Přepis řeči - ITERATIVE ESTIMATION OF STRUCTURES OF MULTIPLE RNA HOMOLOGS: TURBOFOLD

0:00:15	okay the um the next yeah are shown
0:00:19	and and you mentioned a structure
0:00:23	i one
0:00:25	you to long way
0:00:29	monica
0:00:30	i i being here
0:00:33	so i'm to the it was to just to the work and we present a joint work
0:00:36	as is prime the will of most wouldn't a of
0:00:39	how much whose lose you didn't know was don't you
0:00:42	you couldn't be here and if a used to but a written by them
0:00:48	so i don't as a very that was means
0:00:51	so i try and begin with a fairly broad introduction
0:00:53	i don't wanna apologise in advance
0:00:55	for not being able to cover all but details
0:00:57	that
0:00:58	process
0:00:59	that something which i think it's more but if
0:01:01	leave but even understanding of at least what we like to do and why it's important
0:01:05	rather than to try and
0:01:06	of the but is
0:01:07	is it it is to do with you
0:01:10	we also had a as a to P is as to be run this has just recently of your
0:01:15	you B C but my i
0:01:17	provide
0:01:19	i begin can that motivation by don't why is an important problem and talk about are only
0:01:23	non coding not an in particular
0:01:25	and i i structure dance
0:01:26	the not talk about what that's one is a
0:01:28	prediction
0:01:29	both for single and multiple sequence but that
0:01:32	a technique is a what easy but method
0:01:34	and that's with analogy with that what it comes send
0:01:36	so i present a um at a double for
0:01:39	i don't high level
0:01:40	and show how this is an it the probabilistic method for decoding of of much lower on a common
0:01:45	and a very strong analogy would that would be coding in digital communications put in the way you look at
0:01:50	the problem
0:01:51	and the tools we saw all and sort of the stage
0:01:54	all literature also this point at which we don't that
0:01:58	it's quite
0:01:59	and be in the business
0:02:01	a present experiment results in this process
0:02:03	present present how people form as compared to what it out and
0:02:06	which are from the same frame
0:02:08	and finally and but some
0:02:09	right
0:02:11	so
0:02:12	i think everybody's family would be an and the double helix
0:02:15	this is the famous discovery be what's and right
0:02:18	what good
0:02:19	but it discovered that the in a happen the in is found in this W can back to
0:02:24	we have a
0:02:25	i don't a complementary based pairs
0:02:28	a with each other a process he likes
0:02:31	and
0:02:32	on knee
0:02:33	is
0:02:34	very similar to the any
0:02:35	except that
0:02:37	hi mine
0:02:38	is replaced by you
0:02:40	or what i was as
0:02:41	do you think about audrey
0:02:43	as a leading a in one Q
0:02:44	these
0:02:45	you imaging
0:02:46	is a
0:02:47	strong a what i wants
0:02:49	is that what they can look at the mall
0:02:51	as compared to
0:02:52	these point which are are stored row
0:02:54	to to to to look at a mall
0:02:56	and a trend is exponential
0:02:59	in the free energy change
0:03:00	so these points here
0:03:02	are a much harder to break than the one that way
0:03:06	i
0:03:07	so
0:03:08	this
0:03:08	is what was the for to a me structure and usually the a are going from the five point to
0:03:13	three by man
0:03:14	so you listen this
0:03:15	i'm i'm structure
0:03:16	as a sequence of the autonomy
0:03:18	much like a dog of the in human humans you know and the sequence of you don't
0:03:21	you will have a similar by me structure of that i as a sequence of be dies
0:03:25	along the model
0:03:28	i like the in a or what happens in a don't is that it is common for this money to
0:03:31	to for the one so
0:03:33	so you typically have a
0:03:35	single molecule rather than to document model use like to put the point that
0:03:38	forming a variety of
0:03:39	struck
0:03:41	or the longest amount of time people believe
0:03:43	that
0:03:44	a on so just
0:03:45	one function
0:03:46	which will being a transient copy of the information in
0:03:49	this is a a a simple dot mine it
0:03:52	so you you know the size and the
0:03:53	new yes
0:03:54	a of it gets we
0:03:56	transcribe right
0:03:57	and do
0:03:58	a missing and your in which comes out of the side of button
0:04:02	and then it relies
0:04:03	protein synthesis
0:04:04	in that i was on this is you meant factory for producing proteins
0:04:08	and all of that and you know
0:04:11	so the in belief for that information for or you that in and fashion from the a or any on
0:04:16	routine
0:04:18	more recently what is the most how a
0:04:20	that you never realise that are a bit a very active role
0:04:23	in
0:04:24	but
0:04:26	i have realised that these
0:04:28	additional types of are nice
0:04:30	and a characterization of these on an is by what did do not do
0:04:33	not but what
0:04:34	so they do not hold for me
0:04:36	that for to was non coding are nice
0:04:38	because there are providing a function
0:04:40	without really being translated into protein so they not coding and there all
0:04:44	and that in these numbers being really discover
0:04:48	or do not but right of the C O a what a different point in time for a variety of
0:04:52	these
0:04:53	and this is showing but
0:04:54	sort of a can you please bass
0:04:56	can happen
0:04:57	one of these is we must your on it this
0:05:01	i think at less is its own splice same
0:05:04	which is cutting out of a segment
0:05:06	all the concentrate on an a
0:05:08	from this to produce actual a money
0:05:10	it operate in it and down plates
0:05:12	we use probably and the that
0:05:15	reactions actions in the cell
0:05:17	so there's a idea of these functions that are nice
0:05:21	in these rules
0:05:22	a like the role that you're familiar with and protein instances
0:05:26	with a what on coding is the one is is important
0:05:29	in these second roles for a any
0:05:31	the it is the structure was determines
0:05:33	function
0:05:35	and that's almost all in an G for almost all molecules
0:05:38	including proteins
0:05:39	and that's the reason why this problem would be
0:05:42	is a grand challenge problem and science to be
0:05:46	based from the relation of structure
0:05:48	it's quite challenging
0:05:50	it it a well X like
0:05:51	it's a about three
0:05:52	and this is difficult to do because to purify a sample and so the slice that
0:05:56	and then there questions finally with of those conditions actually represent present physiological conditions
0:06:00	the body
0:06:01	in which the what it actually opera
0:06:04	so what will be interested in this computational estimation of
0:06:07	are and it's can be structure
0:06:09	and if you can do this
0:06:11	the kind of things you can on
0:06:12	on so that's what is this function of these non putting on
0:06:15	because once you know structure
0:06:17	you have a we've proving what should be function
0:06:20	you also have a ease of understanding this you know
0:06:22	i do you know so it is a whole sequence based
0:06:25	what one on is
0:06:27	for the reasons that a lot lately to the structure is what
0:06:30	the function
0:06:31	so that we need to work so that you will have more eight sequences were from the same
0:06:35	have the same structure of on the same function
0:06:37	so you would like to be able to figure out which of these are in different out
0:06:41	it's rather than comparing
0:06:43	based on sequence a sequence you'd like to compared based
0:06:46	structure
0:06:47	finally a as the standing close you got like to users
0:06:50	a a it's right i
0:06:51	that's the the quality
0:06:53	of such a prediction
0:06:54	to be able to
0:06:55	some sized it's
0:06:56	rather than
0:06:57	just test
0:06:58	a right but if you know
0:07:02	one thing which was
0:07:03	spherical in at feast for i and six are any collection structure prediction as compared to prove
0:07:09	and that are any have a
0:07:11	our our mention to you that this primary structure consists
0:07:14	or linear chain monte you which is laid out the we have from the pipeline to by and and this
0:07:18	just rolled away way of for but was a fitting the space
0:07:21	is it just the in monte Q
0:07:23	this
0:07:23	for one itself so the formation of these complementary be spare
0:07:28	this is i'm not as to that of mind that in the time i
0:07:31	we N betting with side side
0:07:33	in the N a
0:07:35	a that in this case time as the base that you're so so you have a you
0:07:38	in addition you also have a
0:07:40	do you pairs
0:07:41	and R
0:07:43	is also possible
0:07:45	so this same one you you see from here by frame to prime
0:07:48	is laid out over here
0:07:49	and i don't
0:07:50	i
0:07:51	or for me and the sign to see what you can see that it's going around round i'm coming well
0:07:55	the be an on coming back
0:07:58	this is referred to as the primary structure which is the sequence
0:08:01	and this is done but it is can make a mean this is what high throughput sequencing does
0:08:06	what you interested then is predicting the second we structure once to predicted this the dot or take those three
0:08:10	structure
0:08:11	and you have structure becomes
0:08:13	easier because you already know the interactions that the
0:08:17	and and i think is this
0:08:18	progression of interactions
0:08:20	is that's simply strong
0:08:21	already mentioned this is very strong ones here a one i one
0:08:25	and over here that wants all much speaker
0:08:28	a the trash that this little are given you because
0:08:31	so there's progression of prediction
0:08:34	more prediction provision of formation of structure
0:08:37	also guys the mechanisms by but you pretty
0:08:39	so our goal in this work will be the prediction of segments
0:08:45	this is referred to only as fourteen of a not any you
0:08:49	and
0:08:49	that much greater variety of structure than the in and up
0:08:53	is it an example of an non you
0:08:55	this is
0:08:56	uh are are is P
0:08:58	and you will see that what you have
0:08:59	all these various more piece
0:09:01	which are made up of you D Cs
0:09:03	and loops
0:09:04	do the two
0:09:05	types
0:09:06	to is as bad or just to describe them
0:09:09	so this you can not strong flat as a ladder or would here actually is the he likes
0:09:13	and the way that a wasn't you'd
0:09:15	structure
0:09:16	and then you have these these two
0:09:18	and i was applied for data
0:09:20	what set of base pairings a lotta
0:09:22	given the C
0:09:26	now
0:09:27	like
0:09:28	you stop with just was and dynamics
0:09:30	will be used only in the
0:09:32	dominant
0:09:34	in the room
0:09:35	you can have a variety of different structures
0:09:38	the property of a given structure
0:09:40	or not do
0:09:41	this one quantity
0:09:42	the was meant constant
0:09:44	which is
0:09:45	but actually using
0:09:47	with the free energy change
0:09:48	you have a i have free energy
0:09:50	but have just like a structure energy that that you want
0:09:54	so the most likely structure comes the one which minimize is free energy
0:09:58	and accordingly techniques for prediction of secondary structure
0:10:01	a by
0:10:02	coming up with models
0:10:03	which predict this free energy structure
0:10:05	what of the most efficient models tends to be one which is called the nearest neighbor model
0:10:09	it looks at
0:10:11	a based betting interactions sense to the one nearest neighbouring base pair
0:10:15	and has a we have become the true free energy change
0:10:18	in terms of this
0:10:19	based pairing
0:10:20	right
0:10:22	as work also done
0:10:23	i didn't to protest just to a and not in as lab
0:10:26	who was a chemistry
0:10:27	not for speeding and a work but this
0:10:29	a model which is down use but i mean i
0:10:33	so one can imagine various algorithms buttons now for predicting second structure
0:10:38	by trying to minimize free energy
0:10:39	and that's something which is been done prior to a well
0:10:43	people and understand we programming
0:10:45	the
0:10:46	what want what make a were here is this method does a very much that some to do to be
0:10:50	a
0:10:52	that or what is the minimum free energy structure
0:10:54	a set of possible these the dynamic program you do is an of the yeah
0:10:58	oh for those of you who were go estimation of the coding you also know that that
0:11:03	the or with the south and
0:11:05	the P C G are a button
0:11:07	does this in a soft sense
0:11:09	and it is also the an it was in in the setting which is referred to and the chemistry the
0:11:13	noise he of the partition function
0:11:15	which because of the property
0:11:17	uh
0:11:18	a base their location i would be at a location in based G
0:11:21	and have a lot of what about how this done but
0:11:24	also compute a by using a dynamic program
0:11:29	so that a a a a a a new techniques one is a hard decision would you can note like
0:11:32	to do what is this
0:11:33	think the structure which minimize free energy
0:11:36	are there is a prediction of based pairing properties
0:11:39	in the C
0:11:42	i don't sell
0:11:44	what is the connection the
0:11:48	double for the
0:11:49	yeah but i
0:11:50	a the same day
0:11:52	and then try for white
0:11:53	a joint decoding
0:11:55	you got to joint decoding exactly because
0:11:57	a computationally expensive
0:11:59	so you do
0:12:00	approximate joint decoding by using it usually decoding and probably information from one sequence to that
0:12:06	well
0:12:07	in major don't out
0:12:08	the same structure
0:12:11	which probably a the same function
0:12:13	are encoded as that
0:12:14	sequence
0:12:15	and
0:12:16	that's the connection
0:12:17	do would be
0:12:18	so over here we showing of what is the R any
0:12:22	across different organisms
0:12:24	and that of and and this to do with here
0:12:26	and what you would see is that the structure is the same and when i C C am i mean
0:12:30	of and a logical sense rather than a very exact sense that's
0:12:34	some sir
0:12:34	a some that about
0:12:35	to use
0:12:36	but if you look at these closely you will see that there are
0:12:40	bases which are modified for instance this do you see where is change to and you below
0:12:45	it becomes obvious that as not as you make you patients
0:12:48	in a compensating fashion
0:12:49	each time you change you do any a you jane the corresponding C do you
0:12:53	you can still made in the beast bidding interaction and maintain the integrity of a second be structure
0:12:58	so the structure and still be form a log in as a most to be able
0:13:02	and will therefore be seen in the H
0:13:04	so that multiple encodings
0:13:06	all the same structures are provided to us
0:13:09	by nature or true it's process of
0:13:11	compensating nations
0:13:12	and our goal is to try and predict signal structures by using this model of a lost to get a
0:13:18	that that in as an would be decoding
0:13:20	you want to use them collectively to decode
0:13:23	and
0:13:23	you can now look at a similarity to in them you can see the same you see the responding regions
0:13:28	and
0:13:29	also in addition you have the information from the alignment of these two C
0:13:33	a a of what alignment
0:13:35	but you also have
0:13:36	information about alignment from this
0:13:38	and so the goal of all are structure prediction structure and alignment
0:13:42	i to come up with a production of these structures
0:13:45	and well as a conforming a line
0:13:48	an obvious this constraints from one which impose constraints on what we can do without
0:13:54	so this is
0:13:54	in some sense the frame but our goal is to take a number of input sequences
0:13:58	the model C construct a prediction but that was pretty
0:14:01	structures of these
0:14:02	and also up
0:14:03	or what's an optional and
0:14:07	that's so that you phone like this all so that a programming out there is a mapping by just cycle
0:14:12	this
0:14:13	and again the similarity but that would coding
0:14:16	is very
0:14:17	telling
0:14:17	this is exponential and topics complexity
0:14:19	and the number of sequences
0:14:21	you can the joint
0:14:22	according
0:14:23	two sequences and double
0:14:25	you can give indications and the decoding
0:14:27	the complex exponential and the interleaving that
0:14:30	so this is something which is not feasible
0:14:32	you and for two sequences
0:14:34	this is something you cannot do without think one
0:14:38	so
0:14:38	our goal is to try and come up with a proper stick take me
0:14:42	does this
0:14:43	by iteratively computing
0:14:46	single sequence for like properties
0:14:48	and updating these as to go from iteration preparation
0:14:51	and much the same be as the decoding
0:14:54	for
0:14:54	i
0:14:57	so do talk about this in detail level
0:14:59	present this
0:15:00	but in this
0:15:01	to to form
0:15:03	so the way you can do but this and to real form
0:15:05	is that you have these two sequences
0:15:08	which have this structure but the just can be shown and this
0:15:11	lower triangular matrix a with here
0:15:13	but issuing showing what are the peace betting interaction
0:15:15	so this space
0:15:16	at this location
0:15:18	a with the base at this location
0:15:20	in the screen in this very that we're here
0:15:22	and so on so these
0:15:23	he is of or at least traces of lines as is you here
0:15:27	corresponding to seconds to you have a corresponding
0:15:31	set of
0:15:32	based pair shown over here
0:15:34	and then there is the alignment green the do which is between the two sequence
0:15:38	oh the they just try to predict
0:15:40	the best possible second structures
0:15:42	can be a a a a dynamic program to and figure out
0:15:45	what is the bar
0:15:47	for
0:15:49	alignment
0:15:49	and what are the bearing interactions that way
0:15:51	maximise
0:15:52	you free energy chi
0:15:57	in order to do this and the double frame
0:15:59	we cannot live with hard decisions
0:16:01	so the first thing that you do
0:16:03	to to actually to present this
0:16:04	in
0:16:05	a soft frame but with information is problem
0:16:07	so the base pairing properties become
0:16:10	properties of base pairing
0:16:11	the problem and i and properties become properties of alignment
0:16:15	and then if you sequences as the figure you see
0:16:21	at at this point
0:16:22	you realise that
0:16:24	if there is a very likely
0:16:26	like a base pair in the sequence
0:16:27	and it's highly likely
0:16:29	that
0:16:30	the fight i'm and of that base spare as a line with a given by prime and
0:16:33	a second sequence and G prime and all that base pairs aligned with the T by of the second sequence
0:16:38	it's providing you information
0:16:40	about what of the second sequence
0:16:42	and this is the information that you get a a bit of an alice
0:16:47	so
0:16:47	we can easily see and four
0:16:50	a to your properties
0:16:51	or base pairing
0:16:52	for a second sequence
0:16:54	by using the information
0:16:55	or base baiting one sequence along with the alignment problem
0:16:59	and because everything is a probabilistic
0:17:01	all the information or and saw
0:17:03	and this is something we can now incorporate
0:17:05	in the voting of the sec
0:17:08	you have to a sequence
0:17:09	these sequences
0:17:10	the process is not much different
0:17:12	you can use
0:17:13	the do this information to two sequences
0:17:16	and
0:17:16	in for a what are the properties of base pairing
0:17:19	for the third sequence
0:17:21	same way and you can be this and that is a weighting scheme that we come up with which
0:17:27	so here is essentially a that
0:17:29	a scheme works
0:17:29	if you're trying to predict what is the
0:17:32	extrinsic information what of the information provided you for you for for of a given sequence X M
0:17:37	but other sequences
0:17:39	use use information from all the other sequences the corresponding alignment property matrices
0:17:43	in for these
0:17:44	we them an appropriate uh
0:17:46	combine them to come up with a
0:17:48	extending thing information for paying of a given C
0:17:50	this and the information can be incorporated it a frame but in much the same way at as done that
0:17:55	would coding
0:17:56	it has an interpretation
0:17:57	as
0:17:58	to the posterior property
0:18:00	in the stuff that you have a
0:18:02	in the what decoding coding also
0:18:03	i well it a lot of to drop the you of see and or or or a good also has
0:18:07	the structure of to to the updating the base being property
0:18:13	oh is the summary of this
0:18:15	and then want to a it this you can find pretty how would the high prediction
0:18:18	how
0:18:21	oh present this
0:18:22	before for i make that presentation of what was point out the computational complexity
0:18:26	is also
0:18:27	similar to would decoding
0:18:29	we get the complex be compared to single sequence folding
0:18:32	while i to get the benefits
0:18:34	a a joint sequence training
0:18:35	the joint sequence for would be exponential in the number of sequences this is
0:18:39	you do the for two D and to the part okay
0:18:42	a whereas as a complexity is you square
0:18:46	so we can uh well it is
0:18:48	a look at how these performed
0:18:50	and
0:18:51	i will give you the results quickly
0:18:53	so
0:18:53	we are it this or a benchmark dataset but not structures
0:18:58	and we can value these by looking at a sense be was as P P V
0:19:01	since to really is the number of actual that he predicted directly
0:19:05	P B we had the number of predictions that i
0:19:07	so that
0:19:08	a standard are a of you are what is to be
0:19:10	in the upper right corner of what they're
0:19:13	and here is
0:19:14	double for
0:19:15	for three sequence double for for ten sequences
0:19:18	and
0:19:19	or need techniques this is log on any which is a technique would just probably stick information
0:19:24	on a highly four
0:19:26	and single sequence for
0:19:28	so the message here is that by using this information and initiatives
0:19:32	fashion
0:19:33	you get do significantly better and what is these that
0:19:36	time i'm also be disk
0:19:37	if a better than these on an L for is much faster but then you can always give the wrong
0:19:42	on so
0:19:42	but
0:19:45	got load
0:19:45	at present a double for a multi sequence
0:19:48	structure prediction the
0:19:49	which has strong and is with that would be coding and is motivated by this and as you hear
0:19:54	and provide that is the close to or high that everything is forming
0:19:58	well having like city
0:19:59	similar to sing
0:20:02	i this collection
0:20:03	coding T
0:20:04	and for the
0:20:05	to
0:20:31	yeah
0:20:38	so i given weeks is the one who was on shape based techniques
0:20:43	we you collaborating with him trying to see how we can incorporate shape
0:20:46	i that in addition to the data that incorporating
0:20:49	there's a very strong analogy in the way the she that can also be an now incorporated
0:20:53	it's also was to get property which you can be
0:20:57	in do the forming of sequence
0:20:59	traditionally that has been a a sequence you can single sequence for
0:21:02	you working on trying to see how you and are in that the much as C
0:21:07	that's right
0:21:08	i will
0:21:10	and more recently also like to see how we can apply this to a I V N S I B

ITERATIVE ESTIMATION OF STRUCTURES OF MULTIPLE RNA HOMOLOGS: TURBOFOLD

Biosignal Estimation and Classification

Přednášející: Gaurav Sharma, Autoři: Gaurav Sharma, Arif Harmanci, University of Rochester, United States; David Mathews, University of Rochester Medical Center, United States