1
00:00:00,090 --> 00:00:03,720
My name is Charles Godfrey, I am director of the Oxford Martin School.

2
00:00:03,720 --> 00:00:10,680
And thank you for joining us on an extra Web conversation that we're having over the last term.

3
00:00:10,680 --> 00:00:18,660
We've been having a series of conversations on the topic of what the slogan building back better means after the pandemic.

4
00:00:18,660 --> 00:00:23,070
And that series stopped last week. But something else happened last week.

5
00:00:23,070 --> 00:00:32,600
And this was a really interesting scientific bonds, a potential solution of what is known as a protein folding problem.

6
00:00:32,600 --> 00:00:40,410
And we have a number of emails which asked us, might we have a talk to explore some of the exciting issues around that?

7
00:00:40,410 --> 00:00:46,500
And that is what this talk is today. And I'm joined by two really interesting guess.

8
00:00:46,500 --> 00:00:56,460
One who you can see who's filled begin in front and one who you'll be able to hear, but I'm afraid not see because of a technical issue.

9
00:00:56,460 --> 00:01:01,770
And that is Yvonne Jones. Yvonne, can I just cheque that you are there and listening to us?

10
00:01:01,770 --> 00:01:07,210
I am. And I'm really sorry, everybody. Sorry, Charles and not Brandon.

11
00:01:07,210 --> 00:01:15,180
As being as Charles said, there's something weird has happened to the camera and my laptop and it just refuses to.

12
00:01:15,180 --> 00:01:20,970
There is a gremlin somewhere. So let me begin just by introducing Yvonne.

13
00:01:20,970 --> 00:01:25,740
So Yvonne is professor of Protein Crystallography at Oxford.

14
00:01:25,740 --> 00:01:31,860
She's a fellow of the Royal Society. She co-founded at Oxford with Dave Stewart, the Division of Structural Biology,

15
00:01:31,860 --> 00:01:39,720
which is part of the Nuffield Department of Clinical Met of Clinical Medicine, of which she is Kohat at the moment.

16
00:01:39,720 --> 00:01:48,710
And Phil Felbeck and is professor of computational biochemistry in the Department of Biochemistry, which is part of our medical science division.

17
00:01:48,710 --> 00:01:50,220
And rather than me doing,

18
00:01:50,220 --> 00:01:58,080
I'm going to ask fellin Yvonne in about two sentences just to give us a flavour of the work in their lab at the moment involved.

19
00:01:58,080 --> 00:02:06,280
Might you go first? Sure. I'm interested in the receptors that sit on cell surfaces and taken signals in the

20
00:02:06,280 --> 00:02:11,470
form of proteins that bind to them to allow cells to communicate to each other.

21
00:02:11,470 --> 00:02:21,840
I'm particularly interested in signalling between cells in both in the developing nervous system and also cells involved in the immune system.

22
00:02:21,840 --> 00:02:25,500
And I use protein crystallography to solve those structures,

23
00:02:25,500 --> 00:02:30,730
but I also bring in lots of other techniques because I'm really interested in the way that the proteins interact with each other,

24
00:02:30,730 --> 00:02:36,350
not just in their shape. Thank you. Volume. Yes.

25
00:02:36,350 --> 00:02:39,310
And my lab is also interesting in receptors.

26
00:02:39,310 --> 00:02:49,570
But we predominantly use molecular dynamics simulation technology to look at the underlying dynamics of these proteins.

27
00:02:49,570 --> 00:02:53,350
Looking at things like channels and transport of proteins.

28
00:02:53,350 --> 00:03:02,560
And although we don't do a structure prediction, say, many of the underlying computational methods that they use, we also use as well.

29
00:03:02,560 --> 00:03:08,290
Thank you very much both. So what we're going to do is we're going to just chat very briefly about what a protein is.

30
00:03:08,290 --> 00:03:12,190
And I realise that many people listening will be very familiar with it.

31
00:03:12,190 --> 00:03:15,670
But we're going to have a brief discussion of that just to bring people up to speed.

32
00:03:15,670 --> 00:03:24,310
If they don't know about it. And then we'll go on to what exactly this recent advances and some of the consequences of it.

33
00:03:24,310 --> 00:03:30,910
Now, those of you watching a live in crowd cast will see that darn near the bottom right hand of your screen.

34
00:03:30,910 --> 00:03:36,850
There is a bus. No facility could ask a question. We do really encourage you to ask a question.

35
00:03:36,850 --> 00:03:46,030
There's also the facility to vote questions up so that if someone else has asked the question that you're really interested in seeing answered,

36
00:03:46,030 --> 00:03:48,900
then you you're able to vote it up.

37
00:03:48,900 --> 00:03:58,240
And while it's not exactly a strict democracy when it comes to looking for questions, I see the ones which have got have got the most support.

38
00:03:58,240 --> 00:04:02,860
So let's explore a little bit about the background.

39
00:04:02,860 --> 00:04:14,370
And if on Patch, I go to you just to remind us about what a protein is and why we're worried or why we are interested in its three dimensional shape.

40
00:04:14,370 --> 00:04:22,720
OK, so proteins are the little machines or the little workers in ourselves and also messengers that go between ourselves.

41
00:04:22,720 --> 00:04:36,720
They're encoded by the in the genome, by our DNA sequences, which are translated into what have been described as beads on a string.

42
00:04:36,720 --> 00:04:42,240
So there are a number of different flavours of amino acids.

43
00:04:42,240 --> 00:04:50,820
And these form into polypeptide chains, so long strings of beads, each of which have rather different properties.

44
00:04:50,820 --> 00:04:59,040
Some of them are positively charged and negatively charged or some sticky sort of oily amino acids.

45
00:04:59,040 --> 00:05:08,370
And according to those properties, the beads on the string will fold up into a complex Three-Dimensional Structure.

46
00:05:08,370 --> 00:05:18,840
I heard John Thornton on the on the Today programme, actually, when this announcement about the protein folding problem hit the news week or so ago.

47
00:05:18,840 --> 00:05:21,380
And she described it as. Yes, sort of.

48
00:05:21,380 --> 00:05:28,370
You start out with a shoelace and you're kind of, you know, forming that or tying it up into the three dimensional shape.

49
00:05:28,370 --> 00:05:33,510
So now that kind of origami. Absolutely. Yes.

50
00:05:33,510 --> 00:05:36,720
And why is it interesting the way that it all folds up?

51
00:05:36,720 --> 00:05:41,020
Well, because each protein has its own unique shape, its own unique fold,

52
00:05:41,020 --> 00:05:46,140
and that fold, that shape that it has in three dimensions determines its properties.

53
00:05:46,140 --> 00:05:52,530
What it can do. I just said that that that like little machines in our cells or the messenger is going between cells.

54
00:05:52,530 --> 00:06:00,750
So they are capable of doing all the tasks that we need doing but to do those tasks.

55
00:06:00,750 --> 00:06:07,720
They have a very specific shape. And if you don't know, I'm not explaining this very well.

56
00:06:07,720 --> 00:06:13,170
If you don't know the shape. It's a bit like not knowing what's underneath the bonnet of a car.

57
00:06:13,170 --> 00:06:17,640
You can't work out how they work if you don't know what shape they are.

58
00:06:17,640 --> 00:06:21,620
So understanding a three dimensional shape has really critical.

59
00:06:21,620 --> 00:06:24,330
Now, your professor approaching crystallography,

60
00:06:24,330 --> 00:06:35,120
might you say how crystallography an X-ray diffraction is used to to construct a three to mine to determine a three dimensional shape?

61
00:06:35,120 --> 00:06:46,100
Right. Well, the problem is obviously that that proteins are too small to be able to visualise in using like microscopy.

62
00:06:46,100 --> 00:06:56,300
So there are a number of older its magnitude smaller than than you could do that using using visible light to to to see that detail.

63
00:06:56,300 --> 00:07:09,080
But you can use X-rays. They have a wavelengths that is able to pick up that fine structure.

64
00:07:09,080 --> 00:07:18,020
If you do that, you you can't really just do that by having a single copy of a protein and hitting it with an x ray with x rays.

65
00:07:18,020 --> 00:07:24,440
You need many copies of the protein to all lined up rather nicely in in a crystal structure.

66
00:07:24,440 --> 00:07:29,990
Crystal structures just so that the crystal is just a trick and amplification method.

67
00:07:29,990 --> 00:07:32,990
So you're able to look at many,

68
00:07:32,990 --> 00:07:42,050
many copies of the protein at once and that will give you a strong enough signal to be able to actually solve the structure.

69
00:07:42,050 --> 00:07:46,760
Unfortunately, it's a bit it's a bit more tricky than that because you don't actually manage to

70
00:07:46,760 --> 00:07:52,170
measure all the information you need just from the the amplitudes of the diffract.

71
00:07:52,170 --> 00:07:54,530
Did x rays, as we call them.

72
00:07:54,530 --> 00:08:02,750
There is some other sort of missing bits of information to do with the timing of the phase of the different refracted rays coming off.

73
00:08:02,750 --> 00:08:10,500
A part of the complexity of the problem. Is why so many Nobel prises have gone to x ray crystallography over the years.

74
00:08:10,500 --> 00:08:18,950
I guess. I guess I guess maybe initially that would be a case, because when people like Darcy Hodgkins working in Oxford,

75
00:08:18,950 --> 00:08:28,500
Max Perutz and John Kendrew working in Cambridge, when they started out, it was a mammoth task.

76
00:08:28,500 --> 00:08:33,440
And they had to work out all the methods of how you could how you could even begin to sort out

77
00:08:33,440 --> 00:08:38,360
what a protein structure would look like if they had no idea what it was going to look like.

78
00:08:38,360 --> 00:08:46,180
But. From the beginning, the people who always cheer because it's a difficult thing to do.

79
00:08:46,180 --> 00:08:53,100
Always chosen, structured, trying to do structures that are going to be very enlightening for the biology.

80
00:08:53,100 --> 00:08:57,730
So in the case of Dorothy at she she worked for many years on insulin.

81
00:08:57,730 --> 00:09:01,290
She thought this is going to be an important structure to understand,

82
00:09:01,290 --> 00:09:11,230
because ultimately we want to be able to make forms of insulin that will help diabetics and mobile proteins be crystallised.

83
00:09:11,230 --> 00:09:16,860
Although some proteins that you just can't get a structure out of using this approach.

84
00:09:16,860 --> 00:09:23,630
Yeah, I mean, for many proteins that we have succeeded in in crystallising the various tricks.

85
00:09:23,630 --> 00:09:32,130
Quite a lot of them work over many years to get to persuade them to crystallise using different techniques to get them to crystallise.

86
00:09:32,130 --> 00:09:38,280
But there are also some proteins that have just never crystallised.

87
00:09:38,280 --> 00:09:47,860
Now then quite why that might be one. One whole set of proteins that weren't crystalise are those that are only partially structured.

88
00:09:47,860 --> 00:09:52,230
Don't really become structured when they interact with other proteins.

89
00:09:52,230 --> 00:09:56,460
Well, that's another whole different different problem. But yeah, that we do.

90
00:09:56,460 --> 00:10:00,060
Yeah. It's it's a bit of a blackout persuading them to crystallise.

91
00:10:00,060 --> 00:10:07,510
But of course, you know, there are other methods for doing experimentally solving Three-Dimensional Protein Structures.

92
00:10:07,510 --> 00:10:10,800
I was going to ask you about that. But you have a lot of outcry.

93
00:10:10,800 --> 00:10:18,870
OEM, might you just explain what matters and what different scientists make to a protein that structural studies?

94
00:10:18,870 --> 00:10:25,520
So I was saying that x rays are can can can allow you to probe the fine,

95
00:10:25,520 --> 00:10:35,250
the fine detail that sort of dimensions of the distances between atoms that you need to know to visualise a protein structure.

96
00:10:35,250 --> 00:10:40,850
Well. Electrons can allow you to do that as well.

97
00:10:40,850 --> 00:10:48,100
So an electron microscope will allow you to solve the protein structure.

98
00:10:48,100 --> 00:10:56,860
But for many years, that promise of a big able to solve the protein structure was out of reach.

99
00:10:56,860 --> 00:11:05,620
The first hurdle was, again, that you need to be able to look at a time to gather the information from a lot of different copies of the protein.

100
00:11:05,620 --> 00:11:12,880
And each one individually would be damaged very rapidly by being hit by the electrons in electron microscope.

101
00:11:12,880 --> 00:11:20,020
So you have to be able to do cryopreserved to embed them in in vitreous ice in a way

102
00:11:20,020 --> 00:11:26,350
that people discovered ways of being able to do that without damaging the protein.

103
00:11:26,350 --> 00:11:33,310
But then you need to be able to collect huge amounts of data to do so very sensitively.

104
00:11:33,310 --> 00:11:40,660
So you need really, really good detective systems. And then you need really, really good computer power to crunch them,

105
00:11:40,660 --> 00:11:45,280
because some of the the tricks that we've been able to use in protein crystallography,

106
00:11:45,280 --> 00:11:50,290
the fact that we've got crystal, the computer has to do that for the criterium.

107
00:11:50,290 --> 00:11:55,540
It has to be able to add together all the individual images from many, many, many,

108
00:11:55,540 --> 00:12:01,840
many, many copies of the individual protein molecules that you've got on your bank.

109
00:12:01,840 --> 00:12:09,610
We had a grid get so needed, really good computers needed, really big advances in detectors and bingo, you start to be able to do it.

110
00:12:09,610 --> 00:12:17,380
And that's opened the door for lots of proteins that were recalcitrant when it came to crystallising them.

111
00:12:17,380 --> 00:12:21,580
So now we can use the two techniques in tandem. It's really powerful and exciting.

112
00:12:21,580 --> 00:12:29,020
Thanks a lot. Come on. So that seems sort of physical, direct way of estimating protein structure.

113
00:12:29,020 --> 00:12:35,890
But of course, for most proteins, we have the amino acid sequence. We have the sequence of beads on the string that Yvonne talked about.

114
00:12:35,890 --> 00:12:40,660
And we pretty much know the different ways that the molecules interact.

115
00:12:40,660 --> 00:12:48,370
So, Phil, you have all this information. Why is it not straightforward just to use a big computer to calculate the three dimensional structure?

116
00:12:48,370 --> 00:12:53,980
Why is there a protein folding problem? Yes, it sounds such a simple question, doesn't it?

117
00:12:53,980 --> 00:13:02,710
We got the sequence to should be able to just compute the structure. And the best way probably to explain this was, is that kind of thought exercise.

118
00:13:02,710 --> 00:13:09,970
That is a classic paradox that many people probably have heard of and formulated by serious elemental.

119
00:13:09,970 --> 00:13:16,990
In 1969, when he noted that it would take longer than the age of the known universe to enumerate

120
00:13:16,990 --> 00:13:22,450
all possible configurations of a typical protein by brute force calculation.

121
00:13:22,450 --> 00:13:31,360
So he estimated that something like 10 to the power of 300 possible confirmations for a typical protein.

122
00:13:31,360 --> 00:13:40,780
And, you know, two out of that in each confirmation would require even picosecond of time would mean that actually that require time,

123
00:13:40,780 --> 00:13:43,240
way beyond even the age of the universe.

124
00:13:43,240 --> 00:13:52,330
And just to put that in context, the context, the age of the universe is about 14 billion years, which is about four times 10 to the 17 seconds.

125
00:13:52,330 --> 00:13:56,560
So, yes, that's why it's such a hard problem, because you can't solve it by brute force,

126
00:13:56,560 --> 00:14:02,950
because you'd have to numerical calculate all of the energies of these different potential confirmations.

127
00:14:02,950 --> 00:14:08,130
And even if you had a very fast way of doing that, it would just simply be very intuitive.

128
00:14:08,130 --> 00:14:14,560
You wouldn't have enough computer. And yet, you know, a nature proteins fold spontaneously, something within milliseconds.

129
00:14:14,560 --> 00:14:18,460
And that dichotomy is referred to as the minerals paradox.

130
00:14:18,460 --> 00:14:25,720
So if you can't do it by brute force, what approaches have computational biochemist taken over the last 50 years?

131
00:14:25,720 --> 00:14:32,550
Yes. So that's your question. So I would say you could even go back beyond that, actually, you could probably go back to the 50s.

132
00:14:32,550 --> 00:14:39,080
The first prediction, if you like, can be attributed to Powerlink and Corey, who first suggested this notion that they would be out.

133
00:14:39,080 --> 00:14:43,460
He loses and sheets and they're going to do.

134
00:14:43,460 --> 00:14:49,870
Just to interrupt you there. Alpha. I see. And sheets, apart from protein structures that you get into lots of different proteins that are

135
00:14:49,870 --> 00:14:54,560
sort of common theme that you get three three deaths in Iraq as sort of substructure,

136
00:14:54,560 --> 00:14:58,250
if you like. That's probably the right way to think of them. Yes, exactly, sir.

137
00:14:58,250 --> 00:15:02,960
Yes, sir. Sir. So Alpha Helices had this kind of Ilic ethical kind of confirmation in sheets.

138
00:15:02,960 --> 00:15:06,500
It just kind of extended strand confirmations.

139
00:15:06,500 --> 00:15:13,530
And the 50s was also a time when the first sequence was actually elucidated, which haven't just mentioned incident.

140
00:15:13,530 --> 00:15:18,800
Incident, I think was the first sequence that was elucidated by Fred Syma.

141
00:15:18,800 --> 00:15:28,850
And around this time, Antonsen was also doing this quite famous experiments on RNA and looking at how they would fall.

142
00:15:28,850 --> 00:15:33,590
And that's a now very famous experiment that he did in 1961,

143
00:15:33,590 --> 00:15:37,550
which concluded with a statement which was basically that the nature confirmation of

144
00:15:37,550 --> 00:15:43,690
a protein is determined entirely by the amino acid sequence in a given environment,

145
00:15:43,690 --> 00:15:45,730
but that the latter flies off and left off.

146
00:15:45,730 --> 00:15:55,130
It's actually superimportant that this given environment and then obviously that became a thing to predict, try and predict the 3D structure.

147
00:15:55,130 --> 00:16:02,480
But I would say most effort was focussed on the smaller task of being able to predict the secondary structure.

148
00:16:02,480 --> 00:16:10,790
So by that, I mean he's out helixes in sheets so he could just work out which bits might adopt these helical bits or strand strands.

149
00:16:10,790 --> 00:16:15,670
That would be a good start. And so a lot of the early work sort of dates back to work.

150
00:16:15,670 --> 00:16:25,910
Chewin Fischman in about 1974 was the first, I would say the first real prediction of structure, this decision sheets, if you like.

151
00:16:25,910 --> 00:16:27,320
And they simply said, well,

152
00:16:27,320 --> 00:16:36,410
we'll just use the statistical propensities of the amino acid residues to form a specific structural element or secondary structural element.

153
00:16:36,410 --> 00:16:42,080
And they did that and that was quite successful. And ever since then, people have been trying to extend that.

154
00:16:42,080 --> 00:16:44,210
And people have started to employ, for example,

155
00:16:44,210 --> 00:16:51,170
neural networks in the late 80s and 90s that was first on to secondary structure prediction quite successfully.

156
00:16:51,170 --> 00:16:59,810
And that's still still ongoing. And then I suppose the next thing to note that this particularly important in the context

157
00:16:59,810 --> 00:17:06,290
of a protein structure prediction is the is the establishment of the PDB in 1971.

158
00:17:06,290 --> 00:17:08,570
So that's that protein databank,

159
00:17:08,570 --> 00:17:16,590
which is essentially a database with which is where the structures that Yvonne was talking about that you derive from x

160
00:17:16,590 --> 00:17:23,270
ray crystallography are deposited in such a way that it means that anybody around the planet can actually access them.

161
00:17:23,270 --> 00:17:31,250
And we're gonna be coming onto that when we talk about Alpha fold. And forgive me, felt just hurrying along to it.

162
00:17:31,250 --> 00:17:35,390
Could you tell us a little bit about the CASPA competition? Yes.

163
00:17:35,390 --> 00:17:43,460
So the cash competition was something of a set up in a nineteen ninety four stands

164
00:17:43,460 --> 00:17:48,400
for Past Stands with the critical assessment of structure prediction software.

165
00:17:48,400 --> 00:17:56,750
94 people in the field kind of realised that they needed some way of trying to assess how good these predictions really were.

166
00:17:56,750 --> 00:17:59,990
And the idea was simply to do this in a double blind experiment.

167
00:17:59,990 --> 00:18:11,270
So double blind in the sense that you would release the sequences of rupture and withhold the structure itself back and allow participants to enter.

168
00:18:11,270 --> 00:18:12,620
They were entering a blind way,

169
00:18:12,620 --> 00:18:20,500
meaning that the people who were going to assess the submissions didn't know who the people are making those submissions were.

170
00:18:20,500 --> 00:18:25,970
And it was a very successful competition. There were 33 targets at the time, about 35.

171
00:18:25,970 --> 00:18:30,560
I think groups took part in that and made about 100 predictions.

172
00:18:30,560 --> 00:18:34,250
And that model has been followed up. And it was so successful.

173
00:18:34,250 --> 00:18:38,330
It's still going now. Obviously, it was way as we know, but it was also copied.

174
00:18:38,330 --> 00:18:46,160
So there was also something called Caffery, which is the critical assessment of protein interactions that are set up in 2001.

175
00:18:46,160 --> 00:18:47,690
That's still going as well.

176
00:18:47,690 --> 00:18:56,570
There was also something called Cafer Critical Assessment of function annotation that was set up in 2010 that I stopped in 2014, I think.

177
00:18:56,570 --> 00:19:03,090
But I think the community still active. And another related set of competitions is called Sample,

178
00:19:03,090 --> 00:19:08,600
which stands for this statistical assessment of the modelling of proteins and ligands and the name,

179
00:19:08,600 --> 00:19:15,200
as it suggests that the idea of that is to focus more on problems to do with Miggins and their interactions with parties.

180
00:19:15,200 --> 00:19:20,680
And that's something that my group personally, I'm a little bit more involved with it in the last couple of years.

181
00:19:20,680 --> 00:19:24,910
And it's interesting you say that on a population, biologists and certainly.

182
00:19:24,910 --> 00:19:32,560
The mathematic mathematical population biologist, my area have looked with great envy at Caspa and the success of Bitton.

183
00:19:32,560 --> 00:19:35,970
We've tried to think whether we could set up something equivalent to Snarfing or.

184
00:19:35,970 --> 00:19:41,290
But so far have failed. Maybe all the headlines because it is a bomb.

185
00:19:41,290 --> 00:19:45,880
Can I just interject? No, it's just from the perspective of an experimentalist.

186
00:19:45,880 --> 00:19:51,280
I mean, Caspa. Is it. Is it. Is it the fantastic community effort?

187
00:19:51,280 --> 00:19:54,820
The guys that run cars go round,

188
00:19:54,820 --> 00:20:03,100
get in touch with all the experimentalists in the run up to the competition and ask us if we've got any structures that we haven't yet published.

189
00:20:03,100 --> 00:20:10,980
And so it's there's a real interplay between the experimental experimentalists community and and the, you know,

190
00:20:10,980 --> 00:20:18,040
the protein folders, because that's where they're getting they're getting structures that we haven't yet published.

191
00:20:18,040 --> 00:20:22,570
I myself haven't the timing has never been quite right for me to be able to give them something.

192
00:20:22,570 --> 00:20:27,460
But, you know, these these Emo's pop into my inbox every couple of years.

193
00:20:27,460 --> 00:20:29,470
Twenty years ago, when I worked at Imperial College,

194
00:20:29,470 --> 00:20:38,580
I remember going to a tremendous party when Mike Sternberg's group that had done very well in the cask competition of that year.

195
00:20:38,580 --> 00:20:42,700
Phil, you mentioned machine learning and neural nets.

196
00:20:42,700 --> 00:20:51,040
And the headlines that we saw right at the beginning of the month was that I had solved the protein folding problem.

197
00:20:51,040 --> 00:20:56,410
And you've mentioned that various forms of machine learning.

198
00:20:56,410 --> 00:20:57,820
So using a computer,

199
00:20:57,820 --> 00:21:07,570
particularly based on a neural net to scan many different to learn from many different existing is to try and protect you once had been used before.

200
00:21:07,570 --> 00:21:19,930
Can you tell us a little bit about what is special that the approach that the mind and Alpha fold have taken and what they have achieved that is new?

201
00:21:19,930 --> 00:21:28,900
Yeah. OK. So the first thing I think that just needs to be clarified a little bit here is that they actually haven't solved protein folding per say.

202
00:21:28,900 --> 00:21:34,870
It's more correct to say that they've solved the protein structure prediction problem rather than the folding problem,

203
00:21:34,870 --> 00:21:37,540
which is a bit more complicated, actually.

204
00:21:37,540 --> 00:21:42,880
And if you even want to be more precise, you could argue they solved the prediction of crystal structures of proteins.

205
00:21:42,880 --> 00:21:52,170
So the question really is, how did they do it then? And the real answer is we don't actually know the precise details of how well it works,

206
00:21:52,170 --> 00:21:56,680
because I've been very cagey about what they've been saying in the in the announcements.

207
00:21:56,680 --> 00:22:07,090
But presumably we will learn exactly how that works when you have paper papers appear in due course over presumably the next year.

208
00:22:07,090 --> 00:22:12,160
We do know, however, a little bit about how their previous entry work.

209
00:22:12,160 --> 00:22:19,070
So going back to the previous CAFS version, Alpha fold one, if you want to call it that,

210
00:22:19,070 --> 00:22:26,110
that we do we do know roughly how they work because they did release a version of the code and so you could look at that and tinker with it.

211
00:22:26,110 --> 00:22:32,410
And that kind of works and builds on actually what a lot of other people were doing,

212
00:22:32,410 --> 00:22:39,040
which was to employ this concept of multiple sequence alignments to great effect.

213
00:22:39,040 --> 00:22:47,140
So one of the big developments, which was which was I should also mention, which is terribly important, was the developments in sequencing.

214
00:22:47,140 --> 00:22:50,960
And because off around the two thousands and beyond that,

215
00:22:50,960 --> 00:22:56,200
they suddenly appeared this vast amount of data in terms of the sequences that were available.

216
00:22:56,200 --> 00:23:02,290
And this actually allowed you to by doing a very large, multiple sequence, allow alignment.

217
00:23:02,290 --> 00:23:11,440
You can actually look at which residues in which positions seem to COBOL and the IP, which is a residue of this position, would change over here.

218
00:23:11,440 --> 00:23:17,380
And it looked like it would change to something else over here at roughly the same in graphics, in the same sequence.

219
00:23:17,380 --> 00:23:20,380
And the idea was by looking at how strongly these evolved,

220
00:23:20,380 --> 00:23:25,600
you could actually then predict where the idea was that they might be close together in space.

221
00:23:25,600 --> 00:23:32,050
And you could use that to kind of then create a kind of pairwise distance constraint on the building of the model.

222
00:23:32,050 --> 00:23:36,580
And actually, that's what a lot of people had done prior to the success of our default.

223
00:23:36,580 --> 00:23:41,170
And that's what Apple folded in their first entry, to all intents and purposes,

224
00:23:41,170 --> 00:23:46,310
intents and purposes, combined with a little bit of tweaks around the outside, as it were.

225
00:23:46,310 --> 00:23:50,050
Now, also fold one appears to be a little bit different.

226
00:23:50,050 --> 00:23:57,610
And this is the step change, which we obviously have to wait to generally find out what the the details of that are.

227
00:23:57,610 --> 00:24:01,990
But it appears genuinely end to end in that a sequence goes in one end and our hopes,

228
00:24:01,990 --> 00:24:07,680
how hopes the structure which is which is which is which is incredibly impressive.

229
00:24:07,680 --> 00:24:12,550
But what we do know is that it appears to, four, adopt a kind of much more dynamic,

230
00:24:12,550 --> 00:24:18,520
dynamic learning approach within it to work out kind of as it's going along,

231
00:24:18,520 --> 00:24:24,250
which pieces of information from the sequences are more important than others.

232
00:24:24,250 --> 00:24:31,420
And this. Hands on a new piece of deep, relatively new piece of deep learning architecture, which I don't really know anything about it.

233
00:24:31,420 --> 00:24:35,430
It's called a 3D quivering transformer for those who are interested.

234
00:24:35,430 --> 00:24:43,080
It was also developed by a team at Google in 2017. So it's not surprising that it's been used in Alpha, too.

235
00:24:43,080 --> 00:24:49,350
It's been demonstrated the use of that kind of architecture has been used very successfully in language translators.

236
00:24:49,350 --> 00:24:54,960
That's probably where most people might say they're interested, might have come across it before.

237
00:24:54,960 --> 00:24:57,600
So that was that was definitely one of the big differences.

238
00:24:57,600 --> 00:25:04,680
The other difference is, is that it requires a lot of computational power as well, so that they do require a lot of resource to do it.

239
00:25:04,680 --> 00:25:10,230
I think they use something like 128 vertical tensor processing units.

240
00:25:10,230 --> 00:25:14,760
That's roughly about 200 GP use, I think, over a few weeks.

241
00:25:14,760 --> 00:25:16,500
So they had. So there are some differences.

242
00:25:16,500 --> 00:25:24,270
We don't know quite the full extent of the methodology yet, but it's kind of builds on what people have done before.

243
00:25:24,270 --> 00:25:32,280
But there is a step change, definitely, and an employment of new are new deep learning architectures to teach to make this significant jump,

244
00:25:32,280 --> 00:25:44,510
if you like, in in the predictive power. I've heard that in trying to solve protein structures that the expert humour the you and Yvonne

245
00:25:44,510 --> 00:25:55,150
and sometime provide insight that sort of straight mathematical brute force fails to provide.

246
00:25:55,150 --> 00:26:08,440
And that what might be happening is that the. New autofill machine learning is in some ways replicating that by a chemical insight that humans had.

247
00:26:08,440 --> 00:26:12,410
That is a truth. And that was, after all.

248
00:26:12,410 --> 00:26:16,800
I don't yeah. I sort of heard that sort of thing said before as well.

249
00:26:16,800 --> 00:26:21,480
And I I'm not sure if that's the correct way to think about it.

250
00:26:21,480 --> 00:26:25,980
To be honest, I think it's partly because I don't actually know enough of the underlying

251
00:26:25,980 --> 00:26:32,850
details of the of the new architecture to give you an honest answer about that.

252
00:26:32,850 --> 00:26:37,290
So, yeah, I'm not sure whether that's true or not. That's a not very good answer.

253
00:26:37,290 --> 00:26:42,030
And open. That's right. Oh, yeah.

254
00:26:42,030 --> 00:26:47,040
You've all come from a point of even more ignorance when it comes to the third approach.

255
00:26:47,040 --> 00:26:53,850
But I mean, from what I understand of the the the the previous approach,

256
00:26:53,850 --> 00:27:00,660
the one from two years ago, as Phil was saying, it's it's this idea that, you know,

257
00:27:00,660 --> 00:27:10,360
the yellow bead and sort of the blue eat in a red bead from different parts of the beads on the string, like you actually end up close to each other.

258
00:27:10,360 --> 00:27:16,290
Only when the when the proteins are is actually in its Three-Dimensional Shape.

259
00:27:16,290 --> 00:27:23,520
And actually, as a Yeah. As a as a structural biologist, often we find.

260
00:27:23,520 --> 00:27:28,590
Finally finalising the structure of a protein. I'll be looking at it and I'll be thinking, yeah.

261
00:27:28,590 --> 00:27:35,790
Yeah, that makes sense that an arginine should be should be, you know, close to a despotic acid.

262
00:27:35,790 --> 00:27:39,700
So a blue and a red type B, it's a close together.

263
00:27:39,700 --> 00:27:43,890
So it's a similar thing. Now that the eye is matching,

264
00:27:43,890 --> 00:27:51,630
you could sort of think that that artificial intelligence has learnt that because it's learnt these rules by going through looking at the I

265
00:27:51,630 --> 00:28:03,210
believe they quoted a one hundred and seventy thousand protein structures in the protein database that they were able to use to learn from,

266
00:28:03,210 --> 00:28:09,990
as well as, of course, all these sequence alignments. But what I also read somewhere that that the latest iteration,

267
00:28:09,990 --> 00:28:15,030
the one that's provided the breakthrough is is not just sort of looking at a pair of pairwise,

268
00:28:15,030 --> 00:28:22,320
but something more like building up kind of a jigsaw puzzle and that you you put together things in you know,

269
00:28:22,320 --> 00:28:28,810
you put together all the bits that are sort of the tree and the bits that are the river bits that are the.

270
00:28:28,810 --> 00:28:34,270
Cow crossing the river. And then it sort of clumps it all, but puts it all together, you see what I mean?

271
00:28:34,270 --> 00:28:42,210
And I imagine, again, that, you know, there are blocks of of of areas in proteins that it just then makes sense that these would come together.

272
00:28:42,210 --> 00:28:47,700
But I'm. Yeah. Who knows? It's going to be extremely interesting to see exactly how they did it.

273
00:28:47,700 --> 00:28:55,970
And it's not a related question. Has that. Machine learning, to a certain extent, is a black box.

274
00:28:55,970 --> 00:29:03,980
You have inputs in and outputs out. But there's a lot of work in trying to actually reconstruct what is happening within that black box.

275
00:29:03,980 --> 00:29:13,130
And do you think that when that is done with deep foaled, that it might say anything about how proteins physically fold in nature?

276
00:29:13,130 --> 00:29:16,190
Or do you think you'll be just much more patterned pattern matching?

277
00:29:16,190 --> 00:29:24,400
So is getting an answer without it telling you anything about the physical processes involved?

278
00:29:24,400 --> 00:29:33,130
My feeling is it won't tell you anything about folding. Actually, I think it ultimately is is it is a bit more of just a pattern matching.

279
00:29:33,130 --> 00:29:36,970
I don't think you'll get much information about the actually how it folds.

280
00:29:36,970 --> 00:29:41,270
This is this is a data versus physics problem almost in that sense.

281
00:29:41,270 --> 00:29:42,790
You know, now we are you know,

282
00:29:42,790 --> 00:29:50,830
previously we we were kind of thinking that this problem would be solved by just using our crude understanding of other entry physics.

283
00:29:50,830 --> 00:29:58,340
But this is this is actually the other thing that's now become prevalent, which is making use of all the all the data that's out there.

284
00:29:58,340 --> 00:30:00,610
And I don't think actually you'll get much.

285
00:30:00,610 --> 00:30:10,280
And that's good news for academics working on folding plus eight, because that's that they making probably a bit more comfortable and sleep at night.

286
00:30:10,280 --> 00:30:15,130
But I don't think they will get much. There's much to be had looking at Olding for years.

287
00:30:15,130 --> 00:30:20,590
Thank you. Yvonne, when the news came out, I think it's November 30th and December the 1st,

288
00:30:20,590 --> 00:30:31,420
then it sort of made the front page of newspapers and was on the on and on radio and television within the field.

289
00:30:31,420 --> 00:30:36,790
How much of a surprise was this to you guys? Did you did you know it was coming?

290
00:30:36,790 --> 00:30:48,140
Were you surprised by how good the project predict predictions were in the final cast round?

291
00:30:48,140 --> 00:30:55,460
I guess I am actually, I suppose. Yeah, I would not have predicted that it would be this particular cast.

292
00:30:55,460 --> 00:31:01,880
I mean that because there have been periods where instead of progress every two years,

293
00:31:01,880 --> 00:31:07,640
you know, that they looked at one stage like the field with stalling. And then and then things started happening again.

294
00:31:07,640 --> 00:31:11,210
Then they were sort of new approaches and big jumps forward.

295
00:31:11,210 --> 00:31:14,630
And you could imagine that that artificial intelligence,

296
00:31:14,630 --> 00:31:22,580
that that machine learning was going to be able to to mine these these huge databases now and make a big step forward.

297
00:31:22,580 --> 00:31:26,150
And a lot of people would trying to do this.

298
00:31:26,150 --> 00:31:32,600
But I guess when I first woke up, I thought, well, I've heard this on the Today programme or whatever.

299
00:31:32,600 --> 00:31:38,780
I thought, well, how good is it? And I again, I've been rummaging around a bit trying to find out.

300
00:31:38,780 --> 00:31:42,890
And there are some. Figures.

301
00:31:42,890 --> 00:31:48,950
The pictures provided, which show very impressive, what we call super positions.

302
00:31:48,950 --> 00:31:53,540
So they've taken the experimental structure and they've overlapped.

303
00:31:53,540 --> 00:31:58,340
They've superimposed onto it the the the bottle that they had come up with.

304
00:31:58,340 --> 00:32:05,290
And you know that to sit. But we identify on top of each other.

305
00:32:05,290 --> 00:32:11,540
But the caveats to that are that they that that individual protein structures.

306
00:32:11,540 --> 00:32:18,860
So they managed. To come up with something that, you know, if if we were talking about insulin,

307
00:32:18,860 --> 00:32:27,550
they would have been able to predict what insulin looks like and it would look very much as Dorothy Hodgkins.

308
00:32:27,550 --> 00:32:31,370
So I don't know.

309
00:32:31,370 --> 00:32:37,850
Well, again, I saw one picture that implied that the actually the the real details cause that the beads aren't just a little sort of beads.

310
00:32:37,850 --> 00:32:41,380
Isn't that all the beads on top of each other in the correct positions?

311
00:32:41,380 --> 00:32:51,470
It's that you've you've got actually the the detail of what each residues, each side chains of the residues that wear their position.

312
00:32:51,470 --> 00:33:01,070
Getting that with that. I think that probably the numbers that I've seen quoted are that they're accurate to within

313
00:33:01,070 --> 00:33:08,510
maybe two angstroms overall of the match in position on average of everything in the.

314
00:33:08,510 --> 00:33:13,670
Well, in the in the main chain in the final. But it time that would have to do certain things.

315
00:33:13,670 --> 00:33:18,240
And I think that's going to be fun to just chat over for a few minutes now. They've done nothing.

316
00:33:18,240 --> 00:33:23,150
Got to interrupt that. So strong. That's about the size of an atom. Just to give people.

317
00:33:23,150 --> 00:33:25,820
Oh, sorry. Yeah, absolutely. Yeah. Yeah.

318
00:33:25,820 --> 00:33:38,640
And the reason I'm setting about how accurate it is is that you need things to be accurate to within less of an angstrom actually to be able to.

319
00:33:38,640 --> 00:33:43,740
Too accurate to use structures then to actually act accurately,

320
00:33:43,740 --> 00:33:56,410
design drugs that you might want to to use to to jam some, you know, to to fit into part of of a protein structure.

321
00:33:56,410 --> 00:34:06,180
And, for example, stop it working if, for example, you want to to block a protein machine that's important for for a viral function or something.

322
00:34:06,180 --> 00:34:14,700
So you need a very high level of accuracy for certain sorts of things to be evolved.

323
00:34:14,700 --> 00:34:21,060
I don't know, I. I think they are. I think going to be able to get that they they have clearly got there.

324
00:34:21,060 --> 00:34:26,310
But some of the medium sized proteins, it's extremely impressive.

325
00:34:26,310 --> 00:34:35,680
But the next challenge, and they say this themselves in their press release is, is to be able to then look at not just one protein by itself,

326
00:34:35,680 --> 00:34:40,630
but these complexes, proteins, because proteins that don't actually usually wander around by themselves.

327
00:34:40,630 --> 00:34:47,510
They there are gangs of them. They're not very good at socially isolating proteins together.

328
00:34:47,510 --> 00:34:55,410
So just to cheque, I understand you correctly. You're impressed by the structures that they are producing at the moment,

329
00:34:55,410 --> 00:35:02,430
but you worry that they might not be quite as good to show where every atom is.

330
00:35:02,430 --> 00:35:07,050
And you need to know where every atom is if you're trying to design a drug.

331
00:35:07,050 --> 00:35:12,090
Because typically a drug needs to fit in like a lock, a key in the lock.

332
00:35:12,090 --> 00:35:18,900
And so you really need high accuracy, which at the moment you you do require a physical means.

333
00:35:18,900 --> 00:35:27,510
I'm looking at the structure rather than the computational. I think it's probably a little it's still that's still a little bit off from being

334
00:35:27,510 --> 00:35:34,170
able to to be able to help with that directly with the drug design side of things.

335
00:35:34,170 --> 00:35:41,010
But having said that, there's a lot of protein structures that aren't actually good enough,

336
00:35:41,010 --> 00:35:46,350
experimentally determine protein structures that aren't good enough to be able to really guide drug design.

337
00:35:46,350 --> 00:35:51,990
You need you need only the very best structures are good enough for that.

338
00:35:51,990 --> 00:36:00,930
So it's a I'm setting them a very high bar in saying that. And I think that it's extremely impressive what they've been able to do.

339
00:36:00,930 --> 00:36:07,320
And they are clearly able to, for many protein sequences,

340
00:36:07,320 --> 00:36:18,210
predict what that protein is going to look like to sufficient accuracy to be able to say, give up the idea of maybe how it works or what it does.

341
00:36:18,210 --> 00:36:26,960
But. There are still big questions about how it will be able to interact with other proteins.

342
00:36:26,960 --> 00:36:32,700
At the moment they're not able to predict that whilst we can do structures either using cryo

343
00:36:32,700 --> 00:36:40,410
anymore or X-ray crystallography of the clusters of proteins and see how they fit together.

344
00:36:40,410 --> 00:36:47,230
And it's often in the way that they sit together. That's important for for their function.

345
00:36:47,230 --> 00:36:53,930
For instance. So, Phil, let me go to you as a computational barkhad chemist.

346
00:36:53,930 --> 00:37:00,400
Yeah. And. How are you feeling as well about the relative weather?

347
00:37:00,400 --> 00:37:07,330
I mean, I think actually, if we if it's fair, it's a very impressive result, whichever way you look at it.

348
00:37:07,330 --> 00:37:18,360
I don't think anybody 20 hurt in the cast. 13 would have predicted that the alcohol would have done as well as they have done in this in this event.

349
00:37:18,360 --> 00:37:24,160
This is notable. And the reason is a lot of hype is notable, has some for a lot of for a few reasons, actually.

350
00:37:24,160 --> 00:37:31,840
The first is that alcohol, too, is not just a head, but it is more than twice as good as the next best entry.

351
00:37:31,840 --> 00:37:39,160
And that was true across nearly all the targets. I think I'm correct in saying that not that would be impressive enough, actually.

352
00:37:39,160 --> 00:37:41,980
But this time, the reason why there's a lot of hype, I think,

353
00:37:41,980 --> 00:37:52,150
is actually that for the accuracy, as measured by something called the global distance score,

354
00:37:52,150 --> 00:37:55,300
which is which is another metric which I won't go into,

355
00:37:55,300 --> 00:38:01,920
but anything above 90 is considered informally at least, to be rivalling experimental accuracy.

356
00:38:01,920 --> 00:38:07,030
And Alpha sort of claimed a median score of ninety two point five across all targets.

357
00:38:07,030 --> 00:38:14,560
That's why there's a lot of hype. Now, if Om's right, if you want to delve down and look at the armis, the root mean squared deviation,

358
00:38:14,560 --> 00:38:19,990
how different it is when you superimpose one of the predictions on the target, then.

359
00:38:19,990 --> 00:38:25,930
Yeah, about 50 percent of the time across all atoms. This is you can get on that.

360
00:38:25,930 --> 00:38:29,170
They're under two angstroms, which is what Yvonne was mentioning,

361
00:38:29,170 --> 00:38:34,480
which is still pretty impressive actually, if you make your toddler a little bit slacker.

362
00:38:34,480 --> 00:38:42,520
So less than five angstroms, then 92 and a half percent of the time they get the right answer, which which is quite impressive.

363
00:38:42,520 --> 00:38:45,830
And I think there's scope for improving that. Actually, that's the good news as well.

364
00:38:45,830 --> 00:38:54,310
I think actually, it's not clear whether you could refine that a little bit, making that bit of physics and perhaps do it to improve the models.

365
00:38:54,310 --> 00:39:03,130
And it's not also clear whether some of that difference reflects some of the crystallisation artefacts or some,

366
00:39:03,130 --> 00:39:12,040
so that there's a lot of things that one one could probably delve down to a little bit and consider that consider it a bit more detail, I think.

367
00:39:12,040 --> 00:39:15,360
I think the level of accuracy is useful enough even now, though,

368
00:39:15,360 --> 00:39:20,470
that you might even start to question some experimental results you might be thinking of.

369
00:39:20,470 --> 00:39:23,380
Is it really also worth the experiment? Right.

370
00:39:23,380 --> 00:39:31,450
And then I read on somebody's blog and apologies if you are listening and I'm quoting you from misquoting you here.

371
00:39:31,450 --> 00:39:34,240
But I didn't read it. Actually, that's already happened.

372
00:39:34,240 --> 00:39:39,520
One of the experimental groups had seen the results on the outfall, went back and looked at that data.

373
00:39:39,520 --> 00:39:43,170
And indeed, they have actually mis assigned probably residue.

374
00:39:43,170 --> 00:39:46,660
I think it was so. Yeah.

375
00:39:46,660 --> 00:39:52,200
So. So, yeah, I read I read that blog. So I thought that was a good idea.

376
00:39:52,200 --> 00:39:59,500
That actually now we're at the level where we're actually using the computational predictions to question whether the experiment itself is right.

377
00:39:59,500 --> 00:40:04,100
I think he's very that's an important step change, actually.

378
00:40:04,100 --> 00:40:11,340
I think going forward, it would probably be better for the predictors to compare to the raw electron density, actually, rather than.

379
00:40:11,340 --> 00:40:16,480
Yeah. Will that make Chris Lawrence generate the great theoretical ecologist?

380
00:40:16,480 --> 00:40:19,870
Bob May. Lord May. He died early this year. You still have to quit.

381
00:40:19,870 --> 00:40:24,790
My theory completely disproves your data. And I'm sorry he's not alive.

382
00:40:24,790 --> 00:40:28,870
Enjoy. That's what I would like to go to some questions.

383
00:40:28,870 --> 00:40:36,700
And the first one, I think it's for Yvonne. And this Sunesys has alpha fold.

384
00:40:36,700 --> 00:40:50,270
Is it likely to facilitate or replace crystallography? And how do you think it will become incorporated into the drug design pipeline?

385
00:40:50,270 --> 00:40:55,420
Right. Okay. I think for the moment, anyway,

386
00:40:55,420 --> 00:41:07,030
it's just going to be incredibly complementary and helpful to breaking crystallography and also to to Criolla n that preaching crystallography.

387
00:41:07,030 --> 00:41:14,740
We we often benefit from having a model that we can start from.

388
00:41:14,740 --> 00:41:19,740
That might not be the exact isn't the exact structure that we're trying to determine.

389
00:41:19,740 --> 00:41:31,190
But it gives us a starting point that allows us to interpret our data more rapidly and if we can.

390
00:41:31,190 --> 00:41:36,860
Get those models using alpha fold.

391
00:41:36,860 --> 00:41:46,080
That that could be incredibly helpful. It could also, I think.

392
00:41:46,080 --> 00:41:56,190
Be used to generate structures for individual proteins, which you could then dock together into complexes that we are visualising,

393
00:41:56,190 --> 00:42:04,380
using criterium, which may actually be at a rather lower resolution to rather less detail.

394
00:42:04,380 --> 00:42:12,340
We might not be able to see quite all of the detail in.

395
00:42:12,340 --> 00:42:16,000
Using some of our structural techniques, looking at really big complexes,

396
00:42:16,000 --> 00:42:22,000
it's often very useful to be able to get really detailed structures of maller parts

397
00:42:22,000 --> 00:42:27,020
of those complexes and then dock them into less detailed experimental structures.

398
00:42:27,020 --> 00:42:38,330
And I think that's that could be a very important way forward. That term Alpha Fold is able to open up for us in terms of the drug design.

399
00:42:38,330 --> 00:42:46,870
Yeah, I kind of touched on that earlier, didn't I? I think it is going to be very exciting in probably the only the short term, actually,

400
00:42:46,870 --> 00:42:57,030
that they are going to improve, because I didn't mean to sound too churlish. I think they it's it's incredibly exciting what they're doing already.

401
00:42:57,030 --> 00:43:01,800
Thanks very much. That question was from you. Can the law, I should have said.

402
00:43:01,800 --> 00:43:11,460
I've got a question from Larissa goal, which I think is fullfil machine learning is only as good as the data on which it is trained, trained.

403
00:43:11,460 --> 00:43:18,870
And if there are biases and that's going to affect it. Are you aware of any such biases, biases in the training set?

404
00:43:18,870 --> 00:43:24,750
And if I can expand that. Does this mean that you can't take, say,

405
00:43:24,750 --> 00:43:35,430
a very novel peptide sequence from a poorly understood virus and and feed it into deep fault if it's seen nothing like that in the past?

406
00:43:35,430 --> 00:43:40,530
Yeah. So that's a very good point. And yeah, that is very true.

407
00:43:40,530 --> 00:43:45,900
Machine learning is traditionally only as good as the data you put in.

408
00:43:45,900 --> 00:43:51,270
And there's of course, there is a bias in in in some sense in, for example,

409
00:43:51,270 --> 00:43:59,580
the PDP in those structures that are readily solvable by by definition, almost, because that's what the PDP.

410
00:43:59,580 --> 00:44:05,760
So there is a slight bias there. Now, if you'd asked me the question about alcohol won,

411
00:44:05,760 --> 00:44:11,640
then then I would have said yes to to to Charles's follow up question would would

412
00:44:11,640 --> 00:44:20,770
it would have really struggled to predict something that it's never seen before, but would Alpha fold to the jury is a bit out, I think on this.

413
00:44:20,770 --> 00:44:24,880
It's not clear how well it would actually do on unrelated things.

414
00:44:24,880 --> 00:44:26,280
I think that would be the interest.

415
00:44:26,280 --> 00:44:33,930
A very interesting to see how it does on on stuff that it clearly has no idea about that and ask it to make a prediction on that.

416
00:44:33,930 --> 00:44:40,830
I think that will be an interesting test. I don't think anybody has a feeling for how badly if you're a pessimist or how

417
00:44:40,830 --> 00:44:45,300
well you're not missed how well it will world will do in that kind of scenario.

418
00:44:45,300 --> 00:44:53,800
But certainly a machine learning the input data of the data set that you have for training for some of these things is super important.

419
00:44:53,800 --> 00:44:58,410
And yeah, you have to be aware all the time of biases. Yeah. Okay.

420
00:44:58,410 --> 00:45:06,080
The next question is for Yvonne. And if I'm going to ask you to actually explain the question as well as to answer it.

421
00:45:06,080 --> 00:45:14,340
What are these? What do we think about whether Alpha Fold will be able to solve the structures for G protein coupled receptors,

422
00:45:14,340 --> 00:45:20,050
seeing as they're so hard to crystallise? So might you begin by just saying what the G protein coupled receptor?

423
00:45:20,050 --> 00:45:26,490
It's okay. It's it's a protein, that's all.

424
00:45:26,490 --> 00:45:38,250
A family of huge family protein, huge family proteins that are embedded in the plasma membrane that the membranes that surround ourselves.

425
00:45:38,250 --> 00:45:47,100
So they are proteins that are involved in picking up signals from the outside and signalling into the cell.

426
00:45:47,100 --> 00:46:00,090
And they are major drug targets. Many of these these proteins, because they control all sorts of signalling, not in our brains, for example.

427
00:46:00,090 --> 00:46:04,830
So they they can be they can be targets fully and anti-depressants and such like they can.

428
00:46:04,830 --> 00:46:11,950
But they they control the.

429
00:46:11,950 --> 00:46:22,130
Your activities in our hearts and there are many, many, many of these family members are very potent drug targets.

430
00:46:22,130 --> 00:46:30,890
And for many, many years, people have worked to get structures, effort to understand and to design better drugs against them.

431
00:46:30,890 --> 00:46:36,980
To understand the actions of the drugs we've got and to sign better ones.

432
00:46:36,980 --> 00:46:42,230
But they have proved to be notoriously difficult to crystallise.

433
00:46:42,230 --> 00:46:50,390
And indeed, a clutch of Nobel prises were awarded not that long ago for the very first structures that came out.

434
00:46:50,390 --> 00:46:59,270
And the first structures were by approaching crystallography. But they they needed new techniques to be used to persuade these things to crystallise.

435
00:46:59,270 --> 00:47:08,690
More recently, there's been quite a lot of progress being made with using criterium to solve structures that these proteins.

436
00:47:08,690 --> 00:47:16,760
And so. Yes, that they're a reasonable number of family members now for which structures are known.

437
00:47:16,760 --> 00:47:25,730
But as I was saying, there are many, many, many members in this family, many hundreds protein's for which we don't have structures.

438
00:47:25,730 --> 00:47:33,260
And so the argument would be that indeed we could use the alpha fold to predict these structures.

439
00:47:33,260 --> 00:47:44,410
And it could. One would hope it would do. Perhaps better than than just starting from a known structure and trying to model from that structure.

440
00:47:44,410 --> 00:47:49,020
But I'm since the devil is in the detail, I don't know.

441
00:47:49,020 --> 00:48:04,440
I don't. I hope it's going to help us. But I don't know how well our phone is working at the moment for these highly membrane embedded type proteins.

442
00:48:04,440 --> 00:48:12,410
And also the problem with G protein receptors. And part of the reason why they've been very problematic to crystallise is that they.

443
00:48:12,410 --> 00:48:18,260
Part of their function involves them being very dynamic, so they they shape change, look huge by a huge amount,

444
00:48:18,260 --> 00:48:27,530
but they do shape change a bit and that's important for the way that they work and they interact with other proteins.

445
00:48:27,530 --> 00:48:33,390
As part of that, they pass on the signal to other proteins through their shape change.

446
00:48:33,390 --> 00:48:44,060
And there's a movement so it doesn't work, doesn't tell you you shape changes and doesn't tell you Pruett's in protein interactions and the,

447
00:48:44,060 --> 00:48:48,420
you know, the deep mind people in that in their block, in their in their presence.

448
00:48:48,420 --> 00:48:54,350
What we're seeing these are the next are the challenges that they're going to have to move onto in the longer term.

449
00:48:54,350 --> 00:49:01,980
Phil, you've already asked around about things that will be a use for for drug designs to look for a bridge.

450
00:49:01,980 --> 00:49:06,280
Do you see cars that are going to be perhaps not easily solved of them?

451
00:49:06,280 --> 00:49:14,030
You know, I was just gonna say one of the yeah, that is definitely one of the challenges for other other membrane proteins like transporters.

452
00:49:14,030 --> 00:49:17,420
You know, we know they don't have one unique structure.

453
00:49:17,420 --> 00:49:27,370
They actually have at least two probably more set separate structures that must be Matus stable as part of their part of their function.

454
00:49:27,370 --> 00:49:33,920
And so which one do you predict? You know, so I won't be challenging for alcohol, too, I think,

455
00:49:33,920 --> 00:49:42,410
to predict the different states of transporter proteins or things like PCR, which can exist in multiple different confirmations.

456
00:49:42,410 --> 00:49:48,200
I think transporter's might be a bit easier because that changes are probably a bit more extreme.

457
00:49:48,200 --> 00:49:52,860
So you would think perhaps that that might be a little bit easier G.P.S.

458
00:49:52,860 --> 00:49:57,920
So the movement is quite subtle, actually. Most of them, at least anyway.

459
00:49:57,920 --> 00:50:06,680
But isn't this a problem that's perhaps amenable to the next generation of machine learning in this area that you can.

460
00:50:06,680 --> 00:50:11,060
You do understand these metho stable states so the states that flit between it,

461
00:50:11,060 --> 00:50:16,740
then you can train a A.I. to try and learn that and hopefully tell you new things.

462
00:50:16,740 --> 00:50:21,170
Yeah, I think I think you're right. I think that is actually that.

463
00:50:21,170 --> 00:50:28,340
And the problem legalisation are actually relatively low hanging fruit now in the sense that, you know,

464
00:50:28,340 --> 00:50:34,010
they've done the big problem, which is basically protecting the structure of the give or take lives.

465
00:50:34,010 --> 00:50:41,090
And actually, you know, to to extend that in these other directions, you would kind of think it's not so hard.

466
00:50:41,090 --> 00:50:48,670
Is that the first step? You know, these be second and third steps we're talking about now, I would think would be a little bit less difficult.

467
00:50:48,670 --> 00:50:53,750
But I don't know, maybe it's just then just a follow up question to that.

468
00:50:53,750 --> 00:50:59,510
My friends who work on Glycoprotein is always say that the amino acid sequence

469
00:50:59,510 --> 00:51:03,440
is only the beginning and they see the things you stick to your protein off.

470
00:51:03,440 --> 00:51:12,140
You made it. That give all the activity. Then how does how does that affect what we're talking about now?

471
00:51:12,140 --> 00:51:16,020
Yeah, I suppose what we would call post translational modifications. You're right.

472
00:51:16,020 --> 00:51:21,070
I mean, one of the approaches that I'm really interested in, which is one of the messenger proteins that goes between cells,

473
00:51:21,070 --> 00:51:24,540
it isn't just made out of the beads on the string,

474
00:51:24,540 --> 00:51:32,370
but it's it's also got an extra parmeter late attached to it, which is essential for its its its activity.

475
00:51:32,370 --> 00:51:37,260
So it's it's got a it's not just entirely made of amino acids.

476
00:51:37,260 --> 00:51:42,110
There are other things that add onto proteins that further change,

477
00:51:42,110 --> 00:51:47,470
add to their shape and add to their properties and are essential for their function.

478
00:51:47,470 --> 00:51:53,880
And so there's a whole world there. That Alpha vote isn't exploring.

479
00:51:53,880 --> 00:52:01,680
OK. We've got a few more questions. And so, if possible and I realise it might be a if you can answer them quite briefly.

480
00:52:01,680 --> 00:52:05,700
I don't know who would like to pick this one up, which is the highest vote at the moment.

481
00:52:05,700 --> 00:52:13,710
And we'll need a bit of explanation. How long do you expect before Alpha Fold is going to be capable of predicting hysteric interactions?

482
00:52:13,710 --> 00:52:18,190
So you can. Failed you? Yes.

483
00:52:18,190 --> 00:52:22,910
How long before it predicts, Alastair, it interactions? Just to clarify, Alex.

484
00:52:22,910 --> 00:52:30,260
Derek InterAction's meaning that there would be another part of the protein away from perhaps where the agonist binds,

485
00:52:30,260 --> 00:52:34,750
or that the main leg and binding side is another side or the other Sterrett side.

486
00:52:34,750 --> 00:52:41,000
Which would. Which would. Which would be. Somewhere where something else could buy it and do something.

487
00:52:41,000 --> 00:52:44,800
How old. How far are we away from understanding that? I think that's the thing.

488
00:52:44,800 --> 00:52:51,490
That's a particularly challenging sub problem of when you get down to higher accuracy.

489
00:52:51,490 --> 00:52:59,050
So, you know, I said earlier that 50 percent of the results were under two Angstroms or Mesta.

490
00:52:59,050 --> 00:53:01,990
I think you'd have to be confident the other, you know,

491
00:53:01,990 --> 00:53:09,210
much higher and 50 percent before you could begin to hope to to to sort of tackle that particular question.

492
00:53:09,210 --> 00:53:14,630
All right. Sorry, fella. I interrupted. No, no, I just said that that would be my view.

493
00:53:14,630 --> 00:53:22,060
Yeah. Am I right? The significance of the question is that were you to understand hysteric interactions that new new drug targets,

494
00:53:22,060 --> 00:53:26,320
because you could then look at small molecules that would then.

495
00:53:26,320 --> 00:53:28,450
Exactly. Yeah. It so that the benefit.

496
00:53:28,450 --> 00:53:36,450
Looking at our Starick sites in that way would be that you'd be not looking at the same site as the author Sterrett Ligand.

497
00:53:36,450 --> 00:53:45,430
And that would give you more freedom presumably to to design something new that they didn't have the.

498
00:53:45,430 --> 00:53:48,200
There wouldn't be so constrained as the authors of the authors,

499
00:53:48,200 --> 00:53:55,880
Darkside tends to be constrained by all of the evolutionary properties that are designed to make it the main orthostatic site.

500
00:53:55,880 --> 00:54:00,910
But the historic site presumably can be a bit more freer in terms of how it's evolved and that there could be

501
00:54:00,910 --> 00:54:08,380
potential for many more different compounds to bind and change the activation or otherwise of the target protein.

502
00:54:08,380 --> 00:54:09,670
So it's an interesting question.

503
00:54:09,670 --> 00:54:17,050
Here is what are your understanding of how Alpha file deals with nonsense sequences are completely fictional sequences?

504
00:54:17,050 --> 00:54:21,940
So I guess there are two issues here. If you may comp a string of amino acids.

505
00:54:21,940 --> 00:54:33,660
Will it always assume a fixed three dimensional structure or is it only a subset of sequences that form a structure?

506
00:54:33,660 --> 00:54:46,470
We know that it's only a subset because because there are proteins that the cells make that term that are the don't assume one fixed structure,

507
00:54:46,470 --> 00:54:56,620
and that's part of that function, actually, that they they can shape change and partly according to to what they're.

508
00:54:56,620 --> 00:55:02,020
What are the proteins are interacting with, but also there are regions in them that are just very,

509
00:55:02,020 --> 00:55:09,100
very flexible and floppy and just stay like beads on a string. I mean, Phil, what would you say to that?

510
00:55:09,100 --> 00:55:16,990
Yeah, no, I absolutely I think I if that's the intention of the question, I say, but yeah, you know, there are a lot of it.

511
00:55:16,990 --> 00:55:24,460
The genome is probably unstructured. So that not be different question then and probably not relevant.

512
00:55:24,460 --> 00:55:32,520
Right. I think it does come to can you then turn it the other way around, though, and start to design structures?

513
00:55:32,520 --> 00:55:36,540
So so run it backwards and say, OK. This is the structure that I want.

514
00:55:36,540 --> 00:55:44,100
Can I kind of work out what the sequence would have to be?

515
00:55:44,100 --> 00:55:48,420
To get that structure, because that's how you would then design it.

516
00:55:48,420 --> 00:55:56,220
Enzymes to order, little machines that would chew up plastic. I think they've been saying a lot in the press releases.

517
00:55:56,220 --> 00:56:01,920
You could say that that is something that already David Baker has been putting a lot of effort into.

518
00:56:01,920 --> 00:56:07,320
This is somebody else who's been looking at protein coding problem over the area over the years.

519
00:56:07,320 --> 00:56:16,280
And clearly, it's a direction that deep, so deep mind and announce folder without the filter can be interested in protecting well.

520
00:56:16,280 --> 00:56:19,890
So that nicely leads on to my last question.

521
00:56:19,890 --> 00:56:31,890
We've only got three minutes left, and I want you to throw us aside your shackles as some very careful scientists to sort of project yourself

522
00:56:31,890 --> 00:56:39,690
10 years in the future and give us an example of what you think might be a really interesting result,

523
00:56:39,690 --> 00:56:47,310
either in fundamental science or in medicine or applied science that will come out of both what our four fold has done.

524
00:56:47,310 --> 00:56:52,290
But the whole tenor in advance of of protein biochemistry.

525
00:56:52,290 --> 00:56:59,160
So I am asking you to to unashamedly speculate about if everything went well,

526
00:56:59,160 --> 00:57:03,450
what might be a fantastic thing that we're celebrating 10 years from now.

527
00:57:03,450 --> 00:57:11,260
Horrible question. So I'll go to Phil first, right. Very gentlemanly.

528
00:57:11,260 --> 00:57:19,020
So. Well, I think, you know, you could use it progresses as fast as less than one can imagine that you'll have structures for everything.

529
00:57:19,020 --> 00:57:26,730
And in that case, he structures ever want or need to have structures for things like structural systems,

530
00:57:26,730 --> 00:57:30,430
biology, whatever you want to call it, will become a reality. And I think that's one important thing,

531
00:57:30,430 --> 00:57:37,470
because that could actually then bring know how we unify a lot of cellular level stuff all the way down to molecular.

532
00:57:37,470 --> 00:57:44,280
So we really, really will have a very good, detailed structural understanding right away from molecules,

533
00:57:44,280 --> 00:57:48,120
right up to cells genuinely with no gaps in it.

534
00:57:48,120 --> 00:57:55,680
Totally understanding how all these things start to interact and function together, that that is not an unrealistic goal.

535
00:57:55,680 --> 00:57:56,130
I don't think,

536
00:57:56,130 --> 00:58:06,180
given the progress in the last two casts in that sense and what were the ideas we discussed about interactions between different proteins,

537
00:58:06,180 --> 00:58:11,910
that that obviously is the next one of the next big challenges is predicting how all these proteins will interact with each other,

538
00:58:11,910 --> 00:58:20,670
both in space and time within a cell. So that that that, if you like, is is is what you could aim for probably within 10 years.

539
00:58:20,670 --> 00:58:24,230
That really is exciting, Yvonne. Yeah.

540
00:58:24,230 --> 00:58:26,850
And running from that. Carrying on from there.

541
00:58:26,850 --> 00:58:35,830
I think what would also be incredibly exciting is going to be the ability to to understand then the effect of.

542
00:58:35,830 --> 00:58:41,630
A variance in in in the genome of a single point mutations.

543
00:58:41,630 --> 00:58:51,250
Just one person having one different coloured bead in the string of one of their proteins is alpha fold.

544
00:58:51,250 --> 00:58:59,870
Gonna be able to help us more rapidly understand the effect of that colour change of just one in one bead affecting the shape of a protein

545
00:58:59,870 --> 00:59:10,400
and affecting the way that it interacts with all the other proteins it needs to to talk to in a cell to go back to to Phil's answer.

546
00:59:10,400 --> 00:59:24,000
So will it ultimately help us understand all the information that's now being gathered from the sequencing of the human genome?

547
00:59:24,000 --> 00:59:30,680
And help us to translate that ultimately into medicines if we can understand better what's going wrong,

548
00:59:30,680 --> 00:59:38,520
ongoing, wrong, maybe because of just one very subtle change. Will we be able to then come up more easily with.

549
00:59:38,520 --> 00:59:42,330
Where to intervene therapeutically? Thanks so much indeed.

550
00:59:42,330 --> 00:59:45,120
I'm really sorry we're going to have to come to an end.

551
00:59:45,120 --> 00:59:52,710
Now, it has been a real privilege to listen to and discuss with two of the leading experts in this field.

552
00:59:52,710 --> 00:59:57,580
The exciting result that we had last year, last week.

553
00:59:57,580 --> 01:00:03,440
And if I'm sort of to paraphrase the discussion, much of the hype has been justified.

554
01:00:03,440 --> 01:00:07,800
But there is some real details that one needs to think about there.

555
01:00:07,800 --> 01:00:12,690
Just before thanking you, I'd like to thank everyone who has tuned in.

556
01:00:12,690 --> 01:00:19,320
And also the many questions and do forgive me when I've not been able to come to your questions.

557
01:00:19,320 --> 01:00:24,150
We're not stopping for the holiday break, but do look at the Martin School Web site.

558
01:00:24,150 --> 01:00:29,350
We'll be announcing some more virtual conversations in the new year.

559
01:00:29,350 --> 01:00:36,830
And let me finish by thanking both Phil and Yvonne. Yvonne, you've done magnificently, given that you've had to do this completely blind.

560
01:00:36,830 --> 01:00:42,500
Whereas Philon, I can sort of see each other's movements. So I'm sorry we had this technical.

561
01:00:42,500 --> 01:00:48,120
I'm very sorry. Is beginning to be a woman of mystery. But it's been a lovely hearing, the two of you.

562
01:00:48,120 --> 01:00:52,320
Your voices. And it's been great fun. And it's a it is a very exciting time.

563
01:00:52,320 --> 01:00:57,120
Thank you very much. So thank felon Yvonne very much indeed for taking part.

564
01:00:57,120 --> 01:01:00,538
And goodbye, everyone.