# Lecture 4: Protein Structure Data

[Recording](https://www.youtube.com/watch?v=uVBXCMNSp9Y)

Automatic transcript; technical terms and equations may contain recognition errors.

00:00:00 Okay, let's get started. Howdy. So in the past few lectures, a couple of lectures, we basically

00:00:11 understood the molecules that carry information and whose data are to be analyzed, right? Looking at

00:00:19 the central dogma of molecular biology, we know that there's DNA, RNA, and proteins. We know sort of

00:00:26 what kind of information is carried on these molecules. What is the chemical

00:00:33 foundation of this information? What is the biological utility of this information?

00:00:39 And more importantly, what are the data like for these molecules, right? So we had

00:00:47 still had to talk about some concepts last time because, you know, protein sequence,

00:00:53 secondary structure, which is a local 3D, and then tertiary structure, which is a

00:01:00 global 3D, right? And there are a few additional concepts like, you know, each

00:01:06 protein could consist of multiple chains. Each chain is a contiguous sequence, and

00:01:13 each chain can fold in three-dimensional space and adopt multiple domains. Those

00:01:20 domains are spatially a little bit separated from each other, but in one-dimensional space,

00:01:26 like the sequence, it could be non-contiguous. And really, these domains are probably independently

00:01:34 evolved, and they are having separate functions, right? So nature somehow find out a way to modularize

00:01:43 their designs and then put them together somehow, right? And we also know that each domain can adopt a

00:01:50 different kind of fold, which is like the pattern of the domain, which is like the pattern of how

00:01:57 secondary structures are packed together into the global tertiary structure. So these are the contexts,

00:02:04 and today we're going to talk more about the data, and then, you know, not just the databases, the format,

00:02:13 the representation, the implication, the challenges, some visualization, but I don't think this laptop

00:02:21 interface would work. But, you know, there's also all these references if you are interested in. So I

00:02:34 believe that I have given the link to each book, I think we have the electronic subscription, but I

00:02:44 think there are also shorter ones, like, you know, if you are just interested in more essentials, I give some

00:02:51 links, like how to understand the structured data, what's the format, and then I put this squared bracket

00:02:59 here to show that it's just optional because you may or may not be interested in visualizing the structures

00:03:05 by yourself. In the end, I want to show you just two simple stories, not really simple because they are

00:03:12 high-impact journal papers, but I want to use this to give you a glimpse of how you can use data to answer

00:03:20 questions, okay? So keep in mind that, you know, the focus of the data science is not how to use a fancy

00:03:27 algorithm. The focus is to answer, you know, ask the right questions and answer the right questions in

00:03:34 the right way. So that's what, you know, data science is most about. So I'm going to use two very simple

00:03:40 stories, simple in the sense of how data is being analyzed, but they're not simple, but in the sense

00:03:48 that they show you how you use data to answer the questions, okay?

00:03:54 So let's get started. I believe that this was included last time, right?

00:04:01 We talked about, you know, a lot of databases, including database of databases,

00:04:07 but if you're talking about the sequences, especially for the DNA sequences, the NCBI has some

00:04:16 database of databases, and if you click each link

00:04:18 on your PDF copy, you would be directed.

00:04:24 We also talked about the universe of protein, UniProt, and they can be, if you click this link,

00:04:34 they are actually redundancy removed into different subsets, like UniPARC,

00:04:39 that you won't see identical species from this, sorry, identical sequences from the same species,

00:04:48 or UniRaf 100, UniRaf 90, all the way to 50, which shows the granularity. Basically, you know, you have

00:04:57 less and less granularity for this sequence, and a subset of them are labeled with functions, right?

00:05:09 And then if you talk about the structure, we're going to talk about the protein data bank, which are

00:05:14 experimentally derived. But keep in mind that, you know, there were a lot of databases of,

00:05:22 you know, algorithm-predicted structure models as well, but none of them was really paid attention to that

00:05:31 seriously, especially by a biologist who wanted to answer some questions.

00:05:36 But I think AlphaFold 2 single-handedly changed that scenario. So it's pretty much like the first time

00:05:42 that people start to regard the predicted protein structures as real data, and they are studying

00:05:47 these predicted structures. I just saw a bio-archive paper, which is, you know, AlphaFold has already

00:05:58 predicted all the protein structures for the human proteome, right? And they are doing that for other

00:06:07 species as well. So one paper I just saw on bio-archive, actually, was that they were looking at,

00:06:13 you know, some major cancers, and they look at what protein targets can be targeted in those treating

00:06:21 those cancers. And some of them, you know, never had been solved in structure. We never knew what their

00:06:27 structures are. So it's hard kind of to rationally design a drug to bind to that protein and affect

00:06:38 its behavior or function in some way, right? So I think in this case, in Silico Medicine, which is a company,

00:06:47 looked at one protein target with implication in lung cancer. They took the AlphaFold predicted

00:06:54 structure, then identified that on that structure there's some pocket that you can target with small

00:07:00 molecule. Then they used their typical ligand discovery or small molecule discovery pipeline that you can

00:07:08 start with a predicted pocket and then design some small molecules or filter or screen for some small

00:07:16 molecules. And they have actually found some hits, a very good binding or activity. It's like, you know,

00:07:26 10 millimolar, I'm sorry, micromolar, yeah, millimolar, I'm sorry, yeah, 10 millimolar. It's pretty good.

00:07:33 So in other words, these predicted structures are now treated as data now. I mean, a lot of people are

00:07:38 asking, well, what are the mutation in effect? How can they be explained by the predicted structures? We're

00:07:45 actually getting a glimpse of these kind of stories at the end. But back then, you know, we were looking at the real

00:07:52 experimental structures. We were not looking at the predicted structures.

00:07:57 And then there are protein domains, the families, and the functional size, and then protein-protein interactions, and then

00:08:04 functional genomics, the pathways, and genotype-phenotype association, right?

00:08:10 So I'm curious, anyone, any of you are using

00:08:14 other kind of data?

00:08:15 Yeah? Because this is absolutely incomplete, far from, you know, what you need even for the molecular

00:08:23 biology, right? Any other kind of data sets you want to mention? Just curious.

00:08:31 Pardon me?

00:08:34 TCGA, the Cancer Genome Atlas. Is that right? Yeah, thank you. Good. Thank you, Thauman.

00:08:40 Okay, just curious. Okay, so now let's talk about the structured data. Okay, we know the concept now,

00:08:50 finally. Now let's talk about data. As I said that the major database for experimentally solved protein

00:08:59 structures are called the Protein Data Bank. The link is rcsb.org without PDB, okay? So we can look at,

00:09:08 you know, this website.

00:09:14 And this is, you know, today's rcsb.org. You see that, you know, it has all kinds of

00:09:23 molecules, and months, and feature, and highlights, and everything, okay? So,

00:09:29 for instance, if we are in a, of course, you can search the proteins of interest

00:09:35 using their keyword, like, PDB ID, it's like four letter.

00:09:41 But, sorry, four alphanumerical digits. But if you don't know anything about it, let's say we're

00:09:50 interested in the SARS-CoV-2 spike protein, for instance. Apparently, you already get some hints already,

00:10:02 right? Let's look at what we get.

00:10:11 And there are a lot of structures already, right?

00:10:15 The reason that we have only four structures is because of our keyword, but I just want to show

00:10:20 one example, for instance, right? For instance, this is one example. You notice that there is a four

00:10:29 digits of numeric, 7CWM. This is the PDB ID. That's a keyword that you can find the entry in the database,

00:10:42 right? And what it is, is a complex. Complex is like the quaternary structure formed by multiple proteins, or

00:10:52 in this case, it's a complex formed between the spike protein and a fab. This is an antibody, okay?

00:11:00 So, and this with one RBD, this is a receptor binding domain of the spike protein, and this RBD is in an open state, okay?

00:11:11 So, as you know that this, the spike protein has like three monomers bundled together, right?

00:11:20 It's really amazing. Looks like an anchor. It's really anchoring into our human cell, isn't it?

00:11:26 So, one of the, the tip of the anchor is the, so the tip of the anchor is the receptor binding domain,

00:11:42 and you cannot see it very clearly, but this part is the trimer, the three monomers together,

00:11:50 and then RBD is on this tip, and then two of them are in the closed state, and one of them is in the open state,

00:11:59 right? So, and the, and one of them is in the open state, and then this RBD is neutralized by an antibody.

00:12:10 This is an antibody, just right here, with two chains, heavy chain and light chain. So, this is an antibody, okay?

00:12:18 Okay, so, if you ask what kind of structure you get, what kind of data you have, you know,

00:12:27 you can download, of course, you can display. Remember that I said what is the format of a fast start sequence

00:12:34 last time, right? It's essentially the sequence, but with a header starting with this bigger than,

00:12:40 right? Symbol. And you can put any kind of keywords for your own reference in the header,

00:12:47 separated by the vertical line. So, in this case, it's actually using the PDB keyword, 7CWM chain one,

00:12:57 and, sorry, molecule one, I think, it involves chains A, B, and C. So, that's the spike glycoprotein,

00:13:06 which is in the species of SARS-CoV-2, okay? And then this is the sequence. Then you have two more chains,

00:13:15 7CWM chain two, and 7CWM chain three, that corresponds to the D and G. There are actually

00:13:26 multiple copies, so, DG, EH, and Fi are three pairs. So, one of them is a heavy chain, and one of them is a light

00:13:36 chain, and that's a homo sapien, a human. So, that means it's a human antibody, right? So, and again, you have the sequence right here.

00:13:43 So, okay? So, that's the fast start sequence.

00:13:47 So, for instance, this, remember that each letter is actually corresponding to one amino acid or a residue in the context of a protein,

00:13:58 so it actually consists of a lot of atoms, right? So, now let's look at, for instance, the PDB format, okay?

00:14:09 It has a header, you see that, the header, the title, and the compound. You sort of see that they are like,

00:14:15 you know, the first one, two, three, four, five, six, six columns are like telling you the kind of the input.

00:14:24 It's telling you what's a header, what's a title, what's a compound, what's a source, and so on, right?

00:14:31 And it has additional information like the journal and the remark, right? So, this is a 3.6 Armstrong

00:14:40 crystal structure. You don't know what that means. It's fine for now. And this tells the experimental

00:14:48 protocol to generate the one. And what we are mostly interested in after the remark, excuse me,

00:14:54 it's going to be a very long, is going to be the structure information. So, in other words, we need to know

00:15:02 every atom, every atom's Cartesian coordinates. So, this is a sequence. Now, this is a secondary structure,

00:15:13 helix sheet, what you learned before, link. And then now we start starting with this atom, you see.

00:15:22 So, these headers, or the fields, you know, the concept is called field. The field of atom gives you the

00:15:30 Cartesian coordinates of every atom, okay? Except some hydrogens which are too small to be,

00:15:39 you know, localized, okay? For instance, in the FASTA sequence, at position 14, right, of chain A,

00:15:48 what you see is in the FASTA sequence position 14, one, two, three, four, five, six, seven, eight, nine,

00:15:54 10, one, two, three, four. It's a Q, right? Position 14 at chain A is a Q. So, that's actually the first

00:16:03 residue. Q is one letter code for glycine, no, glutamine, GLN, right? So, that's exactly what is

00:16:13 here. So, it says that I have the 14th residue in chain A, that's a glutamine, Q, okay? If I convert it to a

00:16:23 single letter, but in the PDB, it's a three letter. And for each residue, I know that, you know,

00:16:30 it consists of different atoms. For instance, there's a central atom, C alpha. There is a nitrogen

00:16:36 on the amino group, and then there's a CO on the carboxyl group. Then you have C beta, C gamma, C

00:16:42 delta, oxygen epsilon, and the nitrogen epsilon, which are already known because for topologically

00:16:51 already known because you know what each residue or amino acid consists of. But then here, each atom

00:16:57 is given its coordinates, X, Y, and Z. 232, 167, 153 in the unit of Armstrong, okay? So, that's the idea.

00:17:10 So, now let's come back to the slides and, you know, give us a little bit more introduction. So,

00:17:19 essentially what the data you have here is, of course, those sequences whose structures have

00:17:25 already been solved, right? But you see that my chain A started with position 14, not one. Why?

00:17:31 Because not all residues can be seen in the structures, okay? Some parts can be extremely

00:17:38 flexible and low resolution to identify. Some of them are intrinsically disordered. So, be aware that

00:17:44 it's not just, you know, few sequences have structures, but even for those sequences who have structures,

00:17:52 you may have incomplete structures. Not all the residues can be visualized or, and their atoms

00:18:00 don't necessarily have the Cartesian coordinates.

00:18:02 So, whenever you talk about data, you want to talk about a little bit about how these data are collected.

00:18:12 So, I think it's worthwhile to know just a little bit how these data are collected, okay?

00:18:17 So, there are two, three types of methods mostly, experimentally, X-ray, NMR, and cryo-EM.

00:18:27 Okay? So, the classical method is X-ray crystallography. So, you sort of, you know,

00:18:37 you know, grow crystals of these proteins and then, you know, shine X-ray through the crystal and see

00:18:45 what's the diffraction pattern, right? And then that gives you some, you know, you know that how,

00:18:53 you know, I think a father and a son together won the, did they won them together? Did they won the

00:19:01 Nobel Prize together or they won them separately? I'm talking about the Bragg's, you know, the diffraction,

00:19:09 what you learned in physics? Well, we can look it up. So, you know, this is the physics that we learned,

00:19:16 you know, how the diffraction are. So, in other words, it's like a typical inverse problem.

00:19:24 You look at the diffraction pattern and you infer what's the original object, okay?

00:19:32 And an inappropriate analogy is like looking at the shadow and estimate the 3D shape. But this is not

00:19:39 the shadow. This is the diffraction pattern, okay? So, if you, have you, have you seen this very famous,

00:19:48 I think we can look at this on, the famous diffraction pattern for the DNA?

00:19:54 This one, okay? This is the Rosalind Franklin's x-ray diffraction photo of DNA. So, he had this,

00:20:19 she had this data and her boss shared the data with Ross, uh, Watson and Crick, who had been wondering

00:20:27 what's the shape of the DNA? So, so, when they saw the photos, they were like, oh, a double helix,

00:20:35 okay? Of course, this is not how, it's, your eyes are probably not trained to see, see it that way, but, um,

00:20:42 this is just the one example. So, look at this x-ray diffraction pattern, it's very regular, right? So,

00:20:49 it is consistent with a very regular double helix, okay? For proteins, it's a little bit more complicated, but

00:20:57 it's, it's a well-solved problem. But, of course, there are a lot of, uh, challenges, right?

00:21:04 For instance, crystallization can be very difficult and the limit, uh, can limit protein types.

00:21:10 Years ago, getting the G protein coupled receptor structures were still difficult, right?

00:21:16 So, or some proteins which are, uh, flexible, right? So, they could be. So, it's a, it's a detailed atomic

00:21:23 information, but it, experimentally, it could be difficult, okay?

00:21:27 So, you have the NMR, uh, this is, uh, um, you know, you don't have to grow crystals anymore,

00:21:36 but you put, like, in the proteins in high concentration, okay?

00:21:40 liquid, uh, solvated in, uh, in solutions and in high concentration. So, it works for, even for, uh, flexible proteins.

00:21:50 Excuse me.

00:21:50 Um, and the, the, the, the physics behind it is, you know, you, you sort of put it in some, you know,

00:21:57 electromagnetic field and then you observe the magnetic resonance, okay?

00:22:03 And then, again, you solve an inverse problem, right? Observing the, uh, magnetic resonance and infer

00:22:10 the atoms that are close to each other, okay? So, once you figure out which atoms are close to each other,

00:22:17 and then you can infer the 3D structure, so, right? So, um, so, um, so, this is, uh, often not give you a

00:22:24 single structure, but you give you an ensemble of structures. So, um, so, uh, in the PDB, you may end up

00:22:32 having, like, multiple models. Model is another field. You have model one, two, three, four, five, six,

00:22:38 until, 12 or more. So, you really get an ensemble of structures. So, the resolution is not necessarily as

00:22:46 high as, uh, X-ray, okay? When I say resolution, for instance, you know, a one, one, one, one, one

00:22:52 Anstrom resolution is much higher than, uh, uh, you know, 10 Anstrom resolution, right? So, cryo-EM is, is the

00:23:01 last one, you know, it's a cryo-electron microscopy. This was, uh, you know, nature methods of year,

00:23:10 which year was it? Three, four, five years ago or four? So, if you, if you look at, you know, nature methods,

00:23:17 this journal, the method of that year was cryo-EM. Uh, this is, uh, uh, a wonderful technology combining

00:23:27 the advancement in, you know, physics instrument as well as, uh, data processing, okay? So, so, what it's doing is,

00:23:38 essentially, you know, quickly, cryo is like, you know, quickly freeze these proteins and then,

00:23:44 you know, you, you take a lot of pictures from different, you know, angles and then these different

00:23:51 pictures can be aggregated. A lot of people spend time figuring out, including how to do the Bayesian

00:23:58 aggregation, and then you can get a very high resolution. It's still lower resolution than X-ray

00:24:04 from time to time, but, you know, um, you can overlay a lot of low resolution, um, images and, uh, from

00:24:11 different angles and, and get a higher, higher resolution 3D shape. And, uh, I think, you know,

00:24:19 in the past, there was a race of resolution because X-ray can get, uh, one or two Anstrom, right? That's the

00:24:26 resolution, but the cryo-EM used to only get, um, uh, 15 Anstrom. That's already good, but in recent years,

00:24:34 we keep seeing, like, seven Anstrom or five Anstrom or even, you know, lower, uh, or

00:24:41 in other words, higher resolution images from cryo-EM. So this is very good, and there's another

00:24:46 advantage. You probably noticed that the spike protein I showed you, uh, was, uh, actually, uh,

00:24:53 EM structure. And, uh, it's not, uh, without a reason because, you know, cryo-EM is very good and, uh, for large

00:25:02 macromolecular assemblies, okay? So, so for large molecular assemblies, cryo-EM is, is very, uh, is very

00:25:12 powerful, okay? So, um, if you click this link, uh, you will be directed to three short videos telling

00:25:21 you what these experimental, um, methods are doing. But overall, the big picture is the following, that,

00:25:30 um, you know, these are, um, experimentally valid data, but it's not universally available for all the

00:25:42 protein sequences because we can do, you know, to determine protein sequence, what do we do?

00:25:47 High-throughput sequencing, next generation sequencing, right? That's extremely high-throughput.

00:25:53 But, uh, and you, it's usually universally applicable, right? And now we are, we are, we are seeing what,

00:25:59 nano-pole, uh, sequencing, right? It's even, even more, uh, throughput. But for structure determination,

00:26:06 you, you know, I think a lot of automation is like robotic arms and preparing the samples, right? But other

00:26:12 than that, uh, it's hard to say it's high-throughput, okay? So that's why there are, there's a very few

00:26:18 protein sequences that have been, um, structurally determined. That's why protein folding or protein

00:26:25 structure prediction problem is so computationally significant. That is why did MIND put their eyes

00:26:30 onto this problem. Okay, so now, uh, knowing that what these data are and then we can sort of try to

00:26:40 understand the, the exact data format, right? So again, there are two links, right? A very good one

00:26:47 is understanding PDB. That's also on the RCSB. So, uh, you can take a look at it. And then there's more

00:26:54 detailed explanation of the PDB file format, okay? So, um, but basically, uh, you know, if I want to,

00:27:06 for instance, looking at, uh, this link, understanding PDB,

00:27:17 let's see,

00:27:17 there's no copy. So I have to type it.

00:27:32 And what is the keyword?

00:27:35 Gnkqfns.

00:27:41 FNS.

00:27:42 Gnkqns.

00:27:46 FNS.

00:27:57 Okay, so this is the introduction to the PDB data. It tells you, um, uh, the concept as well as how to

00:28:05 visualize and read things in general. Okay? And then it tells you, for instance, when I say

00:28:10 it's a one-anstrom structure versus a two-anstrom structure, what, I mean, what are behind these

00:28:18 numbers. Okay? Um, and with a one-anstrom structure, you have the highly ordered and easy to see every

00:28:28 atom in the electron density map, but lower resolution, like three-anstrom or higher, shows

00:28:33 only the basic contours of protein chain, like, like this one. Okay? So, um, the sticks are the,

00:28:41 the atoms, but you don't really see it. What you see are the contours only, right? But if this, uh,

00:28:47 contours are a higher resolution, then I can trace the contours and actually figure out

00:28:54 the detailed positions of every atom, right? So, so that's what the resolution means.

00:28:59 And then, uh, there is this, uh, uh, PDB format. Um, I think I gave a link as well. Okay? So this PDB format 33.

00:29:24 I think, is it? It's a wrong, yeah, wrong hyphen. So, hyphen, format.

00:29:43 33. It tells you that there are different title sections. And, uh, uh, you have seen this before.

00:29:54 The header, the title, the compound, and the author journal, these are the headers. Uh, what we care

00:30:01 about the most are the coordinate selection. Uh, as I said, that NMR structures, um, each one contains

00:30:11 different images or different structures. Like, if you have 12 structures, um, in, in one NMR PDB,

00:30:18 you have model 1, 2, 3, 4, until 12. So that's what the model is for. Under each model, you have a lot of

00:30:25 atoms and the heteroatoms, like water or some, uh, ions can be in the heteroatom. Uh, the ter is the

00:30:35 terminal, and then this is the end of the model. Model and end of the model. And what we want to pay

00:30:43 attention to the most is the atom section. Okay? That's what I'm going to talk about. So, um, you know,

00:30:50 you can look at this, um, into more details. But, um, in the slides, I give the same kind of information.

00:30:57 So, in other words, these atom fields, right? So, so, you know, the PDB file is, you know, text lines up to

00:31:09 80 columns. And you have different, uh, each line is, has, has a record name, right? Which is the first

00:31:17 six columns. Uh, and the one we care about the most is the record name, ATOM, atom, right? Because that

00:31:25 stores atomic coordinates for the structures. So, let's see what the atom lines look like. Can you see

00:31:31 the, the image clearly? Or is, is it blurry? Or is it fine? So, you have 80, uh, you know, you see this

00:31:39 index from 1 to, this is 10, this is 80. So, 80 columns in total. The first six columns are the, you know,

00:31:49 the record name, which could be atom or heteroatom. And, um, so in Python, you know, you, you start from

00:31:58 position 0 to 5, right? And, uh, for, and then you have the atom serial number. So, this is the index of

00:32:08 the atom, right? And then, uh, this is, uh, uh, the next five, okay? The next five. Seven, eight, nine,

00:32:19 ten, eleven. Okay? It's right justified. Okay? These, um, are right justified except the first. Okay?

00:32:29 Then, for each serial number of the atom, you have a corresponding atom name, okay? So, this atom name,

00:32:38 again, occupies three columns, uh, actually four, four columns, uh, 13 to 16. And in this case,

00:32:49 it's a nitrogen. It's a backbone nitrogen that's in the amide group. And Ca is the central carbon,

00:32:56 that's the C-alpha. And this C and O are the carboxyl group connected to C-alpha. We don't have the

00:33:03 third, uh, atom connected to the C-alpha. That's a hydrogen, because usually hydrogens are too small

00:33:09 to be observed. And then, uh, in this case, we don't have the side chain either. It's a glycine,

00:33:16 by the way. The next field is the, the, the residue name, uh, position 18 to 20. This is 18 to 20. That's the,

00:33:27 that's the residue name. Glycine, three letter code. Uh, what's the side chain of glycine?

00:33:33 Glycine. Remember, it's the smallest amino acid. So the side chain of glycine is only a hydrogen. So

00:33:39 that's why you don't see hydrogen here either. You don't see that the side chain hydrogen, because the

00:33:43 next one is a proline, right? Okay. And then you have a chain ID, right? Uh, that's, uh, you know, uh,

00:33:54 chain identifier, uh, for, uh, column 22. And then, uh, you have the residue. Okay. You have the residue

00:34:06 sequence number. So I actually started with residue three, four, and so on. That means the first two

00:34:13 residues are probably not included in the structure because they are probably too flexible, right? To be

00:34:21 observed, right? Or for some other reasons, biologically, biochemically, they didn't include

00:34:28 that part. It's also possible. Okay. But one thing you have to be careful about is, uh, you know, you may

00:34:34 want to look into, you may not want to look into these details too much, but this number three does not

00:34:40 necessarily mean that it's a third residue or character in the FASTA sequence. Because in the

00:34:47 FASTA sequence is one kind of residue indexing. But in the PDB, it's uploaded by different authors from

00:34:54 different groups. They don't really form the same standard. So this number three, maybe they're internal

00:35:01 indexing. So you've got to be careful. This doesn't mean that G or glycine is the third character of the

00:35:07 FASTA sequence. Okay. It may be the fifth or the, you know, or the first. Be careful about this indexing.

00:35:16 And then we have, uh, three, um, fields corresponding to X, Y, and Z. Again, right justified. Okay. Be careful that,

00:35:28 uh, I don't think you can go beyond, uh, uh, uh, 9999,

00:35:35 Anstrom, right? So if, if it's, otherwise it could go flow these corresponding fields.

00:35:41 And then lastly, there are a few columns you may not pay too much attention to. For instance, uh, this green one is from 55,

00:35:49 right? Starting from 55 to 60. This is the occupancy.

00:35:54 Uh, you can regard this occupancy as if it's a probability mass function. What you see here is like one.

00:36:00 It means that, oh, this atom is 100% there. But you could have a case like, you know, N

00:36:07 and then 0.5 and then there's another N below another 0.5. Okay. And then they have different, the same atom,

00:36:15 but different coordinates. It says like, it's like 50-50. Why? Because the density is too smeared or ambiguous

00:36:22 to be, you know, nailed down to a single structure or a single coordinate.

00:36:30 And then you have, uh, the next field is, uh, 61 to 60.

00:36:35 And that is a temperature factor. You may not pay too much attention to it,

00:36:39 but it's showing the local flexibility, right? The higher temperature, the more flexible.

00:36:44 And lastly, this column is showing what's the element of the, uh, atom, right? Element was the first

00:36:52 chemistry concept we reviewed. So as you can see, this is very, uh, very simple. So one, um, uh,

00:37:00 caution to make is that this is a sort of the standard PDB format, okay? But not, you know, you see that in

00:37:09 different groups, including the computational groups, they make predictions in the PDB format.

00:37:14 They do follow this 80 column rule, but it doesn't mean that they always include all the columns.

00:37:23 Do you see what I mean? All the fields. So in other words, I could have a PDB structure that does not contain

00:37:30 this, uh, factor or does not contain this temperature, uh, or does not contain this element. Or sometimes, uh,

00:37:38 they, it's a single pro, it's a single chain, so they didn't even put a chain ID here. It's always possible.

00:37:45 So that means, you know, when you are using a computer program to process this PDB data, you shouldn't

00:37:52 split them into multiple columns of an array or a matrix simply based on the space delimiter. Do you see

00:38:03 what I mean? Because, you know, they could, some could be missing. So that means you should stick to

00:38:11 the range, which is very rigorously defined because everybody uses this range index, but I cannot guarantee

00:38:16 that all the fields are in every PDB. So that's something to be aware of, okay?

00:38:22 Okay. Okay. So, um, so again, you know, there's a little bit more details and a nicer or more details,

00:38:34 but I feel, you know, that colored ones was nicer, but you, you want to be more details, it's in that,

00:38:41 uh, format, uh, in that one. I believe that your homework one problem one, um, would involve some

00:38:47 processing of the structured data. I think it would be fun to, to put what you learn into practice. Okay.

00:38:56 Okay. So this is the tertiary structure data. Okay. Any questions so far?

00:39:11 Yes, Nami.

00:39:17 Um, the, the residual numbers, you said it starts from, I guess, three, four, is that my correct?

00:39:25 This is from the three?

00:39:26 You, you, you could see it starting from any number one or three or anything, but.

00:39:31 And you said that the, the first ones, they couldn't like capture it or something?

00:39:35 Right. So there are, you, we have to differentiate two situations. One is that, remember, these residue

00:39:41 indices are author designated. Okay. Let's say that this all, you know, in case one,

00:39:47 author designated residue index number is exactly where it's order in the sequence. Okay. In the text,

00:39:55 bar star sequence. Then in those cases, if you see a three at the beginning, that means the first two

00:40:02 were just not, uh, positionally identified. For instance, why they appear so often for the, you know,

00:40:09 the termini? The first two are in the N terminus, right? Because the terminus are more flexible,

00:40:14 isn't it? They're especially less restricted, right? Um, you could see things missing in the middle as

00:40:21 well as you deal with the structures. You see, you may see three, four, five, and then you jump to,

00:40:26 uh, eight, nine. Right? Because six or seven are missing, you may wonder why six and seven are missing.

00:40:32 Can you imagine what kind of six, seven could be missing?

00:40:36 Uh, the hint is the secondary structure. Alpha helix is stable. Yeah, exactly. Beta street is stable,

00:40:44 but a loop, right? Right? Especially a long loop could be very flexible and hard to be, you know,

00:40:50 in other words, the electron density is extremely low resolution. I couldn't identify, right? Right? I don't

00:40:58 want to use 30 coordinates and each with occupancy of 0.03, right? I don't want to do that. Um,

00:41:12 Do you give small ways to those flexible groups that may be even difficult to characterize

00:41:21 experimentally? That's your question, right? So I guess you are asking, do we give smaller ways in

00:41:27 assessment of protein structure prediction? Is that right? That's a good question. Uh,

00:41:33 the design of the assessment metric for the protein structure prediction problem itself is a research

00:41:40 topic. And in the Olympics of protein structure prediction called CASP, critical assessment of

00:41:47 structure prediction, where deep minds of a fold got its name, they actually change the assessment metrics

00:41:54 metrics every cusp. They even change the assessor every time. Okay? Part of the reason is that they don't want

00:42:06 people to game through the system and just try to make their predictions have with better assessment

00:42:12 metric one. You see what I mean? Because it's hard to pinpoint a single assessment metric that can bring out

00:42:22 the enough significance of the structure toward the biological interpretation. So that's the reason.

00:42:29 So you may see that every time people may choose, you know, a lot of ensemble assessment metrics,

00:42:36 right? And then in each assessment metric, all the groups are ranked, uh, and they are giving a score in,

00:42:42 in, in, in, in, and, and, and ordered. Then a Z, a Z, a Z, a Z, a Z score is calculated for each group under each

00:42:50 assessment metric. Then you add all the Z scores together across all the assessment metrics. Then you rank all the groups again. That's a, that's sort of the final. Okay?

00:43:01 But if you go to the CASP, you can also click each column and

00:43:05 find the assessment metric that, that gives you a, uh, a better advantage. But probably some assessment metrics,

00:43:11 more talking about the local, uh, details like the sidechain orientation. Some are more paying attention to the global fold,

00:43:20 then the loops wouldn't be that important. Right? So very good question. Okay? Yeah, uh, designing the assessment metric is actually a, a very tricky question. Yeah, because

00:43:34 it's easy to think that I just want to do a data science project. I just look at eventually if it's a

00:43:39 classification problem, ARC or PRC, if it's a regression problem, MAE, MSE, and then, uh, but in reality, in a lot of

00:43:48 data science projects, uh, you actually have to design the assessment metrics properly, um, unless it's already well studied by everyone. Okay?

00:43:57 Okay. Good question. So, um, then you see these ones in, uh, squared brackets, uh, just, it means that I feel like,

00:44:07 you know, for the completeness of the knowledge, I give them there, but, you know, it's, it's optional because we talked about

00:44:15 fold, right? Right? Fold is like the pattern of the, uh, 3D tertiary structure for each domain, right?

00:44:24 And we have this, uh, cath or scope, which, uh, basically defines the different, uh, domains classified

00:44:32 into different, uh, folds. You may notice that there are like thousands of folds only. So this is a concept

00:44:39 I want to, I want to talk, talk about. So this is, uh, the growth of the structure data. So, um, uh, from

00:44:48 the bottom to the top, you have, uh, the, uh, chronological order from 1970s to 2015. Okay? I think

00:44:58 I have another figure up to 2018. What you see here in red is the cumulative growths. So this is a

00:45:08 cumulative growth. What you see in the blue is, uh, uh, PDB structures deposit, the number of PDB

00:45:15 structures deposited that year. I think that the data was collected in the middle of 2015. So this blue is,

00:45:22 is not as much as the previous year. That's only because it's half, it's in the middle of 2015.

00:45:29 Okay? So blue is yearly and the red is total. Um, I think now I need to double check. Look, these structures

00:45:38 have redundancies, right? You could have one spike RBD showing multiple structures and you, uh, in complex

00:45:46 with multiple, with different antibodies. I need to double check this number. Is this, uh, is it the number of

00:45:55 protein structures that, you know, a deep mind, I have, we can double check this blog used to justify

00:46:02 its approach of data driven approach. It's saying that now we actually have 180,000 structures

00:46:10 for proteins. So it's sort of justifying the, the, the, the abundance of the data. Um,

00:46:19 um, but one thing I want you to pay attention to is, again, the number of sequences versus number of, uh,

00:46:27 data, uh, structures. Okay? Back then, the number of, uh, sequences in the gem bank is, uh, a hundred,

00:46:36 more than a hundred million back then, right? And then the number of experimentally determined

00:46:42 structures in the PDB is 100,000. So, so, um, so you may, you may feel already that these two orders

00:46:52 of magnitudes are completely, you know, differed by how many orders of, uh, magnitude?

00:47:03 Three, right? Three, right? So one out of one thousand, right? Um, so we know why, right?

00:47:12 Because of the technology to collect data, right? There's high throughput sequencing, uh, for sequence,

00:47:17 but now there's nano power even. But for structure determination, uh, cryo EM is the latest revolution,

00:47:24 but you know, it's also, you know, having its resolution, uh, challenge, and also is more applicable

00:47:30 to the, the large assemblies and so on. And not to mention it's not, uh, high throughput, okay?

00:47:36 Okay. Not in the same, same sense as sequencing. So, um, but I want you to also realize that structure

00:47:48 and the sequence are sort of in different space. When you say that two sequences are so similar to each

00:47:56 other that I don't need to include both of them in my database, I can remove one of them to remove

00:48:03 redundancy or make it, you know, clustered if you think in this way, right? So you may remove some

00:48:10 number of sequences in that case, like, you know, the total number of sequences in the, uh, uh, uniref 100

00:48:17 versus uniref 90, uniref 50, and so on. Then you need to realize that, um, it's not entirely a bad thing,

00:48:28 or it's, it's not like, you know, entirely over dominated, uh, if we see that one out of one

00:48:34 thousand protein sequences are structurally determined. Because a lot of these protein sequences

00:48:39 are actually very similar. Okay? I'll give you one number right away that if two sequences are like

00:48:48 30 percent similar to each other, then their structures are very similar to each other. So we'll, we'll see that,

00:48:53 uh, in next, uh, lecture. So in other words, the protein structure space is already much more conserved

00:49:02 than sequence. So it doesn't mean that you have to structurally determine 90 percent of the sequences

00:49:08 so that you have a glimpse of the protein universe in the structure perspective. Okay? So, so remember

00:49:15 that protein structures are much more conserved than sequence already. So this one of the other one

00:49:21 thousand doesn't mean that the current protein structure data is not well covering the protein universe. Okay?

00:49:30 The question is, uh, is there a limit? Is there a limit? Um, so, let's see. I think it's limited by the, the

00:49:40 technology and also the number of sequences we have, right? But the second limit is much, okay? But there's another

00:49:47 interesting statistic that I want you to see is the number of folds. In other words, the number of patterns

00:49:55 that each domain can adopt. Okay? Again, you have these cumulative growth in red and the, uh, the yearly,

00:50:07 you know, new fold deposited. So what did you notice? Compared to the number of structures, right? We have

00:50:16 hundreds of thousands of thousands. First of all, we only have thousands of folds, right? So we have 100 million sequences,

00:50:26 100,000 structures, but 1,000, 2,000 folds, okay? And that's one observation. The other observation is,

00:50:36 like, unlike the protein structure data, that is, uh, still looks like exponential growth, right?

00:50:43 The growth of the structure folds is saturated. So we, we haven't seen many new folds in recent years,

00:50:53 okay? So, so that's an interesting thing. And we estimate that, uh, the number of protein folds is

00:51:00 between 1,000 and 10,000, okay? So then you may ask, uh, yourself, if I don't know whether there's a limit of the

00:51:08 number of structures or even number of sequences, why do I sort of empirically observed, seemingly, I

00:51:17 empirically observed that the number of folds or the pattern of the structures of each domain, of all the domains,

00:51:26 why do I observe that to be seemingly limited, right? Um, so this is, uh, you know, a little bit, uh,

00:51:34 uh, it's still under investigation. There are different theories, but it's a product of both

00:51:40 convergent evolution and divergent evolution. It's, it's because of both. Let me explain what I mean by

00:51:49 convergent evolution and divergent evolution, okay?

00:51:52 So, uh, first of all, you know, not all folds are equally, uh, frequent. Some of them are biophysically

00:52:04 much more favored, like it's more stable, and it's created, uh, and these folds are created independently,

00:52:11 okay? And then if you think about it, uh, you know, think about each fold like a Lego piece, and then you

00:52:17 can, you know, in this case, you can modify this Lego piece a little bit so that it can adopt a new shape,

00:52:23 or combine it with other Lego pieces to adopt a new shape and adopt a new function, right?

00:52:29 So one seemingly easier way, right? I don't know whether that is easy, is how nature choose to do

00:52:36 things. I don't know, but it seems easy that if I want to adopt a new function, then I can,

00:52:43 instead of create a totally novel fold, I could exist, uh, I could modify existing folds, right?

00:52:50 Then completely recreating a new one. So in that sense, it's so-called the divergent evolution,

00:52:57 right? Right? So you're starting with sort of the common ancestor, and then, you know, um,

00:53:04 you, you, you sort of, um, see that they are branching off to different, uh, specialties, okay?

00:53:14 And then, uh, in the meantime, uh, you could see the, the convergent evolution that, you know,

00:53:22 I don't know whether you have known the fact that at the level of, uh, species, you see that you could

00:53:28 see like two kind of, uh, species resembling each other, but they don't share common ancestors. They

00:53:35 actually have different ancestors, and they can, they evolve through like different trajectories, and

00:53:41 somehow they were, they were getting very similar to each other, right? Probably due to the selective

00:53:46 pressure and other reasons, right? So that's sort of the, the convergent evolution. It's, it's, it's a,

00:53:51 it's a balance of the both. Um, our previous told, told you that a protein sequence will determine the

00:53:59 protein structure. You will, you already know that it's, it's partially wrong because you could have

00:54:05 proteins who don't take any structure called intrinsically disordered, and you also know that

00:54:10 the protein may not adopt a single structure, right? Depending, depending on the environment,

00:54:15 depending on who they talk to, depending on what they do, they may adopt different structures,

00:54:21 or even maybe the, their motion is part of their function, right? Is how the part of, how they achieve

00:54:27 their function. Then you also heard from me that a protein structure determines a shape, you know, that

00:54:33 is partially right only, uh, because, uh, you know, uh, apparently, uh, I said already part of it is, uh, not

00:54:42 just the structure, but also the motion and the interaction. And also the fact that you could have

00:54:48 two proteins with very similar structures, but they have different functions. In other words,

00:54:56 the structures will help you, you know, help you, will help you, will guide the protein function,

00:55:03 but it does not uniquely determine the protein function. So, so, okay, so I had to exaggerate and

00:55:09 simplify at the beginning, but now I think it's time to know a little bit more truth, okay?

00:55:15 Any questions so far?

00:55:18 Okay. Yes?

00:55:21 So, does that mean that you can sort of, you can figure out a way to model in 3D the proteins,

00:55:28 but that does not, like, to assume, like, what function this is?

00:55:33 Yeah, you could, you know, using the protein structures, of course, is giving you a lot of

00:55:38 insights into the 3D, into the function. Like what I said in the, in silico medicine study, once they have

00:55:45 that protein target structure, they can look at the, the structure and find out the pocket and then

00:55:50 design small molecules that target that pocket, right? But I'm saying that it's just, it's just not all of the

00:55:57 story. You see what I mean? There are, one, one way to think about it is, a protein structure, right?

00:56:04 It's like what we have in alpha-fold prediction or x-ray crystal structure. It's just a single snapshot. Do you see what I mean?

00:56:14 Yeah, imagine that, is that, imagine that I'm alien. I just came to the earth. I saw a picture of you.

00:56:25 That's what I call your structure, okay? In other words, it's your, you know, whatever you call,

00:56:34 in, in, in one time point at one, you know, spatial coordinate, right? And that particular, and, and it's

00:56:42 probably in some context, right? So if I try to figure out what is it, right? Just by looking at that

00:56:51 alone, I may, you know, I'm getting an awful lot of amount of information by looking at the photo, but

00:56:57 I can use that photo to guide my understanding, but I cannot use that to determine the understanding. So

00:57:02 that's, that's just a rigorous way to say it, okay? Okay, so, um,

00:57:15 so now let's try to summarize computationally, right? What's the implication? How to work with

00:57:22 proteins, uh, structure data? What is, uh, what is the unique challenge and how do we represent it?

00:57:28 Okay, so apparently it's a noisy and incomplete like others, right? And it's not linear as a sequence.

00:57:37 As I said that if you look at the partition coordinates of the different atoms, they are

00:57:43 intrinsically constrained by each other due to the bounded and non-bounded interactions, right? And then,

00:57:51 um, the search space is, uh, so, so it's not linear as a sequence. That's why, you know, uh, the earlier

00:58:00 bioinformatics, um, breakthroughs occurred with the protein sequence analysis because, you know, the, the,

00:58:07 the, the CS algorithms amenable to, to strings will be directly applicable. So that, you know, for sequence

00:58:14 analysis, those, uh, algorithms are readily applicable. And then in this case, uh, uh, all the atoms can move

00:58:24 in the X, Y, Z. So the search space is intrinsically continuous. Of course, if you want to simplify it,

00:58:31 you can make it to discrete as well, right? And one thing we already knew is that it's not as abundant

00:58:38 as the sequence data, right? So, so, so, so unlike the high throughput sequencing techniques, the structure

00:58:44 determination methods are labor intensive and not universally applicable to all proteins. So let's

00:58:50 look at, uh, over the years, the growth of three databases, okay? Okay, first is the UniPARC. That is, uh,

00:59:01 all the non, uh, so UniPARC basically is a number of sequence, is, I'm sorry, is a database of sequences,

00:59:09 from UniProt. It removes the redundancy by removing the identical sequences in the same species,

00:59:17 okay? That's UniPARC. And UniRef50 is removing the identical sequences, even across species, and then

00:59:26 do some clustering, like, you know, 50 is like the, the threshold for identity, uh, for, for cluster

00:59:32 membership, okay? Uh, UniRef90, I'm sorry, UniRef100 is a big, is the biggest among the UniRef. The UniRef90

00:59:41 and 50 are getting smaller and smaller. So we can look at UniPARC, which are the, you know, non-identical

00:59:47 sequences, um, at least within the, not within the same species. We don't have two identical species,

00:59:53 sequences from the same species. So this is the growth over years, right? We have two times 10 to 8,

00:59:59 okay? 200 million, right? Hundreds of million, that's what we, we learned. But if you, uh, you know, look at,

01:00:08 these, uh, PDB structures, um, so I think they should have taken a log for the y-axis, right? Because I

01:00:16 really cannot see very clearly what's the growth, but we know that it's about 100,000, okay? 10 to 5.

01:00:22 So we are talking about 10 to 5 versus 10 to 8, okay?

01:00:28 Okay, so yeah, I think the number is 180,000 now for, uh, the structures in the PDB, okay? So, so this is one thing.

01:00:37 But again, I said that, uh, you know, you also have to realize that the structure data, structures are

01:00:42 more conserved than sequences. So you cannot just compare those two numbers, okay? Uh, so however, uh,

01:00:52 uh, we have observed a lot of, uh, deep learning technologies are getting, uh, more and more

01:01:01 inspirations and applications as well, of course, into the field by studying the 3D structure.

01:01:06 Now that AlphaFold has, uh, predicted so many structures for proteins with reliability,

01:01:12 the importance of the structure data is just increasing. So you've got to be able to know how

01:01:17 to work with the structure data. So when working with the structure data, you need to respect the

01:01:24 physical and biological principles. But sometimes you don't know all of them or all of the details.

01:01:30 For instance, you know that there's a Schrodinger equation, right? But you cannot really solve

01:01:36 that for all the atoms within a protein, right? Or in some other cases, um, uh, you know, for instance,

01:01:48 AlphaFold2, uh, I remember the, you know, their CEO, uh, DeepMind CEO, when, when giving the talk at, um, uh,

01:01:57 Casp 14, the conference, uh, back then they couldn't talk about all the details, but they can give a,

01:02:03 uh, because the paper was not published yet. Uh, so, but they, uh, the, the CEO did emphasize that

01:02:10 they did not just apply AI to the protein-folding problem. They made AI tools for the protein-folding

01:02:17 problem, and they really respected the physics. And, uh, so, you know, people in the audience did try to

01:02:24 ask again and again, what's the physics, right? So it's not exactly the, you know, what we talked about,

01:02:29 bounded interaction, non-bonded interaction, that kind of physics. What, what he emphasized

01:02:34 essentially is the fact that, you know, if you rotate or translate a protein structure rigidly,

01:02:41 the structure does not change, okay? So if you have a neural network whose input includes the X, Y, Z

01:02:48 coordinates of all the atoms and the output is updating it, then you know that, you know, if, if,

01:02:55 if all the coordinates rotate or translate together, then it's the same quality. So that's, uh, why they

01:03:03 use the roto-translational equivalent neural networks in part of their, uh, model. And sometimes because,

01:03:12 uh, our eyes are so much better at finding the patterns, to see is not to believe, but to see is

01:03:17 to discover. So that means the data visualization and analysis actually matters a lot. I remember IBM has,

01:03:24 has a branch dedicated to data visualization because there's a lot of research about visualization as

01:03:30 well, right? Um, how many of you have used the T-SNE? T hyphen S-N-E, right? So that's, that's, you know,

01:03:43 that's a dimensionality reduction sort of visualization tool, right? And there's a good science behind it.

01:03:48 And, uh, how to represent the protein structures. Remember, you talked, you told me what's the data,

01:03:56 all right? But I need to feed this. Imagine that this is the data I want to learn from,

01:04:01 okay? I need to feed this to, uh, algorithm. Now more likely it's a neural network, right?

01:04:07 How do I feed the data, right? You saw the file with a PDB format, right? How do I feed the data

01:04:15 to a neural network? So in other words, how do you represent the protein structure data? So that's the

01:04:21 next couple of slides we're going to focus on, okay? Roughly, you can say that, uh, you can represent

01:04:27 the protein structures in 1D, 2D, or 3D, but there are way more than that, okay? For instance, what do I

01:04:35 mean by 1D? A 1D is like saying that, uh, it's for instance, right? Uh, we, we do have the sequence,

01:04:43 that's a primary structure, but, you know, it's more often called the sequence. But then you could

01:04:48 have, uh, you know, what it's saying is, uh, alignment exemplified by five sequences. This is

01:04:56 a multiple sequence alignment. So this, each column, 1, 2, 3, 4, 5, are like five siblings of these

01:05:03 protein sequence, P, Q, I, T, and so on, okay? Then you sort of use the sibling sequences, or relative

01:05:12 sequences, and align them together. Then you figure out that every position is not just characterized by

01:05:18 a letter, but a distribution of 20 letters. That's called the profile, right? It's called 1D because

01:05:26 it's for every residue. Then you could have this one. You have learned before this, uh, sequence is

01:05:34 the secondary structure. E stands for a beta strand, right? Then it's missing, it probably means it's loop,

01:05:39 and there's no helix in this portion, okay? And then there's some other numbers, uh, less about the

01:05:47 solvent accessibility. So it's a continuous number, but we could discretize it, right, into buried, um,

01:05:54 half buried, and exposed, right? So in other words, these are like the marginal distribution of the protein

01:06:01 tertiary structures, okay? In the one-dimensional space. You could also represent it in the two-dimensional

01:06:08 space. This is sort of the two-dimensional contact map. So instead of looking at every residue, you look at

01:06:16 every pair of residues. So you have the index 1 through 100, and index 1 through 100. So residue, and then at

01:06:24 in each entry of the matrix, you have the entry ij, that means the pair of residue i and the residue j,

01:06:33 you have the other one-dimensional space. And then if you see a darker color, it means that they're closer

01:06:38 in 3D. If you see a lighter color, you mean that it's further away in 3D. You can even binarize it into

01:06:46 two colors. White means, you know, in contact less than, for instance, 8 or 10 Anstrom. Black means, uh,

01:06:53 sorry, yeah, black means in contact, and white means not in contact, if the distance is bigger than 10

01:06:58 Anstrom, for instance. Okay? Then you can also do it in 3D, but I want to talk a little bit more about what

01:07:06 what the 3D is, okay? Now, uh, knowing what this so-called contact map means, it could be binary, right?

01:07:15 Then each, um, each pixel can be a binary. So, so, so, so, so that's why, um, uh, after deep learning gained its popularity in,

01:07:26 you know, ImageNet, one of the first thing that flew into bioinformatics was the study of the contact map.

01:07:34 Either you predict the contact map, or you start with the contact map and predict something. Why?

01:07:41 This is an image, isn't it? Right? So, it's a grayscale image, right? So, it could be binary, as I said,

01:07:49 but this could be grayscaled from, from, you know, white to black. Um, each pixel could be a

01:07:57 probability mass function, instead of a single number, right? For instance, the chance that the

01:08:04 distance is between 3 and 5 is 1 over 50. The chance that it's between 5 and 6 is 1 over 40.

01:08:12 Anyway, they add up to 1, right? So, each pixel could be a binary number,

01:08:17 could be a probability mass function. It could be also a real value, the exact distance, right?

01:08:24 Or you approximate this distance as a Gaussian distribution, then you can use two numbers to

01:08:31 characterize it, like, like VAE. One mean and one standard deviation, right? Right? So, so, so, so,

01:08:38 those are different kinds of 2D, but in, in, in sense, they are images. So, that's why, you know,

01:08:44 one of the breakthroughs actually popularized by AlphaFold 1, okay, three or four years ago,

01:08:51 but that was already accomplished in academia. That was, you know, treating this,

01:09:00 realizing this contact map as images, and the treating the contact map prediction problem.

01:09:06 You started with one dimensional sequence, and you predict the two-dimensional contact map,

01:09:11 you treat, regard this one as like a pixel level, you know, this labeling problem, okay? So, this is one,

01:09:18 and then the corresponding neural network architecture was ResNet by then. So, you see that,

01:09:24 it's getting, you know, all the nutrients, but also, some unique factors is here is that,

01:09:35 you also have to realize that this is not the, the natural images. There are some special things

01:09:40 you have to, to deal with, okay? So, I don't want to talk more details into that. When we get there,

01:09:46 we'll see, especially in the journal club.

01:09:48 So, now I want to ask one question, let's see whether we can get that. So, this is the diagonal,

01:10:02 i equal to j, right? So, it's always in contact because i equal to j, okay? So, now let's see,

01:10:10 let's try to draw something. Can I draw it? No. Okay. So, if you see this, if I cannot draw it,

01:10:30 if you see these short stretches, which is parallel to, very close to i equal to j, but not exactly at i

01:10:39 equal to j. I can tell you that they are like i equal to j plus four or i equal to j minus four,

01:10:49 okay? They are parallel to i equal to j, very close, like four residues away from the diagonal line,

01:10:57 some short stretches, okay? What do you think they are?

01:11:03 What kind of secondary structure is that?

01:11:05 Alpha helix. Alpha helix because there's a hydrogen bond between residue i and i minus four.

01:11:12 So, if the absolute value of i minus j equal to four, that indicates the second, the alpha,

01:11:19 okay? So, this is in the short stretches in parallel. Okay, what about these ones perpendicular to the

01:11:26 diagonal? It means that i plus j equal to something.

01:11:31 Beta sheet.

01:11:38 What kind of beta sheet?

01:11:45 Anti-parallel, right? Because if you have parallel beta sheet, you would also have i minus j equal to a

01:11:53 constant, but that constant is way bigger than four. So, it's going to be somewhere here, for instance,

01:11:59 right? That distance, if it's bigger than four, it shows what's the length of the loop connecting the

01:12:06 parallel beta sheets, right? But when you have i plus j equal to a fixed number, that usually means it's

01:12:14 the anti-parallel beta sheet, right? Is that easy to understand?

01:12:20 Let's see. Let's look at the,

01:12:29 for instance, anti-parallel.

01:12:33 Yeah.

01:12:35 Like in this case, okay? Let's call this position i, and then the length of the loop is d, and then this

01:12:42 position is hydrogen bounding with i, so it's i plus d plus one, right? And then this, the next pairs are also

01:12:51 interacting. So their indices are i minus one here, and then i plus d plus one, right? This is i minus two,

01:13:01 this is i plus d plus two. So there are some, right? Right? So, right? But where in this case,

01:13:09 right, if there's an i and a j, right, then the next one would be equals to, for instance, if I keep

01:13:18 going that direction, it will be i plus one versus, you know, that j is going to be dependent on the length

01:13:26 of the loop. So it's always going to be a constant. I minus j is going to be a constant, okay? It's separated

01:13:34 by the length of the loop, right? Right? So that's why we see, we can see the parallel and anti-parallel

01:13:46 secondary structure is there, okay? So I haven't finished the representation of a 3D structure.

01:13:54 So point cloud is straightforward, but it may not be the easiest one to learn from, okay? You can say

01:14:00 that, well, let's just use point cloud and 3D CNN, but that's not the straight, naturally the easiest

01:14:06 one. Why? Because as I said, so many high dimensional space with a lot of constraints, hard to learn. If you

01:14:12 ask a neural network to learn the constraints about the bonding and non-bonding interactions, that's very

01:14:17 difficult. So what people usually do is, of course, you can use point clouds, but you can also pay attention to the

01:14:24 the fact that protein interactions are usually at the surface. So you could start, you could, you know,

01:14:32 sort of study the surface. The surface is one kind of non-Euclidean structure, right? So Michael

01:14:39 Brownstein is really interested into this kind of representation and the corresponding neural networks,

01:14:45 right? You could use a volumetric contact map already talked about, tessellation. You can also use a molecular graph. If

01:14:54 two atoms or residues are close to each other, then there is an edge between them and each node is an atom or a residue.

01:15:01 So you can represent it as a molecular graph.
