the same problems but they start first with the basic building blocks like having a biologically plausible neural network model will that imitates basic behavior of neurons. And then they try to scale it up in order to understand more complex cognitive phenomenon and so like these field they really deep. They help each other and they need to work together in order to better understand how the brain works and so after after Android area defeating. Okay I need something a more solid and I really wanted to have this bottom up perspective from competition neuroscience which then I got my masters and so afterwards I I I mean I have to say that I was keen to explore and applications of machine learning because while studying the brain I got really interested more and more into the whole field of data signs and machine learning but and I wanted to apply those techniques but at the same time I wanted to I wanted to still somehow how work with the human body and with humans in general so this is how you how I came about him doing cancer Consume drunk modeling and so the poster is titled Pacman. Tell us about yeah Eh. So pacman is a frame. I mean it's an acronym so spelled with a double double and so it's an acronym. We came up during in my like about a year ago. During my master's thesis for prediction of anticancer compound sensitivity with multimodal attention based neural networks. And and so like when my supervisor came about with this acronym one of very long nights we spend in the lab. We like okay. There's no discussion. This is GonNa be the name for the project. Ah So quite funny how this came about so and we what we're doing in this work at that was the first step step off of the project and presenting at the conference. We were trying to basically forecast the effect the inhibitory effect of emol against a specific type of cancer and so we are treating this problem of predicting cancer drug sensitivity. Not really as the property of a pair and the pair is con- like composed of Itself the chemical the drug that you give to the patient and then the particular to more sell that you want to target because cancer is really like A. It's a family of diseases and the SORTA verse I. I mean there has probably never been two types of cancer that have been exactly alike because the Medicaid of mutations you have they vary like hillbilly inbetween of every individual patients. So it's really unfeasible to try to investigate whether molecule has some onto cancer effects in general. So you really need to treat this problem as the property of pair. So is this drug like hesitant. inhibitory effect against this specific type of cancer patient individually one of the questions. That comes up I is one of the techniques. You're applying here reinforcement learning. How does that play into Into achieving that goal so it comes about in the second step first that was really just trying to predict the sensitivity so the efficacy of Audrey and so what we what we did in consecutive step after we had built this model what we asked ourselves was like. Wow wouldn't it be amazing to have a model that can generate rate new drugs at can like come up and propose new anti-cancer candidate rex. Because in the old pharmaceutical industry there's a huge uh-huh productivity decline in the last few decades and the estimated costs that you have pulled new truck there Estimated to be two three billion Indian USD and most of these drugs that are like FDA approved and approved on the market. So they're really specific only for like very few types of diseases sort of even one disease only so the cost in our indeed that go are like spent in this business. It's just huge and and so we I mean we came up with this framework reinforcement. Learning is really core component. Where we're trying to design anti-cancer cancer drugs specifically for individual patients or groups of patients so we tried to envision the precision medicine perspective here where we're really We're not trying to generically. Come up with new cat. anti-cancer candidate drugs. But we try to like in the design process itself. Both we try to tailor the Monaco the drug specifically to the need of the patient himself or herself and so forth for this framework we use. We're using reinforcement Okay you also mentioned in the title of the poster transcript domain data. What is transcript Tomac data? You're right so you can think about transplant. Tomic data as basically The the expression of every single gene that you have in your body like you do you know about the human genome and so part of the human genome and code for specific proteins and these expression of these proteins. You can measure in the cell. That's different techniques techniques to do that so the most commonly used technique and the technique that was used to measure the data we work with is called are on a sequencing thing. Data we are you measure basically the M. A. Snippets in the cell. And so from this. You can infer basically which genes are expressed to what extent so so you end up if you if you do the sequencing step you end up with a vector of about twenty thousand genes and for each gene you would have an expression value view. This is usually just an integer. Like how many times did you find these Slip it in the sample. And then so this this vector Tori you can really think of it as like a fingerprint of the cell. So it's like it's a proper characterization of the cell there's different types of of comics data. So this is true. Tomic's data right. There's like also genomics data which directly directly measuring gene data and there's also also appropriate mix data actively measuring the the proteins"