AI in Biotechnology: Applications, Challenges, and Future Implications

AI empowers biologists with tools like AlphaFold and PyTorch for protein analysis and model building.

The basic concepts in machine learning are supervised learning, unsupervised learning, deep learning, and reinforcement learning. Reinforcement learning is something I'm really interested in; I really like it. Supervised learning uses labeled data—things that

we know for a fact are that thing. So, if we're trying to classify some images, like pictures, and we show a picture of a cat, we know that's a cat because we labeled that picture as a cat, as opposed to something else. So we

say 'cat' or 'not a cat,' 'dog' or 'not a dog,' that type of thing. In unsupervised learning, we just show a picture of, say, dogs; we don't label them. We just show a bunch of pictures of different things or different objects, and it's able to

distinguish and group them so that all the cat pictures would be together, all the dog pictures would be together, all the tree pictures would be together. It wouldn't know what it is; it would just

say, 'This is different.' Deep learning is similar to those things, but it handles much larger datasets. You've probably heard of neural networks; that might ring a bell for some of you. And Transformers—maybe you guys have heard of

Transformers? Yes, ChatGPT. Yes, okay. Everyone uses that. Okay. And reinforcement learning—that's Q-learning and Markov decision processes. Does anybody know what an MDP is? Okay. And so, what are some roles in biology or AI in bio-U biology? So

It, we can learn things about our AI in genomics and proteomics. Anybody study that? No? Drug discovery. Oh, alright. Um, systems biology. Some of these I don't even know. Um, healthcare and diagnostics. Okay. Uh, synthetic biology. All right. Uh, ecology and

evolution. Uh, microbiome research. Yeah. Okay. Uh, image and signal processing. Okay. And the literature, money...and this is more from like, um, the large language models, and there's a mountain of literature out there and it's virtually impossible to read all of that

and make sense of that. So large language models really help out with that. If you add in a way...who uses ChatGPT? Okay. And that's for research purposes. Okay. Write it in Korean or in English?  Bri...bri...okay. Oh well, does that perform...

think, in your opinion, does it do well? Do you get good answers or...little bit good answers? Okay. And I'm going to the AI in this...per...does anybody know these people? Middle person? Anybody know the...the first one? Nobody knows the first

one. Just talked about him. Okay. And second guy doesn't look familiar. Third person. Okay. So the first person is Sam Altman and um, OpenAI. He made ChatGPT, so you can thank him for doing your homework. And the second guy, uh, he's from

Nvidia, um, CEO of Nvidia. Um, you can thank him for all the uh, GPUs out there. And the last person is Demis Hassabis. And does anybody know...forer...nobody knows. It's the um...this large...it's a...

learn a lot of different classes, get certifications, and things like that. She founded that with Andrew D, also a Stanford professor in computer science, but he moved to the bio field and started Citra, an AI-driven drug discovery platform. You might have heard of Recursion and Exscientia, but they merged, and also another AI German drug company...

...where else is AI? AI is pretty much everywhere. Does anybody know what this is? Has anybody seen Boston Dynamics? Okay, Boston Dynamics, that's Atlas, very advanced robotics. AI is used throughout that software.

I'm sure some of you have an iPhone, and AI is also in Siri; a lot of applications there deal with computer vision and generative AI. If you ever see some of those videos that AI generated, videos or images or text...

...Instagram also has AI, making recommendations and showing you different things in your feed that you may like. Everybody has Instagram? Everybody, yeah, Instagram. Okay, I used to. This last one is Tesla, and this is their FSD, full self-driving platform, that they use in their cars to identify different objects on the road and navigate.

Just recently in the news, we have new quantum chips; quantum computers are up and coming, so the implications are that we can have quantum chemistry, which might be very helpful for you guys. Anybody a chemist in here? No chemists? Okay. And so drug...

Discovery of material design, advanced ML, and possibly even AGI, which is artificial general intelligence.  And they're like super smart, and hopefully they don't take over the world. What can you guys use for, as biologists or bio...

informaticians? Um, anyone use AlphaFold? What do you use AlphaFold for? For prediction of the non-existent protein data bank. So we want to analyze the protein-protein interaction, but if there are some pure structures we can...

use the AlphaFold and analyze. But if there are no PDB crystal structures or cryo-EM structures, we have to predict with AlphaFold.  Oh, okay. And how often do you use that? How do you... almost every day? Oh, almost every day.

Wow, okay. And PyTorch—has anyone ever used PyTorch? Uh, let me go back a second. Has anybody used Python, the programming language? Nobody in here? Okay, good to know. Um, so PyTorch is a library in Python which makes it easier to deploy machine...

learning models, and it's a very powerful tool, very easy to use. And Python is wonderful at doing that. Um, so I suggest you guys learn Python, because it's one of the most human-readable programming languages out there. And what I mean by that is it doesn't look like...

a bunch of gibberish, you know? It's the English language, and you should be able to kind of understand that. And the way it is structured is also very understandable compared to, like, Java or C, which is really, really hard even for...

One other tool is Nvidia's BioNeMo, a resource for pharmaceutical companies and research institutions.  There, you can find many resources, including AlphaFold2.

And some other protein resources; I don't know the specifics, but they have a lot to explore. Hugging Face is also very popular with ML engineers. You can get large language models, already pre-trained,

which you can use and deploy quickly. I'll show you that in a bit. I'll pause now—any questions before I continue? Okay, I'll go through this mini-project I prepared,

since nobody here has used Python before, I'll go slowly. One thing is downloading data. You may know that you can go to the AlphaFold website and download it that way,

or you can do it programmatically with Python. You can download hundreds or thousands of different proteins in minutes.  Just run it, grab a coffee, come back 20 minutes later, and everything's done.

I'll go through an example of that.  RFAM—anyone know RFAM?  Let's see... So, RFAM is a database where you can search for sequences. They also have an FTP site you can use to download them.

Just download their entire database. They also have an API you can download, which is this. This is their API; okay, so this is Python code, and this is what it looks like. You make a variable; you make a request.

And you can see this is the website you're going to, and this is the name of the sequence or protein, or whatever that is.  This end part you don't need to worry about, but this is the most important part—the thing

that you're trying to download.  This is also important. So you download just one line of code and transform it into something more readable because, when you get it, it looks very messy, so you might not

be able to understand it. You have to transform it to a JSON object, which is something different and more readable. I'll skip that part because I need to connect to the

internet, and I'm not on the internet, and I'm not sure if that works. So I'll skip that next part and go into... So there are basically three different processes when you're making a machine learning model, and I kind of

skipped over that. One is to download the data, obtain the data, or create some kind of data pipeline. You can do that manually or through an API, which I just showed you—just that one line of code, and you get the data and you

You can get a lot that way. The next part is where most of your time will be spent: the preprocessing phase. That's where you clean up the data, making it usable. You're filling in missing data, converting labeled data, and deleting data that's not so useful. Then you finally do the training and compare different models.

When I say compare models, I mean different algorithms used to make predictions. You might use a decision tree to come up with a model, a support vector machine, or a neural network, and you want to see which performs best.

You want to compare those different models and choose the best-performing one.  How do you compare them? You split the dataset into a training set and a testing set, compare the models, and choose one.

Here's an example of some data. I'm using housing data because it's easy to explain and captures the challenges of getting data and doing something meaningful with it. This is about houses: we have the size of the house, the number of rooms, the address, and the price.

I prepared this myself; it's not real data. Does anyone see anything wrong with this data?  There's something wrong in the 'size' column...

Oh, this was actually just the index; you can ignore that.

But um, just this right here, we have the values 30, 110, 115, and 113. Is there something wrong there, or do you see something that is potentially wrong? And if you look at the size or the price, it's like 500 million won. Okay, so somebody might have made a

mistake here. They might have just put 'won' instead of—should it be like meters squared? So you'll run into some of that kind of uh, mistakes in data, and you have to find these outliers and you have to fix them. Another thing is the number of rooms we

have, n, which is—there's nothing there. Um, there's—so there's missing data. We don't know how many rooms are in there. How do we handle that? And also some missing data here and some missing data here. Um, so what do we—what would we do

for number three? What would your recommendation be for number three? Anybody have any ideas what you would do? Does anybody do statistical analysis in this room? You guys look at numbers and tables and statistics? Great. What would you normally do in this case? Delete? Yeah, because we would delete

it, uh, because it doesn't—it's not very meaningful. Um, we don't know where it is. Um, the location of a house uh, has big influence on the price, um, and it doesn't even have a price, so that makes it even harder. So

this data is useless and we would just delete that; we get rid of it. And for this one, uh, any ideas what you would do for that one? You might just copy it and put a '2' there from the previous one, or a '3' from the um, from the

Next, we call that a backfill, so we say 'fill' or 'forward fill', and we fill it from the previous one, which I did. I dropped the third one, which is here because it's

useless. And then I came up with this: forward fill fills in any blanks that occur from the previous one. So this is 222, and now we have something a little more meaningful.  The next part is we need to do

something about this address because that doesn't really tell us a lot of information. And this is really the important part of machine learning: you take the information you have and make it more meaningful. Also, it has to be a number,

because these machine learning algorithms don't deal with text; they deal with purely numbers. Everything is just a number to a computer. So what do we do? We can take the address and break it down into several different categories and things that

might influence the price, depending on the location. Is it in a metropolitan city? We would say zero or one, yes or no. So this one is in Desan—it's actually here—and we put 'yes'. This one is in Okan; no, it's not in a big

metropolitan city, so we put zero. Also, one for Seoul. We have distance to elementary school (the closest elementary school is 23 meters away), this one's 750; distance to middle school; distance to high school; and that might have some influence on the price.

of a house, so we would want to include that into this data. A number of interesting things with pin walking distance, and that's something you can kind of create yourself, right? What is walking distance? You might say things like, "like 50 meters away from your house", or

100 meters, uh, before you would decide to drive to that location, right? And that's totally up to you, and that's kind of the beauty of modeling this data—that you can be free to kind of manipulate that in your...or what is a good idea,

right? And you can also come up with a lot more other things. So these things are not explicitly in this data; you have to go out and look, like, search for this information and glean that information just from this address. So you might do that when you're

working with biological data. So you might get like your data set, but you would have to go to outside resources to make further data and include it into your training...train and compare models. Okay, so um, I have that here, and it's pretty easy. You

can download, reprocess, processing...oh, train, compare. Okay, so the data set is pretty short; I just made this randomly. Uh, there's—I'm not sure you can see it—but there's a thousand different data points on this, so it's going to go

pretty fast. If you notice, this is called LazyRegressor, a very useful library, and I totally recommend that because you can train different models with just this one line of code right here, and it'll go through...actually, I'll run it and you'll see how fast this

Here's the results at the bottom, and it's done. So, what did we do here? These are all the machine learning models: Orthogonal Matching Pursuit, Orthogonal Matching Pursuit CV, the Lasso, L1, and all these different Gradient Boosting Regressors. You'll notice some linear regression – maybe you guys know that one – and, since it's only a thousand data points, it took less than a second, all these.

And so it does give a score; this is the highest score, 68% accurate.  And that's for the training data. Something to note, and pretty important when you're working in academics, or yeah, especially academics, you want to use this random seed.

A lot of machine learning models are built on randomness, and I'm not going to go into why that is, but just know this makes things reproducible. So if you are getting some random numbers, it will be the same if you use this seed. And if I gave you this code and you use this random seed, you'll also get exactly the same results that I did.

So when you're publishing a paper or something like that, it's highly important that you list that and mention that and use that. It could be any number, by the way; it could be 101, it could be one or zero or 3,000; it doesn't matter. You just should mention it and be consistent.

And yeah, I'll go through the code real quick. This is just getting the CSV file. Anybody know what a CSV file is?

Right, an Excel file, and we load that into a data frame. I won't go too much into that, but a data frame makes things a lot faster than a traditional list or something like that. I want to predict the price, so we say that's the Y variable, and

everything else is the X variable. So X1 is like the size, X2 is the number of bedrooms, X3 is something else, and whatnot. All those things play a role in the output, which is the price. To get an accurate measurement on machine learning

models, we need to split the training set. So you have a thousand rows of data; you want to use maybe, you know, 70% of that and use that for training, and the other 30% you want to test. This function right

here splits it up in a better way instead of taking the first 300 and using the last 700. It randomly picks the 30% or 70%, and gets that data. That's how it's split

up. This next part is the lazy regressor, but I'm going to comment that out. Okay, so I found out that the linear regression model is one of the highest-scoring ones in the test set, so I'm going to use that one from this

list. So we find that it's one of the top-performing models, so we choose that one and we test it. We print off the scores, which I'll do again this way, and we got 68%, which is pretty good. Then we save it into a

Uh, what we call it a pickle? I don't know why it's called a pickle, but some computer scientists thought it would be funny to call it a pickle. And um, anyways, it's called a pickle. It's saved into there and you can read it again.

Take out the pickle and with just a few lines of code, you load the model from the pickle and um, just pick some random variables. So I'm going to pick a house that has um 70 square meters, one bedroom, uh, no pool, and an area

as oh, elementary school, so about a kilometer away from an elementary school, 1.3 km from a middle school, and 1.2 km away from a high school. And it's uh, has two interesting things around the place. And what do we get? We get a house that's

about $80,000. I don't know how accurate that is, but uh, that's what we predicted. Any questions so far about how to deploy a machine learning model? Does it look easy? No, it's only a few lines of code and um, like all this extra stuff is just

some things I just put in there just for um, teaching purposes. Uh, but uh, some things like this is just one line right here and one line right here, and you print out the results. Um, so it doesn't take too much. And if you look online, even ask ChatGPT to write you some

code, and they'll probably give you something like this. And um, so if you guys are interested in doing those kinds of things, um, uh, don't be intimidated by it. Um, it's fairly easy to understand and there's really not much to it. You don't need

To know the core details of the algorithms being used—that's what I do. I have a deep understanding of the algorithms and know how to do some parameter tuning. But for simple things, to get it out there, it doesn't take too much work.  Yeah, and here are the results; I actually got better results than others. Okay. Oh, um, I will skip this actually... no, I'll just go to the last one.

This one's not so important, but this last one is about biases in your data. There are all sorts of biases in data, and it's really hard to mitigate some of them. So you'll get—especially for generative AI—you can't just suddenly make this new compound out of thin air; that's very difficult.

Because you're training on data that's already there—somebody already made that compound—so it's going to be biased towards those things that already exist. So those are things you have to watch out for with your results when you're doing these machine learning models. There are different strategies, like normalizing data and making some kind of function to make that less important. And yeah, that's kind of all I have to say.

Is there any questions and answers? Wonderful! Oh, thank you. I was curious when you mentioned ChatGPT. Yes, so what do you think of when you do a conversation in English? What do you think of—is it very important, is it important when you have a conversation typically in English?

Or Korean? Which language is better for getting more information? I mean, because the church is based on the Creator, the speaking English native speakers, we, as Koreans, so do we get more information as to the English question? Uh, I think Korean is pretty

good. Um, I don't know many different languages, but we are not... we are not speaking... I'm not sure how they interpret that in, um, ChatGPT. I'm not sure if they first translate it into English and then they find the answer and then

retranslate back into Korean. Um, that's another thing, but I imagine English being better because the ChatGPT... I mean, the chatbots, they don't have to train themselves; they already have English language. Yes, other languages, so they somebody has to train

the language. Train, yeah. So how that works is they... um, I'm sure some of the underlying technology, especially for Google Translate, they use these algorithms called word-to-vec, word2vec, and so they vectorize these words into a long list of numbers and they correlate that

with other words that are related to that word, and that's also a vector. So if you look at a graph, an X and Y graph, and you have the word like 'mud' kind of titles, and you have the word 'mud', and it's going off in this vector,

water might be close to that other vector because water is associated with mud, and so is dirt. But if you said 'fire', that vector would be like somewhere off in a different direction. So for ChatGPT to kind of understand those um different languages, I think... I'm not sure, but it

We would have those kinds of calculations first, and then into English. I think, yeah, your question... um, when you talk in English, I mean, the official language... it's okay. But you know, Korean and Japanese, many different speakers... in conversation, but official conversation is okay, but open...

conversation, they still understand what they... what it means. Oh, yeah, yeah. Um, ChatGPT is really good at those nuances in spoken language. So if I just say, "Hey, what's up? How you doing?" you would understand that; they would understand to an extent, like, oh...

they're asking about my physical well-being. So you think... I mean, you're asking the question in English, you can get more detailed information? I would guess so. Uh, I would... because there are a few reasons for that, and I think one of the biggest reasons

would be that there's a lot more English literature out there than in any other language. Korean is just... Korean is only in Korea, and there's... if you look at, like, Wikipedia, uh, most of that is in English. And I'm sure that they used

information from Wikipedia to train their model. Um, they're just OpenAI; they go to different websites, they scrape the web and get all this data, and a lot of that exists in English. So I would imagine the answers that you would get from ChatGPT would be better than any

other... I have a question. Yeah, so you are a digital person. In my case, I'm an analog person. Okay, so in your case, do you want to suggest to CH or students? In my case, I'm native, so... um, yeah, there are a couple of different things. Um, so if you're working in a

Company? I would say no, um, because of security reasons. Um, because you're essentially training, you're giving that information to another company. OpenAI is a company; it's not like, um, you know, like a crib or an academic institution. It is a company that

makes... but, but, but they're not supposed to um, disclose their private information. They... us... but your... it is their policy, but they also take that information from you. And I could see like if you um, ask something about, uh, if you had a question about some um,

biomaterial or something, it's got to get information that's already published, so if... or from their, uh, all the data that they've been collecting. So in theory, I could ask in such a way that um, I could get information from what you put in previously, maybe a year ago.

It's possible. Um, so it's not... they don't publicly disclose that information, but if somebody... when they retrain their model again, um, which they do from time to time, but you know that they don't do it too often, maybe like four or five times

a year, they retrain their model, um, because it takes like two to three weeks to, uh, make those kinds of accommodations, even with all the resources that they have. Um, but they will have that information and it will be in there. And if I was to ask in a certain way, maybe I can get

20241212 101137

김민우

Introduction to Machine Learning Concepts

Overview of supervised, unsupervised, deep, and reinforcement learning.

Supervised learning uses labeled data for classification (e.g., cat vs. not cat).

Unsupervised learning groups unlabeled data based on similarities.

Deep learning handles large datasets using neural networks and transformers.

Reinforcement learning includes Q-learning and Markov decision processes.

AI applications in genomics, proteomics, drug discovery, systems biology, healthcare, diagnostics, synthetic biology, ecology, evolution, and microbiome research.

Large language models assist in literature analysis.

Examples of AI applications in drug discovery companies (Citra, Recursion, Exscientia).

AI's presence in robotics (Boston Dynamics), personal assistants (Siri), social media (Instagram), and autonomous vehicles (Tesla).

Emerging Technologies and Future Implications

Quantum computing's potential impact on quantum chemistry and drug discovery.

Use of AlphaFold for protein structure prediction.

Introduction to PyTorch and Python programming language.

Nvidia's BioNeMo and Hugging Face as resources for machine learning.

Practical Example: Machine Learning Model Development

Data acquisition methods (AlphaFold website, programmatic download).

Using RFAM database and API for sequence data.

Three processes in machine learning model development: data acquisition, preprocessing, and model training/comparison.

Example using housing data to illustrate data preprocessing challenges (outliers, missing data).

Using LazyRegressor for efficient model training and comparison.

Deploying the trained model to make predictions.

Discussion of biases in data and their impact on model results.

Discussion about the influence of language (English vs. Korean) on ChatGPT's responses.

Advice and Discussion on AI in the Workplace and Academia

Discussion regarding the ethical considerations of using ChatGPT in professional settings.

Discussion about the speaker's work in analyzing protein structures and binding.

Challenges in applying AI to biotechnology, focusing on terminology, data interpretation, and convincing experienced researchers.

Discussion about job security in data science in the face of AI advancements.