Lecture 3: Protein Structure Concepts
Recording. Download the text transcript.
Automatic transcript; technical terms and equations may contain recognition errors.
00:00:21 Okay, let's get started. Howdy. Okay, so welcome to the second week of ESSEN 766.
00:00:34 So in the previous lecture, we talked about the chemistry and biology underlying the biomolecules
00:00:46 that carry information and whose data need to be learned, right? So that's what bioinformatics is about.
00:00:56 So I think so far you should have a better idea about the chemical basis of these molecules, right?
00:01:05 So the reason we are talking about this is the following. Why do we think that the DNAs,
00:01:11 RNAs, and proteins can be represented in the one-dimensional text? The fundamental reason is
00:01:18 that they're all biopolymers. They are all like repeating units, right? Nucleotides or amino acids
00:01:27 linked together in the chain, right? And what links them together? The covalent bonds, right?
00:01:33 So in other words, if you were to use a language model to learn the texts written in the protein language,
00:01:46 right, or nucleotide or RNA language, right, what you will not learn is chemistry. You will not learn
00:01:54 chemistry out of it because you can have any kind of bonds, covalent bonds between any kind of neighboring
00:02:00 amino acids or nucleotides, right? Then if that is the case, what are we learning
00:02:09 by using, for instance, a language model on protein sequence data? What are we learning?
00:02:14 If it's not chemistry?
00:02:16 It's the distribution for the next period.
00:02:19 The distribution. What determines the distribution if it's not chemistry? What determines the distribution?
00:02:25 What is the distribution? Good point.
00:02:28 Based on our training data, what will be the highest probability for the next one?
00:02:35 Right. So what is underlying such distribution which is to be learned by any kind of model?
00:02:46 Underlying rule is not chemistry as we said that, you know, any kind of covalent bonds, right,
00:02:50 can link any kind of pairs of nucleotides or amino acids. And that means what you are learning is sort of,
00:02:57 you know, the biological distribution. Is that right? Because all the sequences in your training sets are
00:03:03 protein sequences that have been sampled by nature, right? So how did nature out of all these
00:03:09 20 to the power of L, where L is the length of the protein length, out of all these, you know,
00:03:15 astronomical number of protein sequences? Only some of them are sampled, right? So in other words,
00:03:22 you can imagine there is a fitness landscape of all these protein universe. And then somehow we are,
00:03:29 what we have seen, right, in the training data are probably the sequences that have been evolved and
00:03:37 survived so far, right? So something to keep in mind. Whenever you do data science, you first have to
00:03:43 ask yourself what data, what is behind the data, okay? Because at best, you can mine from the data is,
00:03:53 you know, the information encoded in the data. So if you know what to dig, then it might help you to
00:04:00 come up with a better model to dig such information, okay? And we also said that these texts are actually,
00:04:08 you know, ordered, right? So some alphabets, human language, are written left to right or right to left.
00:04:16 Here we say there is an order. Why? Well, the order is largely because of the chemistry. Again, you know,
00:04:24 in the case of the nuclear, in the case of the RNAs and DNAs, right,
00:04:30 the order is going from the five prime end to the three prime end, right? If you remember, there is a
00:04:37 there's a central sugar, right, called deoxyribose or ribose, right? Then there's, you know, if you go
00:04:46 around that five-membered sugar, there are three prime and five prime position, right? Right? Five prime is
00:04:53 trying to grab the phosphofate group, right? The three prime is OH group, right, which is connected to the next
00:05:00 phosphate, right? In the case of the proteins, you go from the N-terminus to the C-terminus because,
00:05:06 you know, in each amino acid, there is a central carbon, right? And then, you know, it's connected to four
00:05:12 partners, right? Two of them are
00:05:15 amino group and the carboxyl group, right? And this and the first one is always a hydrogen.
00:05:20 And the third one is the one that differs the 20 different amino acids. That's the side chain, right?
00:05:25 So there is an order as well going from the N to C-terminus, right?
00:05:29 I'm saying this is because, you know, even as recent as our own
00:05:33 language models research lately, you know, sometimes we have to use
00:05:37 the forward and backward models to learn both the five, the N to C and C to N, you know, especially
00:05:44 if you are familiar with auto-regressive language models, which are
00:05:48 modeling the distribution of Xi, where Xi is the ice positions of a letter,
00:05:54 given X1, X2, all the way to Xi minus one, right? Given all the previous
00:06:02 positions, choices. So that's auto-regressive model. So in that case, the order matters,
00:06:08 but, you know, biologically, you think you would rather go from N to C alone, but,
00:06:14 you know, empirically, because of the limitation model, sometimes you may find it is also useful
00:06:18 to go from the other direction and use both models. So certain things to think about.
00:06:25 We also talked about, you know, the four levels of protein structures.
00:06:30 At the primary level, it's actually just a protein sequence, right? So it's like, you know,
00:06:37 a beads, amino acids threaded together, right? And there are 20 different colors of beads, and
00:06:43 some of them are, you know, positively charged, some are negatively charged, some are not charged, but polar,
00:06:50 and some are hydrophobic, who like to, you know, mingle with each other and hide from water. So they're
00:06:56 usually hiding the core of the protein. Because of all these, you know, different characters of amino acids,
00:07:05 because of the non-bonded interactions between them,
00:07:09 together with, you know, covalent bonds that thread them into one chain, these proteins actually fold into
00:07:18 a 3D structure, right? So locally, you have these, you know, alpha helix and beta sheets,
00:07:26 which are local structures. I didn't explain what they mean exactly. So that's part of the concept today.
00:07:34 And then globally, these secondary structures are going to pack together into a global tertiary
00:07:39 structure. So usually, when we talk about protein structure prediction, we're talking about tertiary
00:07:44 structure. And we call the primary structure the sequence, okay? The multiple chains or multiple
00:07:50 proteins may interact with each other. In this case, two copies of molecule A and two copies of molecule B
00:07:59 come together to form a hemoglobin that carries oxygen in your blood, okay? So, and then, similarly,
00:08:11 there are, you know, secondary structures for RNAs as well. So I'm not giving more details into it. But pay
00:08:18 attention that unlike proteins where you have the non-bonded interactions, like van der Waals,
00:08:25 electrostatics, hydrogen bonds, and so on. In the case of the RNA or DNA, you know, I said that the RNA
00:08:33 double helix, this, you know, double helix, why you have a twisted double helix, that's because of the
00:08:39 hydrogen bonds between the nitrogenous bases across the chains, right? So, but the RNA is usually a single
00:08:46 chain. But then they form a very sophisticated secondary structures, again, because of the hydrogen bonds
00:08:53 between these nitrogenous bases. In other words, we call them base pairing, A with T, C with G. So,
00:09:01 so in other words, it's not a surprise at all that the, you know, one of the earliest
00:09:09 biomolecule text generation or generating models to design molecules first occurred with RNAs, okay?
00:09:18 Because they are a little bit more structured in the pattern because there is a very strong base pairing
00:09:24 pattern that determines the underlying secondary structure, whereas in the case of the proteins,
00:09:29 not only they are much longer, but also the non-bonded interactions are much more sophisticated than
00:09:35 the base pairing pattern in RNAs, okay? That's something to pay attention to. And as I said that,
00:09:41 because as you go from the sequence all the way to individual tertiary structure or the
00:09:49 multiple proteins quaternary structure, the experimental technologies throughput is just decreasing
00:09:58 exponentially. So that's why there is a goal of starting from the sequence and predicting the structure
00:10:07 of individual proteins. That was the breakthrough made by AlphaFold last year. And they are now trying
00:10:14 to solve this, they already have one paper already in bioarchive, predicting the complex protein, protein
00:10:22 complex structures, okay? But inversely, you can also say that, well, in this case, you can only tell me to
00:10:31 predict how a nature sampled sequence may fold into a three-dimensional structure or interact with its
00:10:41 partner in three-dimensional space. What about those large space of sequence universe that hasn't been sampled by nature yet?
00:10:51 So inversely, we have the design task that, you know, you may want to start with the ideal function or structure of a protein,
00:10:59 because we said, you know, a naive assumption is that the shape or the structure of the protein determines
00:11:06 its function. So if I say that I need a protein that can do this function extremely well and that function can be encoded by this shape,
00:11:15 then I would start with this shape and try to design a sequence that optimally adopts that shape, right?
00:11:22 So that's a design question, that's an inverse question, right? Or in the case of the drug design,
00:11:27 you want actually your drug molecule to bind a target molecule in the human body very well or bind it to the virus, like the COVID-19 spike protein, right?
00:11:41 You want antibodies that can neutralize, right? Those virus proteins. In those cases, you start with the protein target in human or in virus,
00:11:51 and you try to design some sequence or small molecule formula that can optimally bind or neutralize those targets.
00:12:01 So these are the questions. Some of them will be addressed, but I will focus mostly on structure prediction and design and use this as
00:12:12 focal examples to show you how we address, don't forget our goal, multimodal data learning and integration, right?
00:12:21 So that's our goal. So today I still need to, before, you know, this week we focus on the protein structures,
00:12:29 and I still need to spend one more lecture to talk about the concepts before on Thursday I will talk about the structured data.
00:12:38 Okay, so bear with me because there's still a lot of the concepts to be covered before we can address the structure data, okay?
00:12:47 So the first concept is the secondary structure. You already have seen it, right?
00:12:52 But I didn't really explain, right, how come you have a helix and you have a beta sheet, and what does this helix or beta strand or sheet means,
00:13:01 and why do they matter, okay?
00:13:03 So the secondary structures are, you know, due to the bounded and non-bounded interactions between the residues, right?
00:13:11 So the secondary structures, right? You have seen these local three-dimensional structures, okay?
00:13:16 So alpha helix, what we are looking at in a cartoon representation is like this.
00:13:22 What we are showing is a cartoon representation of the backbone. Remember, the backbone does not consist of a side chain.
00:13:30 It only consists of the central carbon, amino group, and the carboxyl group, right?
00:13:36 So this cartoon representation of the helix actually corresponds to a lot of atoms just interacting with each other through covalent bonds or non-bonded interactions.
00:13:51 So bear with me.
00:13:54 So I want you to focus on this first, this line. If you see this line going all the way, right?
00:14:02 So remember, it's a one-dimensional chain of polymers. So these backbones actually trace in this line, okay?
00:14:15 And you see that this line is sort of giving you a helix impression, is that right?
00:14:22 Right? So I'm going to show them in three-dimensional space in this figure, and at the end I will overlay
00:14:27 these two views of cartoons and ball and sticks, okay?
00:14:32 So let's look at the ball and sticks where each ball is an atom, or except the yellow one, okay?
00:14:38 Let's forget about that for the moment.
00:14:41 So I want you to pay attention to these black spheres, which are carbons, okay?
00:14:50 And let's look at these carbons, right? So for instance, this carbon, right?
00:15:01 So it also have different colors, like nitrogen, which is blue, and oxygen, that is red, and hydrogen is
00:15:10 oxygen is green, and hydrogen is red. It's not a typical choice. Usually you use red for oxygen for
00:15:16 apparent reasons why, because of the color of the blood, right? Oxygen is carried a lot in the blood, and circulated through our body.
00:15:24 So let's look at these atoms, right? So this one is, let's see, without a blocked view,
00:15:32 without a blocked view, so let's look at CO. Okay, so this is a central carbon, right? So this is a central carbon,
00:15:50 it's connected to a side chain, right? And then this is the CO, we are losing an OH, right? So this is a carboxyl group,
00:16:02 right? And on the, it's also connecting to a hydrogen, right? And then the fourth one here is connecting to an amino group, right?
00:16:14 So in other words, we see a single residue or a single amino acid, right? Now we can keep going. In this case,
00:16:22 you know, we have these,
00:16:26 again, this
00:16:28 amino group, right? Central carbon, and then this, the red one is a carboxyl group, CO, right? And then this is a hydrogen, right?
00:16:39 So in other words,
00:16:41 can you tell me
00:16:43 which direction is the carboxyl group, which direction is, sorry, which direction is a carboxyl C-end,
00:16:50 which direction is the amino N-end, N-terminus? Can you see that?
00:17:05 Okay, we can leave that question later on, but you may wonder, okay, why isn't this
00:17:12 chain, one-dimensional chain, be stretched out? What makes them forming this helix?
00:17:23 Okay, you might notice that something, this kind of, you know, this sticks, this solid sticks are
00:17:29 covalent bonds, but we also have these, you know, red, these
00:17:34 dashed lines, showing non-bonded interactions. It's exactly the non-bonded interactions makes this
00:17:42 one-dimensional chain forming into a three-dimensional helix. Okay, let me give you one example.
00:17:49 Right? This is oxygen, the green one, oxygen. This is on the carboxyl group of one residue, right?
00:17:57 Then here, what you have is, you know, the R group, and then there is a, a better view is here, probably.
00:18:06 Yeah, let's look at this oxygen, right? The green. It's forming a interaction, let me call it interaction
00:18:14 for the moment, with a hydrogen attached to nitrogen. So this is the amino group, right? So in other words,
00:18:25 the oxygen on the carboxyl group of residue I is interacting with the hydrogen on the amino group
00:18:35 of residue J. That is a special interaction called hydrogen bond, remember, right? And the hydrogen bond,
00:18:45 as we see, that is, you know, contributing to a lot of the special properties, like in the water,
00:18:50 because there are so many of them, right? So if you see that there are many of such hydrogen bonds,
00:18:57 all right, along the axial direction of the alpha helix, or along the chain. So those hydrogen bonds
00:19:04 are the reason that this, you know, one-dimensional chain is not stretched, but actually forming a
00:19:10 three-dimensional helix. Okay? So now let's overlay these two views together. We have the half-translucent
00:19:20 cartoon representation of the helix with the ball and the stick representation of these helix together.
00:19:28 And you can see clearly what we are showing in the cartoon is sort of, you know, following the direction
00:19:37 of the backbone, isn't it? Right? You see that, right? This is our backbone, right? And the reason
00:19:45 that you have this helix is because of, you know, these hydrogen bonds formed between residue I and residue J.
00:19:54 And as I said, it's a carbon-neur oxygen. That's oxygen on the carboxyl group of residue I.
00:20:01 And the amide proton, that's a hydrogen on the amino group of residue J. But there is a pattern.
00:20:11 The difference between I and J is very regular. The difference between I and J is always four.
00:20:17 Okay? So because of that reason, you have a very patterned alpha helix. So if you look on the top,
00:20:25 you will see that, you know, every 360 degrees, like every turn, involves 3.6 residues.
00:20:34 Okay? So that is the reason that you have a very patterned alpha helix. Again, we see that hydrogen bonds is playing a major role, forming this alpha helix.
00:20:43 Okay? And this is another illustration, right? So again, you know, this is probably a clearer view.
00:20:52 In this case, we have the standard coloring where the oxygen is in red and nitrogen is still in blue, right?
00:21:02 So this is, you know, central carbon, right? This is also numbered from 1 to 9. So this says that there are 9 amino acids starting from 1 to 9.
00:21:19 Okay? So for instance, here you have the C-alpha of the first residue, and then you have, you know, these COOH.
00:21:32 That's the carbon-neur group of, or the C-terminus of the first residue, right?
00:21:37 Then it's bonded with the second residue's amino group, this blue side nitrogen, right? Amino group.
00:21:45 So this is the peptide bond connecting the two neighboring residues. So if you follow this, you see this nice helix.
00:21:56 Okay? So this gives you a better view because I put the numbers there as well. Okay?
00:22:02 These numbers are the index, right? Then you see that, for instance, this is the oxygen on the carboxyl group of residue 5, or so-called carbon-neur oxygen.
00:22:16 So this is the oxygen on residue 5, which is actually forming a hydrogen bond with the hydrogen on the amino group of residue 9.
00:22:27 So in this case, the hydrogen is in white. This is more standard coloring that oxygen is red and hydrogen is in light gray or white.
00:22:38 Okay? So this is another illustration. And then if you remove, I was asking one question earlier. I was asking what's the direction of this protein chain?
00:22:53 Which direction is N-terminus? Which direction is C-terminus? Can you see that?
00:23:08 So it's from 1 to 9, right? So even without the numbering, you see that, you know, you start with amino group and end with a carboxyl group.
00:23:18 Amino carboxyl, amino carboxyl, amino carboxyl, right?
00:23:23 And if you follow that direction, what you see is that the polarity in the one-dimensional space, or in this case, three-dimensional case, is going up from the first to the last residue.
00:23:34 Is that right? Right? And then if you look at this one, what do you see?
00:23:39 What does this remind you? When you learn, you know, the electromagnetics,
00:23:45 what you have learned is something called the right-hand rule, isn't it?
00:23:50 This is actually a right-hand alpha helix.
00:23:55 Because if you, you know, if you do a right hand like this, your four fingers are following the direction of the backbone,
00:24:02 and your thumb is following the direction of the, you know, N-term to C-term. Did you see that? Right?
00:24:08 So it's alpha helix, right? You may wonder whether there are other kind of helix.
00:24:14 Yes, there are a lot of other kind of helix in nature. But there are reasons that they are less sampled by nature.
00:24:22 So almost all helices found in proteins are right-handed.
00:24:28 The reason is that what is not shown very clearly is the side chain. For instance, I never showed the side chains very clear.
00:24:34 This one threw away the side chain, right? In the previous figure, I showed the side chain as a single collapsed sphere.
00:24:42 Right? But remember, a side chain consists of multiple atoms, right?
00:24:46 So if you want to form a nice backbone structure, you're going to make sure that the side chains,
00:24:52 not only you have these attractions like hydrogen bonds between the backbone atoms,
00:24:57 you also want to make sure that the side chains, you know, which are group, which are big groups of atoms,
00:25:03 which are very similar to the side chains, which are very similar to the side chains.
00:25:08 So that's the reason that in nature, you see most of the, almost all helices found in proteins are right-handed,
00:25:18 because, you know, the steric restrictions imposed by the side chains in the left-handed helices.
00:25:26 Okay? And there are other right-handed helices like 310 and other things, but alpha is the most prevalent ones.
00:25:37 And there are databases of the secondary structures such as DSSP and so on. Okay?
00:25:43 So before I go to the secondary structure with the beta sheet, you may think about this.
00:25:49 You know, secondary structure, what's the big deal of the secondary structures?
00:25:53 So, to be honest, in the early days, when you want to make machine learning predictions for, you know, protein structures,
00:26:02 this kind of secondary structure prediction is almost the first thing they started. Okay? I will talk about it later on. Why?
00:26:10 Because it's not 3D prediction, actually. It's only 1D, right? Because every position you can say,
00:26:17 oh, all these nine residues are in alpha helix, right? Then maybe the next five residues are in beta sheets.
00:26:25 So I can predict a sequence of secondary structures, which is 1D, which is not that 3D. So it's,
00:26:33 it's, in other words, it's like machine translation, isn't it? Right? You start with a text written in the
00:26:39 alphabet of 20 letters, and you translate it into a text of secondary structures.
00:26:44 Depending on how many secondary structures you have, if you have only alpha helix, beta sheets,
00:26:50 and coils, or loops, then you have an alphabet of three letters for secondary structures, right?
00:26:57 And it's like machine translation from, you know, one language to another, isn't it? Right? So,
00:27:05 there's no wonder that the secondary structure prediction was one of the first application of
00:27:10 machine learning in the world of protein structure prediction, okay? Is it still being done now?
00:27:19 They are actually used more and more often as features or input to your machine learning models,
00:27:24 because they carry additional information, okay? So the second, let's see, yeah. So the second,
00:27:33 if you think about there are three major secondary structures, right? The last one is just the
00:27:37 coils. Then the second one is a beta sheet. So, usually we draw them like this kind of, you know,
00:27:45 zigzagging sheets, and then they are like, you know, in parallel, or anti-parallel in this case, right?
00:27:53 With arrow directions, right? You probably figured out what those arrows are saying. It's basically showing the direction
00:28:00 from N-terminus to C-terminus, okay? But how could I draw a cartoon representation like this? What are the
00:28:08 underlying, you know, 3D structures, right? Because eventually we want to predict the structures.
00:28:14 So again, I use the same kind of illustration. What you see here is if you follow the black sticks,
00:28:20 these are the covalent bonds that are forming the whole backbone. You see these black sticks, right?
00:28:31 These sticks are covalent bonds forming, connecting the amino acids. And if you look at the up and downs,
00:28:39 or in the left and the right, or bends and twists of these black sticks, it's essentially
00:28:46 what this cartoon representation of backbone is showing. Is that right? So this cartoon representation
00:28:52 is actually showing this, you know, backbone structures, okay, in an ideal way, okay? So it is indeed
00:29:01 zigzagging, right? Right? Zigzag, zigzag, zigzag, zigzag, right? Just like what you see here, zigzag, zigzag, zigzag, zigzag,
00:29:10 right? But what is underlying, what is underlying, why wouldn't they form a nice one-dimensional thread,
00:29:19 and they would rather form something like this in 3D space. Well, let's first, you know, look at,
00:29:28 you know, here you have multiple beta strand, and together, the strands are anti-parallel and forming
00:29:36 a sheet. So let's talk about apparently the reason is that there's something
00:29:41 going on between the neighboring strands, right? Some neighboring strands that
00:29:48 instead of you have a first segment of the sequence
00:29:55 going this way but instead of keep going that direction you would rather make a
00:29:58 turn and then forming this you know neighboring beta strand. Why is that?
00:30:05 Again I want you to pay attention to those not the sticks anymore which shows
00:30:14 covalent bonds but those dashed red lines which are showing the non-covalent
00:30:21 interactions in this case specifically hydrogen bonds. What you see here is
00:30:26 again sorry there are two coloring schemes but nitrogen is always yellow I'm
00:30:32 sorry always blue and in this case you know nitrogen is connected to a is
00:30:38 bonded with a hydrogen somehow is in red in this case and this hydrogen is making
00:30:44 a hydrogen bond with an oxygen in the neighboring strand. So in other words
00:30:50 again there are hydrogen bonds being formed between these you know hydrogen in the
00:30:58 amino group and these oxygen in the carboxyl group but unlike the case of the
00:31:06 alpha helix where the you know you see that these hydrogen bonds are being formed
00:31:13 you know along this whole single thread of the alpha helix but in this case of beta sheet you
00:31:21 actually have the hydrogen bonds not between a you know adjacent segments sorry actually you
00:31:30 know not within each segment so you know if you look at the alpha helix right within
00:31:35 this segment this is alpha helix within this segment we have the hydrogen bonds stabilizing the
00:31:41 segment segment but here we have the three segments and the segments were put
00:31:47 together anti-parallel because of the hydrogen bonds between the adjacent
00:31:54 segments so that's the difference between the alpha helix and beta sheet they are
00:31:59 both stabilized by hydrogen bonds but in one case is stabilized by hydrogen bonds
00:32:04 within the segment that's our helix in the other case is stabilized by hydrogen bonds
00:32:09 between adjacent segments that's a beta strand okay what I didn't show here if you you know if
00:32:16 you go this way and then before you enter the next segment there are some loop very loose
00:32:23 non-stable secondary structures here as well okay I just didn't plot it if you wonder how come you have
00:32:31 a broken you know protein right actually it's not broken we don't we just showed that the portion of the beta
00:32:39 sheet but if you start from the N terminus you go like you know first and then there's a loop and then the second beta strand
00:32:48 then there's another loop and then you enter the third beta strand and then eventually go to the C terminus okay
00:32:54 this is due to R group not hydrogen
00:32:58 so that's a good point I haven't talked about bending yet I only talked about so far why the adjacent segments like to stick to each other because of the hydrogen bonds right
00:33:09 right but if you know someone asked asked a good question is why are they zigzagging in this way
00:33:18 why do they have to zigzag it's a it's a fantastic question if you now overlay the translucent cartoon representation
00:33:28 and the stick and ball representation of the protein right and ask yourself how can I place the very bulky multi atom
00:33:41 side chain these you know currently represented like a yellow sphere but remember it consists of multiple atoms connected to each other right
00:33:50 some of them can be lengthy right so again like what I said why there are few you know left-handed helices in nature
00:34:00 same here that the conclusion was that you know the the side chains right being able to adopt side chains in 3d space without clashing them with each other or the backbone was a major reason of the dominance of the right hand helices here it's the same that if you notice
00:34:19 if you notice this zigzagging lead to a one pattern that is the orientation of the side chain here you see it's like the side chains are always at the ridge of the zigzagging
00:34:33 and here is pointing inside here is pointing outside here is pointing inside here is pointing outside inside alternatively isn't it
00:34:42 right so it's really a nice pattern you know you don't you don't know who figured this out it's really nice pattern to to because you want to use the building block of you know building blocks of 20 amino acids and make sure that they can adopt different sequences and fold into different shapes and do different kind of functions right but how can you make that mechanism efficient right it's like a very sophisticated Lego right so in this case you know this kind of
00:35:12 secondary structure is in this shape because of the side chains is you know is affecting the turns in this case the zigzag of the backbone okay so that's that's the reason okay and in this case you probably noticed that I said before you know you have to put the loops in between and in this case the neighborings are anti parallel
00:35:42 right but you could have to put the parallel right but you could have this kind of parallel beta sheets as well you know if you were to design these proteins would you rather use more anti parallel or parallel beta sheet or do you think it doesn't matter what do you think
00:36:03 if you're if you're if you're if you're if you're if you're if you're if you're the designer of the proteins in nature and you know you want to use this beta sheets right and you can arrange them in this anti parallel way the neighboring segments are anti parallel or this parallel way
00:36:19 maybe anti parallel because
00:36:21 we have less unstable
00:36:24 stability is one thing that you're hitting that is good right yeah so so so for instance if you if you look at this one this parallel case did you notice that you have to make a very long term
00:36:40 a very long term that is occupying a lot of space right so the stability and and and efficiency is one concern okay um so
00:36:54 so uh you know there are more anti parallel than parallel uh beta sheets but it's not like the dominance of right-handed helices okay there there there are still quite some uh anti parallel uh i'm sorry quite some parallel beta sheets because that's not the only concern yes questions
00:37:13 right right right right good question did you notice that each segment has a direction the arrow shows the direction that goes from n terminus to c terminus the first letter of the protein sequence to the last letter of the protein sequence right
00:37:32 so if you look at the arrow or the arrow or the directions of the neighboring segments in this case it's anti parallel right and in this case it's parallel
00:37:47 uh it looks like helix to you
00:37:49 because they have the connection
00:37:50 oh i see i see what you mean yeah i think it's called a beta sheet because you because of the the neighboring beta strands together form a very flat sheet
00:38:00 So that's you know unlike the the alpha helix
00:38:03 okay
00:38:04 yeah again when we are talking about our beta sheet we are not looking at the turns
00:38:09 okay the turns are the third secondary structure which are just turns or loops very unstable alpha helices and the beta sheets are more stable
00:38:19 okay
00:38:20 okay
00:38:21 All right so now that we have learned so many concepts i gonna i'm gonna ask you you know how do we analyze this data imagine that this is the first time that
00:38:29 first time that protein structure is discovered,
00:38:32 and you found even that there are some patterns
00:38:35 of alpha helix and beta sheaths, right?
00:38:38 But you're a data scientist.
00:38:40 How are you gonna analyze?
00:38:41 Are you gonna design some algorithm or software
00:38:44 to analyze the protein structures?
00:38:47 The question is, how do you do that?
00:38:50 Because so far what you learned is the concept
00:38:52 or pattern by eyes, right?
00:38:55 You learned this by eyes.
00:38:56 The question is, well, how do we analyze this
00:39:00 in a renewable way, right?
00:39:02 So I wanna tell you a very simple algorithm, okay?
00:39:08 A very simple algorithm.
00:39:09 I think I like it because it shows you how do you do,
00:39:16 dimensionality reduction in the real world, okay?
00:39:20 Because this is a three-dimensional structure
00:39:22 of thousands of atoms.
00:39:24 Each atom has three degrees of freedom.
00:39:27 And their collective pattern determines whether each
00:39:31 of the residue is in alpha helix or beta sheet or turn, right?
00:39:35 And it's very hard for me to, you know, translate that
00:39:39 thousands times three coordinates into alpha, beta,
00:39:43 or helix, or turns, right?
00:39:46 If I want to see that, how do I analyze it?
00:39:49 So that was the question, okay?
00:39:51 So I wanna show you a very simple algorithm, okay?
00:39:53 I wouldn't even, let's see what, so again,
00:39:56 the goal is analysis.
00:39:59 Here specifically, it's a secondary structure analysis, okay?
00:40:02 The question is what are the patterns
00:40:03 and why are there such patterns, okay?
00:40:07 So as a reminder, right?
00:40:09 So these are how peptides, bonds are connecting
00:40:12 the neighboring residues, right?
00:40:14 Right, okay?
00:40:16 So, you know, something to keep in mind is that not all the peptides,
00:40:24 I'm sorry, not all the covalent bonds are equal in chemistry, okay?
00:40:28 So if you have this ball and the sticks, and you can rotate these,
00:40:33 for instance, I can fix this part and rotate the remaining
00:40:38 by rotating around this bond, okay?
00:40:40 So in other words, these bonds are rotatable.
00:40:43 Those are the torsional angles, but some bonds,
00:40:47 due to the way the electrons are being shared,
00:40:50 are less rotatable.
00:40:51 So that includes the peptide bond, okay?
00:40:56 So if you want to analyze the secondary structure of every residue,
00:41:02 what chemistry tells you is that,
00:41:05 well, this is residue one, this is residue two, I should break,
00:41:11 you know, if I want to tell whether each residue is in alpha,
00:41:15 helix, or beta sheets, I should break them here
00:41:19 into, you know, different residues.
00:41:21 This is what the chemistry would tell you, is that right?
00:41:24 Right?
00:41:26 But that is actually, you know, not what this algorithm
00:41:29 I'm going to tell you about is doing, okay?
00:41:32 So, what another way, actually a more effective way,
00:41:40 if you think about the thousands times three dimensions
00:41:42 of degrees of freedom, right?
00:41:44 Not all the degrees of freedoms are equal and they are coupled.
00:41:48 So, what another way of decomposing this, you know,
00:41:55 polypeptide chain or protein is actually not at the peptide bond,
00:42:00 but at the central carbon.
00:42:04 Did you see these gray planes, right?
00:42:08 Where they are connected with each other,
00:42:11 are at the central carbon actually, C alpha, right?
00:42:15 We'll have one, two, three, again, three residues
00:42:19 going this direction, okay?
00:42:21 So, actually, you can, you know, decompose this protein backbone
00:42:28 using the so-called polypeptide unit,
00:42:32 this covered by this gray, okay?
00:42:35 So, you may wonder why do I do that, right?
00:42:38 So, it turns out that, you know, these atoms within each peptide unit
00:42:44 as seen in this gray plane are actually relatively rigid to each other, right?
00:42:54 Because of the chemistry, they are not allowed to, you know,
00:42:58 move around with each other, with regard to each other too much
00:43:02 because of the peptide bond is usually regarded as not rotatable, okay?
00:43:08 So, in this case, in other words, you have reduced the degrees of freedom
00:43:14 because within each peptide unit, there's no much relative degrees of freedom
00:43:21 because they cannot move relative to each other.
00:43:23 It's like a rigid piece, right?
00:43:25 Then, you can see that, oh, each peptide unit is really like a, you know,
00:43:29 rigid Lego piece now, and you can now connect with,
00:43:33 connect to the individual pieces or the individual peptide units, right?
00:43:38 So, the degrees of freedoms, what is remaining?
00:43:42 When you are describing the degrees of freedoms for a lot of atoms
00:43:46 connected with each other through bonded or non-bonded interactions,
00:43:50 you can, of course, use thousands times three Cartesian coordinates.
00:43:54 You can do that, right?
00:43:56 But what is the difficulty?
00:43:58 One, it's a lot of degrees of freedoms, right?
00:44:01 It's a huge number.
00:44:03 What is the second problem?
00:44:06 There are a lot of constraints across the degrees of freedom, right?
00:44:11 Because it's not like every atom, three degrees of freedom,
00:44:14 can move freely relative to each other.
00:44:17 If that's the case, we are not talking about protein,
00:44:20 we are talking about gas, isn't it?
00:44:22 Yeah.
00:44:23 So, in that case, what, if you are going to use thousands times three degrees of freedom
00:44:30 in Cartesian coordinates, what you have is extremely high dimensional space,
00:44:34 and in this space, you don't have very nice distribution.
00:44:40 In other words, it's like, you know, on the Pacific,
00:44:44 there are a lot of scattered islands.
00:44:46 Only on those islands, you know, these degrees of freedoms are possible.
00:44:52 But outside of the islands, most of the Pacific is, you know, is not island,
00:44:58 meaning that the combination of the degrees of freedom wouldn't be possible.
00:45:04 Because they are violating the chemistry, or the bounded or non-bonded interactions.
00:45:09 Does that make sense to you?
00:45:10 In other words, you have very high dimensional space with a lot of constraints.
00:45:15 So, an analogy is like, you know, it's a three dimension, you know, on the Pacific,
00:45:20 only some scattered regions are allowed.
00:45:23 So, if that's the case, it's really difficult for you to analyze,
00:45:27 it's really difficult for you to search, right?
00:45:30 So, can you change the coordinates to something lower degrees and more contiguous?
00:45:39 That's our goal, okay?
00:45:41 So, instead of using the Cartesian coordinates, you can use internal coordinates,
00:45:46 like this, the length of the stick, that's a bounded length,
00:45:51 or the angle formed between one, two, three, three atoms, right?
00:45:55 Bound angles.
00:45:57 And these dihedral, that are rotations around each backbone bond, okay?
00:46:04 As I said that this peptide bond here, between the amino group of the previous residue,
00:46:12 right?
00:46:13 And the carboxyl group, I'm sorry, between the carboxyl group of the previous residue,
00:46:18 and the amino group of the next residue, this peptide bond is not rotatable urally.
00:46:25 So, if that is the case, then if we think that the bounded lengths are fixed,
00:46:32 the bounded angles are fixed, and the torsional angles involving four atoms are fixed for the peptide bonds,
00:46:39 then the only remaining degrees of freedoms are what?
00:46:45 Only the torsional angles around this bonds, right?
00:46:49 One before C alpha and one after C alpha, isn't it?
00:46:55 So, in other words, what you end up having is, you know, notice that this, and I call it a phi or a psi,
00:47:03 that's, you know, two means the index of the C alpha.
00:47:07 Phi is the torsional angle around the bond right before the C alpha, and this psi is the torsional angle around the bond right after the C alpha.
00:47:18 Again, because the direction is going one, two, three, okay?
00:47:22 From N terminus to the C terminus.
00:47:24 So, specifically, phi is around the bond between nitrogen and the C alpha, so nitrogen and the C alpha.
00:47:35 And the psi is between, is the dihedral around the C alpha and carbon, right?
00:47:42 The C prime, right?
00:47:43 That's the carbon in the carboxyl group, right?
00:47:46 So now, if we fix the bounded lengths and the angles, then what we end up having is no longer thousands of degrees,
00:47:53 thousands of degrees of freedoms that are all strongly coupled with each other.
00:47:58 We only have, what?
00:48:00 Two times L degrees of freedom left.
00:48:04 Well, L is the length of the amino acid.
00:48:07 Isn't it nice?
00:48:08 Okay, so now let's look at this so-called Ramachandran plot, okay?
00:48:16 So, in other words, we can look at the phi and the psi of every residue together and look at its pattern, okay?
00:48:22 So, I don't want to talk too much details about it because this is high school geometry, okay?
00:48:28 But, so that you understand the concept, I'm just going to say one thing, right?
00:48:32 So, again, you know, this is a phi angle, right?
00:48:36 The phi angle is between the nitrogen of the residue I and the C alpha of residue I, okay?
00:48:45 And to calculate this torsional angle, you are basically calculating the plane, sorry, calculating the angle formed by the two planes, right?
00:48:55 This is atom one, two, three.
00:48:57 This is the plane formed by atom two, three, four, right?
00:49:01 Right?
00:49:02 And atom one, in this case, is the carboxyl carbon of the previous residue I minus one, okay?
00:49:09 So, in the protein structure data, you have the three Cartesian coordinates of every atom and you need to calculate these, the dihedral angle, okay?
00:49:20 So, this is high school geometry, right?
00:49:23 So, you just need to calculate, for instance, in this case, you can calculate the unit normal of these three, of the plane formed by the atom one, two, three, right?
00:49:35 So, this involves what?
00:49:37 This involves cross products of two vectors, isn't it?
00:49:42 Because you have this, this vector, line one, and this vector, line two.
00:49:47 You want to calculate the unit normal, right?
00:49:51 To the plane formed by these two lines that involves the cross product of these two vectors, right?
00:49:57 And you can do the same to get the unit normal of the second plane and then you can find out the angle formed between the two,
00:50:05 that is exactly this torsional angle phi, okay?
00:50:11 So, you may want to try some of this.
00:50:13 I give some tutorials, not tutorials, just additional contents on the canvas, and we may do some simple homework on this.
00:50:24 And be careful that the sign matters, okay?
00:50:30 And similarly, you can calculate size.
00:50:33 So, this is just the high school geometry, but in the end, so this is how you determine the sign.
00:50:40 It's just some convention.
00:50:41 I don't want to emphasize it too much.
00:50:43 But in the end, what you end up having is, you're looking at phi and psi, a two-dimensional visualization.
00:50:52 Phi and psi only, okay?
00:50:55 And you may have a lot of scattered plots or density distributions in this phi-psi plot.
00:51:04 So, this is called the Ramachandran plot.
00:51:07 So, in other words, you are actually now visualizing the 3D structures with regard to its secondary structure pattern,
00:51:19 using a very straightforward, simple illustration.
00:51:23 Because you have, you know, decomposed the protein into peptide units,
00:51:28 you figure out that the degrees of freedom can be reduced down to phi and psi for each residue.
00:51:34 So, if I calculate the phi and psi of each residue and show the scattered plot or density distribution,
00:51:42 you will actually find that some, you know, distributions, some regions of the phi-psi distributions are corresponding to the alpha helis.
00:51:54 Some are corresponding to the beta sheets, and some of them are corresponding to the left-handed helis.
00:52:03 Okay?
00:52:04 So, it's a very simple way.
00:52:06 So, in fact, even right now, if you use alpha-fold, for instance, to make a powerful prediction,
00:52:13 you still want to somehow do a sanity check.
00:52:15 One of the sanity check is to see whether the distribution of the phi-sine angles in the predicted structures
00:52:22 are more or less, you know, agreeing with the distribution in the current data.
00:52:29 Okay?
00:52:30 So, this is one way to analyze this.
00:52:34 Okay?
00:52:35 So, this is the scattered plot.
00:52:40 Okay?
00:52:41 So, I want to ask you a few questions.
00:52:44 In this case, it's a scattered plot.
00:52:45 Okay?
00:52:46 For instance, I take, there's a homework, I can look at all the PDB structures or a subset of them.
00:52:52 All right?
00:52:53 I have to calculate the phi and psi for each residue of each structure and each,
00:52:58 phi and psi pair of one residue in one structure is one dot.
00:53:04 And then I show a scattered plot, for instance, in this case.
00:53:07 Okay?
00:53:08 But, in this case, I show them in two different cases because there are 20 amino acids, right?
00:53:14 20 types of amino acids, right?
00:53:17 So, glycine is one of them.
00:53:19 In this case, I'm showing the Ramachandran plot for all the 19, all residues except glycine.
00:53:26 So, 19 amino acids versus glycine.
00:53:32 What do you see the difference?
00:53:34 And, you know, what's the difference in the pattern and why?
00:53:39 First, we can talk about what's the difference because, you know, for all the observed, for all the residues except glycine, you see this, you know, density distribution centered around alpha and beta regions and some small number of left helices.
00:53:58 That is, according to our understanding, correct, right?
00:54:02 But what do you see from the Ramachandran plot for glycines?
00:54:08 What do you see?
00:54:09 What's the pattern?
00:54:10 Do you see a difference?
00:54:12 What's the difference?
00:54:19 What's the difference?
00:54:28 No difference?
00:54:34 So, more regions in this phi-sci plane are sampled in the case of glycine, isn't it?
00:54:43 For instance, this region is represented a lot, right?
00:54:47 Right?
00:54:48 And this region as well compared to the other 19 amino acids, isn't it?
00:54:55 So, in other words, glycines are more free to adopt all kinds of phi-sci combination, aka, backbone structure.
00:55:08 Right?
00:55:09 Why?
00:55:10 Is that surprising?
00:55:12 Any guess?
00:55:14 Why?
00:55:15 Why are glycines more free to adopt backbone structures?
00:55:23 There's no side chain.
00:55:26 There's, Ruge's answer is there's no side chain on the glycine.
00:55:30 Yeah.
00:55:31 So, I think that the answer is correct.
00:55:33 But to make it more precise, glycine's side chain is the smallest.
00:55:38 It only has one hydrogen.
00:55:40 Okay?
00:55:41 So, in other words, glycine's side chain has the least restriction, right, on the structural choices of the backbone.
00:55:53 Right?
00:55:54 So, that's why the backbones are more free to adopt all kinds of combinations, right?
00:55:59 Right?
00:56:00 So, maybe in your homework when you do it by yourself and plot a Ramachandran plot for, for instance, prolines, you will see a very different distribution as well and you may ask why.
00:56:13 Okay?
00:56:14 So, that's one kind of, another kind of exercise.
00:56:18 Yes, question?
00:56:19 So, I'm a little confused.
00:56:20 So, like the three secondary structures that we learned, like, they are always either like beta sheets or like alpha etates or the loops or like there are other like structures as well.
00:56:38 Yeah.
00:56:39 Yeah.
00:56:40 Yeah, very good point.
00:56:41 So, usually we classify secondary structures into three classes or eight classes.
00:56:48 So, if you classify them into three classes, then it's just alpha helix, beta sheets, and the loops.
00:56:54 But if you want to, you know, look into them more carefully and show more higher resolution, then you may consider eight classes.
00:57:06 For instance, the three ten helix is one of the eight.
00:57:09 Okay?
00:57:10 So, in other words, it's just how many classes of the label you are adopting.
00:57:15 It could be three, it could be ten.
00:57:16 Or, you know, what's the size of the alphabet for the secondary structure text?
00:57:21 Okay?
00:57:22 Good question.
00:57:23 Any other questions?
00:57:26 Okay.
00:57:27 So, now, we talked about a sequence, we talked about one analysis algorithm, right?
00:57:36 So simple, but it's really not that easy to figure out, okay?
00:57:39 So, that's why it's named after the scientist, the Ramachandran, okay?
00:57:43 And it's still used very often, even in the post-alpha fold era.
00:57:48 So, now let me show a second algorithm.
00:57:51 I'm not showing an algorithm.
00:57:53 Actually, I'm showing a second utility of the algorithm.
00:57:57 That is a prediction.
00:57:58 The question, you know, usually the prediction kind of task is asking what would happen given this, okay?
00:58:03 So, our question is what, you know, given the one-dimensional sequence of amino acid, what are the secondary structures of each position?
00:58:12 Okay?
00:58:13 Which portion is alpha, which portion is beta, and so on.
00:58:16 So, that's a secondary structure prediction question, right?
00:58:19 So, as I said, this is one of the earliest machine learning problem.
00:58:25 And if you don't know who Martin Koplas is, I regard him as my academic grandfather.
00:58:31 Anyway, so, also a Nobel Laureate at 2013 in chemistry.
00:58:38 I just want to remind you that Martin Koplas already used the neural network in 1989, probably before you were born, to predict the secondary structure, okay?
00:58:49 So, neural network is nothing new, okay?
00:58:52 But deep learning is slightly different.
00:58:54 Anyway, and also back then there were not a lot of data.
00:58:57 There were not enough computing resources.
00:58:59 But anyway, I mean, these are the, you know, founding fathers of this field.
00:59:06 But anyway, one of the elementary tasks, prediction tasks, still remain in secondary structure, in the structure of bioinformatics.
00:59:19 That is to predict the secondary structure, okay?
00:59:21 I just said that, I think now with the advancing of data, right?
00:59:30 We have way more sequences already, and we have more sequences that are labeled with three-class or eight-class secondary structures.
00:59:37 This question can be answered much, you know, more effectively now than 1989, right?
00:59:45 So, I mentioned the machine translation formulation.
00:59:48 I think that is one way, right?
00:59:49 So, you can use recurrent neural networks, for instance, right?
00:59:55 And machine translation, I think you can use other things as well, right?
00:59:59 So, one of the recent papers are actually using this RNN framework to start with a one-dimensional sequence and predict this, you know, secondary, sorry, the secondary structure sequence.
01:00:13 So, it's a sequence-to-sequence translation formulation.
01:00:15 It's still relatively easy.
01:00:17 And we'll briefly revisit the topics when we go to the data-driven approach.
01:00:26 So, you can see the secondary structure prediction problem.
01:00:29 And there are some, you know, review that you can look at.
01:00:34 Now, what is the secondary structure prediction problem?
01:00:38 The input is a sequence.
01:00:40 It's a letter, it's a text written alphabet of 20 letters.
01:00:44 But, sometimes you can see something, a format for the sequence called a FASTA format.
01:00:50 So, what is a FASTA format?
01:00:52 Actually, you can just regard it as, you know, something right before the actual text, you have a header.
01:01:00 You have a header starting with a bigger than this symbol, okay?
01:01:04 So, this, right before the line of sequence data, you have a single line description header.
01:01:09 And you can determine it's your own format.
01:01:12 For instance, this is example one.
01:01:14 This example one, this FASTA sequence is taken from the NCBI entrance.
01:01:20 So, you see that they have a lot of vertical lines separating some keywords.
01:01:25 So, these keywords can be used to cross-reference different datasets, okay?
01:01:30 For instance, in this case, it can include the entrance gene ID, protein ID, and the uniprod ID.
01:01:39 So, these are different datasets, and these are the different IDs so that you can cross-reference, okay?
01:01:45 And in the case of the protein databank, we'll talk more about it next time.
01:01:49 So, this is the first time you see that, you know, these vertical lines are separating the four-letter code of the PDBID, and then colon, and then the chain ID, and then this, you know, PDBID chain sequence.
01:02:03 So, this is kind of the header.
01:02:05 You can decide, you know, what is the format, but usually, you know, that is where you can put the additional information, okay?
01:02:14 We recently designed some sequences, and then, you know, we want to still convert them into FASTA format, but we didn't really need any information in the, in the, in the, much information in the header.
01:02:26 So, we just wrote them as bigger than, designed by what model, with what setting, and then, you know, with what filter, for instance, okay?
01:02:36 So, this is the FASTA sequence.
01:02:38 The output, as I said, is, because our labels, in this case, are discrete, right?
01:02:46 It's a three classes or eight classes.
01:02:49 So, it's a classification problem, right?
01:02:51 Right?
01:02:52 Your labels are discrete.
01:02:53 So, it's a classification problem.
01:02:54 So, in the early days, they are really solved, you know, like, you know, shallow machine learning models using classification, right?
01:03:01 So, for instance, this is one server called Raptor, right?
01:03:07 So, you see that the first line is the sequence.
01:03:12 P, I, V, K, R, L, L, G, W, K, right?
01:03:17 Proline, isoleucine, valine, lysine, arginine, leucine, leucine, and so on.
01:03:22 Okay?
01:03:23 Glycine, and so on.
01:03:24 And there's position one, position 11, and so on.
01:03:27 Okay?
01:03:28 So, the first row is the input.
01:03:31 That is a sequence of the protein.
01:03:34 And the next four rows are the predicted secondary structure and other labels.
01:03:42 Okay?
01:03:43 For instance, this is SS3, meaning secondary structure, three-class secondary structure.
01:03:51 So, Nami was asking about whether you have three classes or other more classes, right?
01:03:58 So, you see that in this case, we predict these labels in both three-class and eight-class, okay?
01:04:07 Secondary structures.
01:04:08 So, we see that in the case of the three-class secondary structures, position one is a proline.
01:04:14 It's predicted to be in the C.
01:04:17 That's a coil.
01:04:18 Okay?
01:04:19 Position IVKRLG is predicted to form a helix segment, H.
01:04:27 Okay?
01:04:28 Then when you see an E, that means it's a beta strand.
01:04:32 Okay?
01:04:33 And similarly, you can look at the secondary structure in the eight classes.
01:04:39 Okay?
01:04:40 And you can also predict other things that are not secondary structure but related to the structure.
01:04:45 For instance, you can predict which position or residue is hidden in the core of the protein
01:04:52 and which are closer to the surface of the protein.
01:04:56 So, in other words, you can predict the context.
01:04:59 So, in this case, E or M are, you know, showing whether it's, you know, buried in the B.
01:05:10 I think they just show that whether it's, you know, buried or in the surface or somewhere in between.
01:05:16 Okay?
01:05:17 So, there are three kinds of labels showing where they are from the core to the surface of the protein.
01:05:24 And lastly, you are predicting which positions are disordered because so far I'm showing that everything is nicely ordered.
01:05:31 But the reality is that in our bodies there are a lot of proteins that are not structured.
01:05:37 They are intrinsically disordered.
01:05:39 You may wonder, well, you told me that a protein needs to adapt into a three-dimensional shape to function.
01:05:45 Now you're telling me that a lot of proteins don't have structures.
01:05:48 So, what are you doing?
01:05:49 Well, it's still, you know, an ongoing study and there are a lot of discussions, but even about, you know, what is the definition of intrinsic disordered proteins.
01:06:01 But one theory is that if you have a protein with some intrinsically disordered regions, then that region is actually quite flexible to do different things in different ways.
01:06:14 So, you know, different things in different environments or contexts.
01:06:17 Okay?
01:06:18 So, that's one of the theory.
01:06:21 So, lastly, I want to talk about the tertiary structures.
01:06:27 Now that you have learned the secondary structures, local secondary structures, right?
01:06:31 And you see that they are classified in three or eight types and why they can be treated as classification problem.
01:06:41 And intrinsically, it's a one-dimensional sequence as well.
01:06:45 So, that's why you can do this classification as if you are doing the translation, right?
01:06:52 Like translating one to another.
01:06:53 Actually, it's easier than translating from English to France.
01:06:57 Why?
01:06:58 Because there's a one-to-one correspondence between the characters.
01:07:04 It's actually much easier.
01:07:06 Okay?
01:07:07 So, if you are thinking that, oh, some of the most advanced human language models for translation needed to be done in that way, you have to ask yourself how much of that is needed for our case.
01:07:18 And how much of our special properties can be utilized better in your new algorithms?
01:07:27 Okay?
01:07:28 So, there's something to think about.
01:07:30 That's what I mean.
01:07:31 AI is targeting these applications, but it's also inspired from these applications.
01:07:38 So, lastly, the tertiary structures.
01:07:41 So, the tertiary structure is really the global 3D structure, right?
01:07:45 But there are two things, still two concepts I want you to be aware of.
01:07:50 One is what is a domain and one is what is a fold, right?
01:07:54 So, each single tertiary structure corresponds to one chain and then it can form spatially separate regions.
01:08:04 And they are called domains.
01:08:06 I will show you later on an example.
01:08:08 So, these domains are spatially far away from each other, relatively stable and well-folded.
01:08:20 But, you know, they are also conserved parts that can evolve function and exist independently.
01:08:27 Okay?
01:08:28 And then, depending on, you know, within each domain, depending on how the secondary structures are arranged relative to each other, the pattern is called a fold.
01:08:41 Okay?
01:08:42 So, I'll give you a visualization.
01:08:44 So, this is one protein called a pyruvate kinase.
01:08:48 What it's doing is catalyze the following chemical reaction.
01:08:54 That is, you know, it's an enzyme, right?
01:08:57 It's, sorry, it's a pyruvate kinase.
01:09:00 It's an enzyme that catalyzes the following chemical reaction, right?
01:09:03 To convert the, you know, the reactants to the products in this case, right?
01:09:10 With the conversion of ADP to ATP, right?
01:09:15 So, if you ask what is this protein, right?
01:09:18 This protein actually, if you look at this, the code is 1pkn.
01:09:22 I show the backbone cartoon and then already color them in three colors.
01:09:29 So, you can sort of see that there are three sub components of these three-dimensional shape that are relatively, you know, separate from each other.
01:09:41 And each of them is packed, right?
01:09:43 You have very packed individual domains that are relatively loosely connected with each other, like here.
01:09:50 Okay? You have this, right?
01:09:53 And for each domain, there are some patterns of how alpha helix or beta sheets are packed to each other.
01:09:58 That's called the fold, okay?
01:10:00 So, it's very fascinating.
01:10:02 If you look at the, you know, the top domain, this is actually where the protein binds to the nucleotides, okay?
01:10:12 So, and the secondary structures are mostly beta sheets.
01:10:16 You don't have alpha helix here.
01:10:18 So, it's all beta kind of fold, okay?
01:10:22 And the central domain, the gray one, is actually binding to the substrate and helping this chemical reaction, okay?
01:10:31 And it actually contains both alpha helix and beta sheets.
01:10:37 And the alpha and beta are alternating, alternating, okay?
01:10:42 One alpha, one beta, one alpha, one beta strand.
01:10:45 And the beta strands are actually mostly parallel because they are separated by the alpha helix.
01:10:50 So, this is the alpha slash beta, alternating alpha beta kind of fold.
01:10:55 And the bottom domain is more for regulatory functions.
01:10:59 And, you know, again, it's alternating alpha and beta, okay?
01:11:03 So, you see that there are three domains, right?
01:11:06 They are, you know, structurally separate.
01:11:09 They are relatively compact within each other and structurally separate.
01:11:13 But they are actually performing different functions.
01:11:17 Actually, if you wonder how nature made this protein, actually, each of these domains might have evolved independently already.
01:11:25 And then they are put together, assembled together somehow, okay?
01:11:29 So, that's the amazing part.
01:11:32 So, I want you to pay one attention to one thing.
01:11:36 I'm not saying that from N to C terminus, if you follow the sequence, you go through the domain one, then domain two, and domain three, and then that's it.
01:11:46 No, if you see here, you see that you're entering domain three, and then you're finishing domain three, and you're entering back to domain one.
01:11:56 So, in three-dimensional space, they are separated from each other relatively easy.
01:12:01 But in one-dimensional sequence, you may see domain one, domain two, domain one, domain two, then domain three, then domain two again.
01:12:08 You see what I mean?
01:12:09 So, that's a little bit complicating, okay?
01:12:12 So, in other words, the domains are compact in 3D, but not contiguous in 1D, okay?
01:12:22 So, you know, if you look at what these patterns, how these are in 1D, okay?
01:12:30 How these, you know, alpha or beta secondary structures are packed together, and each pattern is like a fold.
01:12:37 I'm just giving you one example, one fold example for each category.
01:12:42 In the case of all alpha, you have this globbing fold, right?
01:12:47 In the case of all beta, you see this top view, which is like, you know, called a beta propeller, a very nice structure.
01:12:57 You can have this all beta structure forming this beta propeller.
01:13:02 And this is a team barrel.
01:13:04 You have the top view and the side view.
01:13:09 It consists of alternating alpha beta.
01:13:12 For instance, if you follow this, let's see.
01:13:15 If you follow this N terminus, you go through a beta first.
01:13:19 Then you come back, you see alpha.
01:13:23 Then you follow through.
01:13:24 The next one is beta and so on, okay?
01:13:27 I'm not counting the loop, but the stable secondary structures are alternating between alpha and beta.
01:13:33 You could have alpha plus beta, for instance, in this case start with, usually this, you know, we are using this rainbow coloring.
01:13:41 So the blue is N terminus and the red is C terminus, okay?
01:13:47 So coldest is N terminus.
01:13:49 So if we follow in the order, we go through beta, alpha, beta, beta, alpha, beta.
01:14:02 So it's not alternating alpha, beta anymore because we have this, you know, two betas already neighboring each other forming a sheet.
01:14:10 So in this case, it's alpha plus beta, okay?
01:14:14 So each individual fold has its three-dimensional pattern and it's also very often related to a function or some functions, okay?
01:14:26 So in the end, let's summarize and see whether there's any questions.
01:14:31 Today we still talked about a little bit more concepts about the structures so that we can understand the tasks and the algorithms, okay?
01:14:41 So we talked about the secondary structures, mostly the three-class definition.
01:14:46 We talked about what made alpha helices and beta sheets because of the hydrogen bonds within a segment or across adjacent segments that form alpha helices and beta strands, sheets, excuse me, respectively, right?
01:15:03 We talked about...
01:15:04 ...