Data Science for Bioinformatics

Key Takeaways

are bioinformatics and so before we begin let me ask all of you guys have you heard of the term data science and the term bioinformatics okay can you share your understanding of of the two terms what is data science what is bioinformatics okay all right that's good that's a good can we have your viewpoint wait let me summarize what you have just said in one sentence so you said that we're collecting information and then you're using an example of collecting information on patient and then you're gonna analyze the information in order to make a decision for therapeutic uh purpose or application okay very good how about you very good definition yeah so you probably have heard of bioinformatics before and so you're essentially telling us that it's like the merger between computer science and using it for understanding analyzing biological data okay very good and how about you okay very good all right so all of you probably have a good understanding already what bioinformatics is and what you're going to be learning today okay so i'm actually recording this lecture okay would it be okay if i post it on youtube okay yes okay very good okay so this is my favorite quote i will argue that computational thinking and computational methods are so central to the quest of understanding life that today our biology is computational biology because as you have told us as scientists we collect data and we don't collect small amounts of data we collect large amounts of data and because we have robotics and because storing information is now cheap right and we have access to a lot of computational power computational resources to analyze data therefore we could harness all of that in order to find knowledge insights from the data right so it's not only that data science or bioinformatics are using computer to make sense of biological data but both are a service to the broader scientific community right because typically if they don't use data science before they would have to use statistics and before statistics right that would we would just have hypothesis we would just have guess guesses right because if you cannot prove it by experimentation and in order to have a reliable results or finding from the study you need to have larger sample size right you need to have it validated by statistical values like p-value right so data science is becoming a very big thing lately and a lot of industries are making use of data and data science okay so what exactly is bioinformatic so before it might seem impossible you know to make sense of all of the large exponential amounts of data but then it is now possible because we can harness statistics computer science information theory in order to make sense of the big biological data okay so bioinformatics it allows you to understand the molecular basis you know like the the small minute detail of how disease happened and so i provide you an example here that you could identify which gene is responsible for a disease by comparing the gene frequency between two populations one having it and one not having it but then this is a simplistic view okay and i'm sure in systems biology in bioinformatics there's more complicated pathways okay it's not simply the upregulation of genes maybe the genes is already there maybe there's no change in the frequency of the gene but maybe there's other changes in the body metabolite changes and metabolite changes could reroute and influence the production of the gene product which is the protein okay so biology is very complicated and if we don't make use of computer science computational approaches it will be quite difficult to handle the vast amount of data right so as one of you have mentioned already bioinformatics is an area where computers are used to make sense of big biological data and so bioinformatics is growing exponentially after the genome project right in the early 2000s we have the human genome project which at the time we hoped that it would unlock the mystery of life but then it just started to open what he called the pandora's box and it's just the first tip of the iceberg or the first piece of the domino all right and there's so many pieces afterward we have proteome metabolome microbiome interact home glycomic and there's so many other omics data if we could summarize the task that bioinformatics can do we could summarize into four major tasks okay so the first one is search so the most obvious thing is when you have a question you want to get understanding and how do you gain understanding by searching because normally you would google right google for the information that you would like to find out same thing if you have a protein or if you have a gene or a compound you would like to google it but then you google it in specific public database right you have genbank for genes you have chambo for compounds you have uniprot for proteins you have protein data bank for protein structure right there's so many more you have keck database kegg for the biochemical pathway and afterward you want to compare right you have your gene you have your protein and then you need to compare it to other protein in the database or you have a gene and you want to compare it to other genes in a database right and in in terms of the protein sequence or the dna sequence the comparison i think you would have heard of the term blast you have heard of it before and you have heard of the term sequence alignment right right where you compare the sequence from multiple species and they want to see what is the consensus right and then typically you would use it to design your primer right or you would understand okay what what is the the consensus region they're probably conserved for the function to happen and what is the variable region and the variable region could give rise to drug resistance okay if you say it's a microbial protein and it's virulent and number three a very common is to build models right models will allow you to understand complex information because you're simulating it right like for example if you don't have the crystal structure you would build a homology model you would predict the protein structure or if you already have it from experiments like the x-ray crystal structure but then how are you going to visualize it when the protein is so small it's smaller smaller than a micron because bacteria are in the micron size right micrometer and if it's so small okay you can't you cannot even look at it under the microscope okay the proteins so how can you visualize it therefore you need models right so they develop molecular models of proteins or compounds right and models can mean many different things you could have the structure model you could have the prediction model right and you could have so so many many other types of model so model is a generic term which is used to represent a collective a collection of all of the structure and relationship the data types of a data set right like what is the relationship between the chemical structure and the biological activity if you build a model the model will store all of that information right it will store the information about the relationship between the structure which is quantitative through numerical descriptor okay which you will learn in this presentation and number four very important what i've told you the first three search compare model it might occupy your 20 of the time that you use to do bioinformatics task or data science has number four integrate and curate this will take you a lot of time and many people give up during this stage because the data the information will be very dirty there will be a lot of redundancy and at the first time when you have a look at the data it might not look like it could be modeled right so you have to create the data set you have to combine data sets or or information raw data from many sources of database and each database are heterogeneous meaning that if you compare them they will not look the same they will have different columns and even if they have the same column they will be called different things and so the trick is how do you combine it how do you how do you know that compound id from database a molecule id from database b are exactly the same thing right so you have to so a human will have to evaluate that and then normalize information across all of the databases that they want to combine okay and afterward you would need to clean the data right because data are dirty because they might have redundant columns they might have missing values so you have to clean that you have to impute meaning you want to replace the missing value or you want to cut the redundant column out because as you will soon see during model building if you have redundant information your model will be biased yeah so you want to have your independent variable or the x variables which i will show you later on to be independent from one another okay all right so these are the common tasks in bioinformatics and i've summarized it in the form of a infographic okay so same information so data scientists has been called the sexiest job of the 21st century and it was actually the title of an article published in the harvard business review by thomas davenport and dj patel in 2012 okay so about nine years ago and after since the term data scientist has become on the upward trend so everybody wants to become a data scientist and therefore companies are competing to acquire the best talent the best data scientists and therefore they're providing pretty competitive salary right in the us you would make upwards of six figure with the undergraduate degree okay and it's increasing over time but what what is to be loved about data science it's not only the monetary gain that you will acquire but data science allows you to as illustration demonstrate convert data to insights via the data science process okay and so the data science process is actually kind of like an art form there is a setup guideline for you to follow but then it's flexible enough for you to innovate okay so if you have 10 different scientists data scientists analyzing the same information i'm quite confident that all of you okay let's say that you're all data scientists and i give you the same data sets and i'm sure all of you will analyze it in a different way you're gonna use different tools it depends on what tools you are familiar with and you're gonna get different insights out of it and if we can combine the insights the unique insights and different insight from each of you we will have a complete view of the problem so more and more data scientists if they are analyzing the same information we get to see the complete picture because if you imagine if i have an apple and if you analyze the data you will see one face of the apple and you will see another face of the apple and if you could piece it together in all possible rotation we will see the complete picture of the apple because different people will have different view and different perspective okay and that's the beauty of data science okay okay so i told you that bioinformatics could be summarized into four major tasks so data science could be summarized into three major tasks and the first one is exploration so what's the first thing that you do when you acquire a new data you want to understand the data right so don't jump to the conclusion don't jump to the prediction model building process but first start by exploring the data what is the data about how many columns how many rows how many missing data how many missing values do they have what's the data type is it categorical is it numerical right quantitative or qualitative um what is it about what is the category about what is the topic about what is the relationship between each column what is the distribution of the data or the data binary having zeros and one are they continuous floating number like 1.58 98.58 okay so you could do that by using descriptive statistics right where you calculate what is the minimum value of the particular column what is the maximum value of column a what is the standard deviation okay and then you could also compute like what is the intercorrelation between each column how are they interrelated is x1 variable related to x2 is x2 related to y variable okay so you could use intercorrelation or pearson's correlation coefficients and another great way to understand the data is to perform data visualization you would make plots scatter plot bar bar plots pie charts violin plots scatter plot contour plots there's so many right i mean i could list a whole different types of plots for you and you will see that one of the most intriguing thing of data science i would say that makes it an art is the data visualization right as they say a picture is worth a thousand words you could easily spend the entire week to craft one single plot which i've done on some of the papers that we have published can you take you one week or two weeks to make because the first iteration when you make it you have to think about the the reader when they see it oh they understand what are we trying to tell them and then you try to make it as digestible as possible you're trying to simplify the plot as much as possible so it's not just dumping it into microsoft excel spending what 10 seconds to make the plot that you literally spent six months to get the data you're not doing your data justice right you spent half a year making the data through experiments and then you spent 10 seconds in microsoft excel to make the plot you're not doing justice for your data right yeah so as i've said data science is kind of like an arts and therefore it could take you a couple of months to learn but it will take a lifetime to master because data science is so vast okay so there's so many sub areas of data science which is overwhelming and if you jump right in i mean the first day could be like wow this is so fascinating and then you could spend two months three months in and afterward you realize how much information there is to learn about okay and don't get discouraged because discouraged will only stop your further progress into the field okay but just be aware that everyone else are like you right they don't know everything but important thing is you use the information in data science the tool available to make sense of your data that's all that matters if you could add value to your data right that's the most important thing right second task is inference right so inference allows you to use statistical tools to draw conclusion from your data right because you do t-tests you do anova you do inter intercorrelation correlation coefficient you want to look at the trend right so you apply statistics to understand your data to understand whether two variables are different whether they are significantly different via statistical tests in order to add reliability or confidence to your information that you're presenting right you're saying that a is better than b with a p-value of less than 0.05 and so that adds credibility to your data right prediction we see this a lot right now they have self-driving cars they have amazon they have this supermarket where you just walk in grab everything into your bag just walk out you don't even have to check out because they have cameras around the store and they know who you are right they use computer vision they capture your face they compare it into a database they identify your identity you already have a credit card on the record so whatever you grab from the shelf the computer vision visualizes that converts that into information okay maybe you grab an apple you grab coca-cola can and so they mark information down into the database at the moment that you step out of the store right they make the transaction right so very convenient right so prediction can do many things i will show you a list of what predictions can do but there's two phases of it one phase is that it improves our life but another one is the ethical issue the responsibility use of ai right and you probably have seen it from movies such as terminator skynet right and how computer or ai how evil they are in the movies right so that is something that we have to be aware of okay and consider the ethical issues and the responsible use of ai okay because peter parker was told in spider-man that with great power comes great responsibility so ai is very powerful you have to be wise with how you use it right it could be it's a sword it could be used for good or it could be used for evil and you know that they have deep fake i mean they have fake news and sometimes you see celebrities talking but actually they're not because they use deep learning to create a fake version of the person or even they have a voice synthesizer right so this is becoming very scary and therefore ethical use and responsible use of ai is very important okay so this is the overview of the process in data science so you start with the collection of the data and then once you collect the information the information could be a specific column from a particular database for example you have you have hundreds of database you're cherry picking the information you're chopping are you going to the supermarket you're taking items off the shelf you're selecting which variable which feature you want to include in your analysis but before you do that what do you need you need to have a hypothesis you need to have a research question that you want to find out and in order to answer that question you will try to find the information that are relevant to your question and so you are cherry picking columns feature from different database and then you have to normalize it you have to clean it right to make it comparable to make it compatible and then after that you perform exploratory data analysis remember exploring the data you're computing descriptive statistics you're making plus visualization and after that the eda right the exploratory data analysis part will give you a better understanding of your data and now you know that columns one through 500 is relevant maybe maybe relevant for predicting the price of cryptocurrency or important for predicting the deceased outcome of patients are important for predicting the success of graduate students right so you have a series of x that you can use to predict why right the outcome and therefore you do you do model building and then once you get the results you evaluate the model performance and you see if it is sound maybe you could go back to eda again to identify other factor build another set of models maybe you discovered that well you need additional data therefore you go back to data collection get more data clean the data some more perform some additional eda model build and so this is an iterative process you could loop it to whichever component that you need right you could go back to collection you could go back to eda and then it would be iterative finally in all of the information are in place when you're ready to move on the final step is model deployment it is when you want to put it to the market similar to how like for example if you want to release a product to the supermarket you have to do a lot of things you have to conceptualize the idea you have to test the market survey market research does the market demand that create the products optimize the product do the branding put it to the shelf and then once it's out in the shelf what do you do you have to monitor it same thing here you build the model you make it into a application everybody could use your application but underneath it is your model right and so this is the life cycle of the data science model therefore you will see that it's not only data scientists maybe this illustration is misleading actually data scientists could do many of the tasks here but if you take a cross-section data scientists could do maybe a minimal part of data collection but data engineers are more skillful with databases with all of the data infrastructure so you have to work with data engineer so data scientists could know a little bit of data engineering but then in order to collect sophisticated data you will need to talk to the experts the data engineer right so data scientists might know a little bit but then they have to collaborate with the data engineer data scientists might know a little bit about cleaning performing eda but then they have to talk up to the data analysts right because data analysts have more of a business orientation they have more business acumen they could have more domain knowledge and therefore you have to talk to them to get insights from the data in order for you to build the model so you see that it is a team effort and data scientists could then build the model but then they could build a basic model probably but then they have to talk to the machine learning engineer they are more skillful in developing state-of-the-art machine learning workflow right so as you can see data scientists can do a little bit of everything but it also depends on the person as well right some could come from a data engineer background and become a data scientist and it also depends on the company that they're working at are they working at a big company big tech company like facebook amazon if yes then the roles will be more modular but if they're working in a small startup company one person will have to do everything right so it depends on where are you working at okay so you might be wondering what is the skill set that are required to become a data scientist so i created this infographic and i call it the data science landscape and i released it on february 14 of last year 2020 so it shows you that the skill sets that you need to become a data scientist include programming statistics data pre-processing software engineer mathematics data visualization machine learning and also very important is the soft skill so all of this are actually obtained by analyzing the linkedin profile so as you might already know linkedin is a it's kind of like a social network platform similar to facebook but then for working professionals to network to find a job right okay so here are some use cases of data science number one computers like the ibm deep blue have defeated humans in jeopardy and chess and then quite recently they have the alpha goal right and also they have the alpha fold for predicting the protein structure and now the alpha photo2 has are now able to predict with the accuracy rivaling the x-ray crystal structures okay you might know google tesla you're working on a self-driving car nasa uses computer to simulate space mission aircrafts are designed by using computer simulation to look at the aerodynamic supermarkets shopping malls companies they're analyzing customer data to find promotion for you they send you coupons because they know you buy product a and they know that you're also likely to buy product b based on your prior shopping behavior so why don't you use it for for bioinformatics for biology why not use it for improving quality quality of life for designing developing a new drug right so data science and drug discovery because in bioinformatics there's so much sub-discipline so let me give you an example of a subdiscipline in bioinformatics which is drop discovery so number one right data science particularly machine learning will be able to help you to look at the relationship between the structure of the compound and the biological activity right when you get the relationship in the form of a model the model could make a prediction on new data okay because the thing is individual data are very limited it might be difficult to assay it it might be costly and it would probably take a lot of time considering that you might need to have information for millions of compounds available for decision making but then maybe experimentally you already have a couple of thousand but for example if you're working at like a regulation governmental agency where you have to handle let's say the companies are submitting their compound for use but then the government needs to regulate the compound right they first need to determine whether the company is safe or not safe right in order to have assigning them like a material data sheet msds or a for safety toxicity evaluate the toxicity of the data so it's quite impossible to provide justification for all new compound but then if you have a database of known compound you build a model and let's say that your new compound have a similarity to an existing compound in the database the model could make a prediction and say that this new compound is similar to compound y and therefore it is toxic with a prediction probability of 0.98 which is quite high because the structure looks alike right if one compound is known to be toxic and another compound looks almost the same then they're also likely to be toxic as well okay so this concept has been used for discovering a new therapeutic indication for an existing approved drug therefore they call this drug repurposing or drug repositioning when you're teaching an old drug a new trait let's say for example you have an anti-cancer drug you have never tested that it also has antimicrobial activity you never know that let's say that five years down the line someone else tested your anti-cancer drug and they discovered that it also has an effect let's say for against antimicrobial infection therefore they could reposition the anti-cancer drug to be also an antimicrobial drug as well okay so that is called drug repurposing and how can you do that you could do it using machine learning similarity between the compound and the protein which is also applicable to personalized medicine because for example all of you have cytochrome p450s enzyme but then the sequence might look different and therefore you may react differently to the same drug and if we know your sequence we could predict which drug variant will be better for you which will give you less side effects okay so this is something for the future for you to ponder about the possibilities of data science in drug discovery so when you apply data science in drug discovery one form of it is called quantitative structure activity relationship or shortly as qsar so it tries to find mathematical relationship between the chemical structure information and the biological activity so the chemical structure tells you the unique molecular feature of the compound so if you think of a molecule as kind of like a lego block you break down the molecule you get a different building block like a lego you have the different functional group therefore you want to see which functional group does it have does it have a carboxylate group does it have the benzene does it have aldehyde and there's so many there's hundreds of substructure right the the small component they also call it the substructure or the functional group so these are then called molecular fingerprints okay so the molecular fingerprint will be called the local feature because it tells you the the constituents the composition but then at the global feature you have one value molecular weight solubility right if you represent a molecule with one value you call it a global feature like you represent a human being by one value heights weights blood pressure but if you can represent it with many value you're zooming in to the detail one molecule many value therefore each value you call it the local feature right you want to have it as much detailed as possible like for example what is your liver function test you have several other parameters well so collectively all of the biomedical record each of them will be the local feature okay all right so okay so this is a data set when you want to build a model the fundamental unit of data science one of them if you have tabular data it's a data set and so what is it what is tabular data i think all of you know this already it looks like an excel spreadsheet you have the column and you have the rows so in here the rows each row will represent a sample like a molecule molecule a b c d patient a b c d okay so they represent by the rows each column for example the orange column you see x1 x2 x3 these are the input variables input variables because when you build a model you need to have variables that are inputs to the model the data are the input variable synonymous term they are also known as in statistic independent variable because each variable would have to be independent from one another because if you have two variables which have the same correlation coefficient they are similar you need to drop one of them in order to be independent or you could call it the feature you can call it the features another word is you can just call it x you probably have heard y equal f of x if i tell you this you recall back to your mathematics 101 course in undergraduate right you already have foundation for data science okay why variable which is also called the outputs why why do you call it output because when you put in the input variable into the model the model will process the input variable make a prediction and outputs the output variable so they call it the output variable you could also call it the dependent variable you could also call it the class label or you could simply call it the y variable so they are the columns right so your data if you read a paper maybe you call it the m by n matrix just know that m represents the row n represents the column okay so this is a tabular data set right because you want to use the x variables and the y variable to make a prediction model okay if you use both x and y we call it supervised learning if you use only x we call it unsupervised learning i'll show you that just a moment eda okay i mentioned already it allows you to get an understanding of your data sets right by doing descriptive statistic mean value median mode standard deviation by creating data visualization and also by reshaping your data you could you could perform pivoting you know like it might be the values in the column will be split into different columns okay or you could transpose your data from horizontal to vertical or vertical to horizontal okay it seems confusing right like for example if i have one column and inside that column i have the values okay let's say that i have a column called grades i will have a b c d f in there let's say that i want to i want to amplify that i take out the values a b c d f i created five new columns a column a b c d f and if the person has an a i mark it with a one and then the rest will be zero zero zero zero and the person who typically have the value of b i will mark the b column to have a value of one all right so this process called hot encoding hot encoding okay so you're making like a column to be five different columns and then you would have it binary zero and one okay okay data splitting so let's say that you start with the data here in the gray color actually all of the illustrations i've shown here are taken from my medium blog post i'm also a blogger i blog about data science okay and then i also do youtube videos about data science okay so all of these are actually in the medium article if you go to medium search for uh data professor and you will find this okay in one of the article i think this is in the data science process article okay so you have a data set here represented by the gray dots let's say that you use your entire data sets to build a model okay let's go let's go with the reason why you need to split the data okay you could use the entire data to build a model and finish but you want to evaluate whether your model is usable for the future will it be any good for future prediction so what do you do you split your data one subset of the data will be used to train the model and another subset will be used to represent the unknown it will be used to represent the future data what do you do you do a a very common ratio is to do what the 80 20 splits 80 will be used as the training data to build a prediction model and then you take the model to predict on the 20 and then you will get the performance that will be an indicator of your future performance okay and on the test sets right they call it the testing set for the 20 and the training set for the 80 okay so it it is used to make an extra collation okay if you make the performance for training sets you use a training set to build a model and then the performance is only internal so you call it model interpolation but if your model could make outside inference right because we're making inference right if you could make outside then you call it model extrapolation extrapolation okay but then 80 20 is not the only ratio it could be 60 20 20 it could be 80 10 10. it really depends you could be 50 50. okay and also you would also do cross validation for example you take the 80 the training set you take it to do cross validation but then you still keep the 20 to act as your unknown and then you do the cross validation on your 80 okay so let's take a moment to understand this infographic what is cross validation let's start with the term cross why do we have the term cross yes we have many of these roles what did each row represent they represent a single validation so that's the validation term okay so you have cross because you have many many rows and they are crossing one another i will show you how they are crossing one another okay for that i need to jump here okay so iteration one you're split the data into five fold you say that you want to do five-fold cross-validation therefore because you want to do five fold you split your data into five equal components five equal splits or you could call it five equal fold and for the five fold you would take one fold you leave it out the left out will be like a testing step remember the data splits we do training and testing eighty percent twenty percent and then we take the eighty percent to do another round of cross validation whereby we if we do five fold we will split the 80 into five fold we take one fold we leave it out we use the remaining four-fold to build a model and then we take that model to make a prediction on the left out fold iteration one iteration two the fold that we left out will now be included in iteration two we will now take a new fold and leave it out and we we use the remaining four to build a model so you can see that each fold will be left in left out left and left out so we call it cross because each fold are crossing one another sometimes they are used as the x as the test set sometimes they are used for the training sets okay but you you could also do a 10 fold cross validation as well and there's so many different types you could also do monte carlo two twofold crosswalk validation right there's so many you could do a nested cross validation where you do a cross validation you divide it into five group and then for each of the group that you combine you do another cross validation so it's like a loop nested right at the outer level you do it in the inner level you have to do it and sometimes they call it the double loop cross validation so it sounds very complicated and that is research that's a part of research in machine learning okay let's imagine you have a spreadsheet let's say you have 100 role role 1 through world 20 you leave it as fold 1. in the first iteration you want to take it out they are left out and the remaining will be used for training by different yes they are different they will not be here when they are taken out they're taken out and then second iteration they move back then and we take a new fold roll number 21 until 40 will be taken out is it random or not random it is random but you have to specify the value uh most important thing about doing machine learning when you're using like python or r when you do cross validation it will be random every time it is very important for you to set the seat number otherwise every time you run it you will get a different result and maybe you wonder why therefore you have to set the seed number okay it might be confusing what is a seed number a seed number is kind of like let's say that you could influence the outcome of everything to be the same when you set the seed number if you split the data 100 times using the same seed number you will get the same split 100 times but if you use a different seed number you split the data into two components the the constituents will be different if you have different sheet number it's kind of like shuffling a deck of cards let's say that you take a deck of card you split it into two components and you put it back and you do it again but you have to ensure that you take out equal number of cards each time and then you put it back take it out put it back same seed number but if you do different seed number you have to shuffle the card and then you take it out but if you set the same seat number again you will get the same shuffled okay so let's think of it as random shuffling if you do the random seed number okay so i told you that there's two major learning algorithms and then actually there's three right i told you about supervised learning if you have x and y it's supervised what is supervised let's say that you you're preparing for a mathematics exam the teacher gives you a practice exam okay so you get the practice you do the practice and let's say that from the practice exam the teacher will use almost the same as the practice exam but maybe they will shift the number a bit modify the number therefore you're able to practice when you take the exam okay if the teacher used the same exact practice exam and they gave it to you on the exam day if you are able to memorize all of the exam question and if hypothetically the teacher gives you the same practice exam for the actual exam you can score a hundred that is supervised learning you're learning by examples but that is called training the model if the teacher use the same exam on the real exam the same thing as a training set remember how we split the data training testing we use the training set build a model take the model predict the training set right so if we take the trained model to predict the same sample that was used to train the model we also expect that the performance will be pretty high that is the first term the model interpolation how reliable is the model in predicting the data sample that was used to build the model in the first place and then we also applied the model to make a prediction on new data that it has never been trained or seen before okay but in either way training or testing you call it the supervised learning because you are allowing the model to learn by example given x and y pair learn allow the model to learn and once they learn they can build model and make a prediction unsupervised learning you only have the x you don't have the y and a lot of things in life you only have the x you don't know the outcome what happens you could do unsupervised learning how can you analyze data without a label you could group them right group them according to similarity right so you call that clustering clustering or you could find relationship between the variables so you call that association analysis so those two are the common type of unsupervised learning when you have x but you don't have an outcome you could find a similarity you could group them together according to similarity okay like this group is similar the other group are similar but both groups are different because they are clustered differently sometimes you don't know the outcome of the patient you just measure their blood profile based on their data you cluster them and then you notice there's three cluster and then you have to make sense of the cluster then you go back analyze the data pull additional data in maybe you put another data in and maybe that could serve as your class label therefore you can see that model building is an iterative process right when you build the model you figure out there must be a missing link imagine yourself like in a movie um maybe you're you're in the matrix or you're in the minority reports right you're trying to piece in together information right this this scene in the movie of minority report report i don't really like it where tom cruise he has his computer his moving information with his hands moving from this database this database this database to combine the information just by dragging and dropping different information and seeing whether the information provides value to the existing model that he is investigating right so this is unsupervised learning and then we have another one reinforcement learning any of you play video games starcraft you know starcraft probably dot a uh what do they have now rog pubg okay what if you could teach an ai to play and what if it could learn and win it would lose maybe a million times like a starcraft someone created this reinforcement learning so what the ultimate concept of reinforcement learning is the reward given the tasks that they're doing they learn by trial and error the ai will learn that okay doing this they lose doing this they lose they lose a million times but each loss does not goes to waste each loss they know okay they know the different pathway let's imagine you have a grid a 10 by 10 very simple you go to grid let's imagine the row will be one two three four five like an excel spreadsheet the column will be a b c d e f g first iteration you go to a one you lose second iteration you go to a one you lose third iteration you go to a one you lose so the computer figure out if it keeps on going to a1 it will lose what does it do a2 it won b2 it lose b3 it won now we've noticed that it has to go to a1 and b3 so it will learn this slowly over time the ultimate is the reward is to win that's the last function in model building the ultimate is to build a model with low error with high accuracy with high uh predictability okay so in reinforcement learning you don't teach it anything you allow the computer to learn by its own and that is the concept of alphago alpha fold it learns by itself okay therefore that's the magic of ai you don't even need to teach it just let it learn and you can learn and then once it learns you want to extract knowledge out of it you want to allow the ai to explain so what did you learn explain to us so that we as humans can understand okay like for example to play chess or go to master it it requires maybe a lifetime right to be a chess master but now the computer could master it you just maybe allow it a couple of months right if you have a super computer it could even finish the test in a shorter amount of time right so this is reinforcement learning very exciting stuff right like self-driving it also learns but from the environment the car have sensors right you have the robotic vacuum cleaner the circular robot it goes around the room it has sensors if it goes near it measures the distance maybe it has like 10 sensor if this sensor is the distance is like 10 9 8 7 6 maybe you have a threshold if the distance is less than 2 trigger it to move in another direction if it goes like that 10 9 8 7 6 5 4 3 2 it passes the threshold then it has to move in another direction and there is another wall so you have to ping pong off the other direction that it goes to and now with each of that they will probably understand the room version two the first version they probably don't don't memorize but version two of the vacuum and if the creator want to do it it could memorize the path of the room and therefore it would do it harmoniously even you have a what if you have cats walking around the room or how about a human being carbon dioxide what if they have carbon dioxide sensor i don't know maybe they could create something new right it could be a version two version three improvements so you can see that ai is like experiments when you're doing experiment what are the factor how can you improve it and that's a fun thing you could improve it by coding by programming right so i don't know like for me the data i've learned to code it practically changed the way i view the world because i feel like i have more control of the world in that i could automate tasks that are boring and it could be done automatically like for example in collecting a data set i could automate the task automate the downloading automate the filtering out of the redundant row automate the removal of the missing value let's say that you have to do it manually but if you automate it maybe you spent five hours ten hours coding it maybe it might seem like a big time investment but in that five hour ten hour you don't have to do that again for the rest of your life unless the algorithm change then you modify the code again and then you can do other tasks and therefore you could do more meaningful tasks instead of click click download removing column manually because humans were also error prone you have heard of human error maybe we're sleepy we click the wrong column then the entire data set is useless right okay so in supervised learning there's two major type classification and regression so the same thing because it is supervised it means that you have x and y how does classification and regression differ do you know can some of you provide answers what is the difference between classification and regression okay okay for what for which variable for why exactly so quantity of when the y variable is a quantity like a numerical value then it will be regression but if it's quality like the class label or a categorical value like yes no okay zero one two and let's say that zero one two has a meaning zero could be don't have one could be in half two could mean have a lot right when you do when you do bacterial grading or biochemical tests you have one plus two plus three plus four plus you have the grading level right you have the intensity low medium high very high then you would use categorical therefore you could do classification because if you do regression it doesn't mean anything if you predict the value to be 0.5 what is it is it one or is it zero 0.5 doesn't mean anything unless you make it into categorical then it has to decide if it's 0.5 will it bump up or bump down so that is where the uh they call it the the threshold threshold if the predictive probability is at the middle of the threshold we when you predict it to be class a or class b it also depends on the threshold maybe we lower the threshold to be 0.4 that is why we have to do rock curve have you heard of the rock curve before r-o-r-o-c you could you could change the threshold and therefore you get you could get better performance let's say that if you use the direction of 0.5 and whenever the value is 0.5 it it's confused but if it's 0.4 it bumps up right 0.5 will be above 0.4 it bumps up it gets better accuracy right so the threshold could be another important part of the modeling process okay so now you know classification and regression okay all right so let's have a look at classification so you have x and y and your y is categorical maybe you have class a b maybe you have great abc you have class a b c right here now let's say that you you could also take the same data here x so nothing is stopping you from using unsupervised learning if you have the y variable you can just strip out the y and use only x to do the visualization of the cluster like if you have x and y right the first thing to do is okay are you gonna do classification or regression depending on your y and let's say that you take away y you can use x to build a cluster model like here and because you know what class each sample belong to you color it a different color and that's the beauty of doing this visualization you make a cluster plot using for example principal component analysis and then you color the data sample by the class label for now you could color it by okay that's the person let's say that y is having a disease or not having a disease it helps you to analyze the cluster much more and for here we can see that there is a clear distinction between each of the cluster but in a real practical setting sometimes there might be some clear differentiation between cluster but sometimes some clusters are similar and they might be closer to one another maybe they overlap and we can't tell the difference between cluster and b maybe they overlap by 50 like advanced diagram you have overlap right and aside from coloring the data sample you could modify the size to and for according to another variable back to the regression okay so in the regression you have x and y and your y is a numerical value okay it's like 1.05 95.18 and then your regression is essentially a y equal to f of x y equal to function of x and x could be many x and therefore your your simple linear regression will start by y equal x plus b like for example you could have an equation like y equal five x plus five and if you know x if x is one you could calculate y if x is one five x five multiply one you get five plus five you get ten y equals ten right if you plug in x equal to two you get the value of y right so this is regression however your equation is the linear line that you see here but in reality the data sample are distributed around the trend line above or below because the trend line is an approximation okay now every time when you say something you say 5.5 with an sd of 1.8 so 5.5 1.8 up that's the upper bound plus 1.8 5.5 plus 1.8 5x5 minus 1.8 that's the lower bound of your data right so you know the upper and lower bound of 5.5 that's your mean right so therefore if you have x you could calculate y okay so this is an infographic showing how to do qsar quantitative structure activity relationship right in prior lecture you have learned about how to draw a chemical structure already right here you know how to draw this molecule and therefore these are the molecule right molecule one molecule two molecule one and two how do they differ i showed it in the color it differ from this one the red atom that you see here and the green atom in multi-one it has a ch3 at the position here but in molecule two represented by the red dots it doesn't have that but instead it has cs3 at this position which in molecule one it doesn't have therefore it is represented by a green dot if you zoom in you see the green and the red but what's important is that the different chemical structure when you represent it in a computer you do this process of calling a performing molecular descriptor calculation so it allows you to convert a molecule into binary form into numerical form which is called the molecular descriptor and this is a type of descriptor called the molecular fingerprints so it tells you whether you have it or you don't have the functional group if you have it it will be one you use a value of one if you don't have it you will use a value of zero okay and therefore you make a data set each row like the first row here the first row is white color right that's the the name of the variable the name of the x1 x2 x3 the name of the y variable the blue color that you see is molecule one the yellow color that you see are molecule two why column is the class label what is the output is the molecule active or is the molecule inactive all of this is called the data sets and then you use the data sets comprising of the x and y to build a model a prediction model but because your y is a class label it is categorical qualitative therefore your model will be classification once you build a model you get this gray box right you have a new molecule molecule three purple color the first thing you need to do is take the molecule chemical structure that you have drawn and then you want to convert it into molecular fingerprints after you get the fingerprint you use it as an input to the model the model will make a prediction on the inputs you get a predicted value and then that's your prediction your model will also generate a feature important plot that tells you which feature are important shown here in the future importance you can see that okay x8 x10 are the most important feature for a molecule okay but it doesn't tell you important for what important for active or people are for inactive then you have to do one more thing you take that feature and you calculate the box plots but you have to stratify the data you have to separate it into active inactive remember when i tell you that you have a column you call it the activity and then the value under that one single column will be active inactive active inactive but if you could split that into two separate components you would have a hundred rows being active only then you would have another maybe 150 row being inactive you separate them you separate the active and inactive in statistical analysis you call that the stratification like for example if you would like to see the effects of smoking on the population health right smoking status column is stratified according to smoker non-smoker otherwise your data will be combined but you have to separate it into smoker non-smoker and for each group you do your analysis you you do your p you you do your t test you compare each column because our machine learning model say variable x 8 and x 10 are important you do a t test between the active and inactive for x8 and x10 is it statistically significant okay chances are they are and you want to see how much okay so therefore you can see that actually no one tells you how to do this it's an arts you have to learn the tools of the trade and then once you understand to do each of them you have to figure out in what sequence you want to do it you want to do this first build the model get the future importance get the top important feature top feature is important do t-test separated stratified into active inactive you get the p-value that they are significantly different right so you could do many technology stacking right apply method one method two method three in order to explain your hypothesis that x8 has an impact on the activity therefore you do t-tests you visualize it using box fonts you could use a bar plot you use t-test p-value right so there's many ways and yeah so that that's the major way that i would do so we have already covered the entire process here all right so now let's take a look at this if you want to do this yourself you have three routes round number one no code okay you don't have to do any coding you don't have to learn any r or python and there is a low code meaning it requires you to use python but it requires you only minimal usage of coding maybe one or two lines and mostly if it's low code it will be auto ml automatic machine learning or automated machine learning it means that what you just need to do you just need to use r or python read in the data clean the data and once when the data is clean you use the the library like pi carrots you use only one or two lines of code or three lines depending what you want to do and then you're able to generate like 20 30 models in only one or two lines of code but in option three code this is more hardcore so either python or r you have to write your own workflow how will you impute missing value replace missing values do you want to replace the missing value by the the mean value or the median value or do you want to delete the entire column if there's a missing value or you want to delete the entire row that's for you to decide okay if you have only one missing value you want to delete the entire column it's a waste delete the row if more than half of your data is having missing value in that particular column you could either delete it or you could also put in the mean value or the median value okay that's the impute they call it imputation you could search that and you could easily spend the entire day reading about it how do you replace missing value in a column or in a data sets so when you analyze the data you will see hundreds of problems therefore the important thing is how can you iterate through that very quickly okay so number one there's several ways that will that will make you slow number one you get overwhelmed when you get overwhelmed what you do you could complain it's too difficult there's too much data and if you complain then therefore you will not do it and when you don't do it you waste time and then maybe one or two months later you figure out you have to do it otherwise you won't graduate your graduate degree then you go back to doing it but then you wasted two months of overthinking right overthinking when you see the problem you don't do it you're procrastinating you figure out to find an easy solution you look back there's no easy solution it's only you because when you're doing a graduate study the thing is you have to own your projects the project is yours right if you don't do it the project won't be complete it won't have any progress and that will only hinder your progress right complaining over thinking what else so how can you do it you just need to take action just do it make errors fail learn from the errors iterate back right there's this concept called the ooda right so i've written a blog about that how to learn seven effective ways to learn data science one of them is about using understanding the ooda loop i think you probably have heard in in clinical laboratory they have the pdca plan do check x so we learned about no code low code and also coding python or r and how that could help your your analysis let's go back here here so i talked about learning data science i i think it could be applied to anything you just take out the term data science you put in net tech you put in the topic that you want to learn microbiology first step is planning your your curriculum what do you want to learn about therefore you'll be proactive in your learning make a list of topics that you need to learn about and then for the list of topic you have to figure out okay where where's the resources you have to make your learning effortless you know the book from atomic habits change clear how can you make it effortless for example if you want to read a book you put the book next to your pillow so therefore before you go to bed you see the book is already there and you can just read it or how about making goals posting it everywhere look at that post-it notes put it in the fridge in the shower the door study microbiology study chapter 2 study whenever you walk around the room it's a constant reminder you could use it for consistency reminding you what you need to do because sometimes when it is out of sight what do you call it it's out of mind but if you see it it's a constant reminder that you need to do it so that is all of the planning knowing what you need to know knowing what you need to do and having a schedule and if you have a study buddy accountability tell your friend i'm gonna finish the chapter two today and then your friend tomorrow will ask did you finish chapter two and then you answer yes i did and you feel proud if you say no i did not you feel miserable so your friend will be accountability partner okay so you can study in a group learning tips technique technology tools there's so many i told you about calendar notepad and you know like notion and there's so many more you can even buy a pomodoro timer you know the tomato timer you you set like 40 minutes to study or 45 minutes to study and you can spend 10 days 10 minutes 15 minutes for break and you can repeat the cycle over and over okay because your attention span will lose like this lecture i mean you might get sleepy over if it passes 45 minutes right so number two learn right there's so many resources for you to learn about make a list of what is the resources that you will use or you could write an article make a note make it as detailed as possible make a mind map so if you create what happens is that when you want to create something when you want to draw something you need to have an understanding of that topic if you don't have it then you have to go back to the notes read it understand it draw the schematic and when you can picture that maybe the art of drawing it will allow you to synthesize it materialize all of the knowledge which is floating in the air putting into concrete form objects when you have the notes or the infographic then that's your knowledge materialized in the infographic okay so actually the art of making it the craft of making it it's like a revision you're already revising for your exam by making that artwork right and the most important is to explain to someone else teaching your friend about the topic writing about it right it allows you to make sure that you understood the topic okay that's all and okay so thanks for your attention