Speech Transcript - BIOLOGICAL PATHWAY INFERENCE USING MANIFOLD EMBEDDING

0:00:13	uh thank you uh first of all of the uh organized
0:00:16	five
0:00:18	speak
0:00:18	uh on uh
0:00:20	uh biological pathway inference
0:00:23	and uh
0:00:24	what i'm going to do is i'm going to describe
0:00:27	a approach
0:00:29	to this problem
0:00:30	of
0:00:31	combining
0:00:33	in this case gene expression data so this
0:00:35	continuously value
0:00:37	uh
0:00:38	abundance
0:00:39	of uh
0:00:40	a messenger or are a as measured uh on a V microarray array chip
0:00:46	by now with ontological
0:00:48	data of which describes
0:00:51	something about
0:00:52	gene function
0:00:53	in particular
0:00:55	uh the products
0:00:56	that the
0:00:57	are generated uh or induced
0:01:00	by gene expression
0:01:02	okay so
0:01:03	the standard
0:01:04	approach is to uh i one for matt at uh uh
0:01:08	yeah i data analysis
0:01:10	is
0:01:10	to be data rip
0:01:12	ignore
0:01:13	any kind of functional
0:01:15	uh or biological system
0:01:17	a type of uh of of priors
0:01:20	and then after to you
0:01:21	john your analysis
0:01:22	holdout out particular
0:01:24	i i i a gene uh a uh uh factors let's say from the data
0:01:28	then you go and try of uh you know to validate validated or to make some inferences on
0:01:32	what's actually going on what of the functional relationships between these genes and can you
0:01:36	somehow in court the pork core icsi's make calm incorporate them into a functional path
0:01:42	so we do the simultaneous
0:01:44	so here we're going to
0:01:45	simultaneously do
0:01:47	uh uh clustering variable selection
0:01:50	and the
0:01:52	and then
0:01:52	uh
0:01:53	functional
0:01:54	i don't shen
0:01:56	so
0:01:57	uh i think everybody here has a least some
0:02:00	vague notion at a minimum
0:02:02	of uh the fact
0:02:03	the gene is a segment
0:02:05	of the N A
0:02:06	uh that uh codes for protein
0:02:09	and uh
0:02:10	uh course
0:02:11	uh not all of the uh
0:02:13	uh
0:02:14	all the go nucleotides uh
0:02:16	on the D a on the D N A um as trans
0:02:19	code for proteins but the genes
0:02:21	uh in particular
0:02:22	are are the ones that uh while just understand and can describe some sort of function to
0:02:27	uh
0:02:28	they
0:02:29	these functions which are
0:02:31	a primarily production of proteins true the pro is all
0:02:34	and the poor uh be process of a translation
0:02:37	or organised into what by all just call pathway so pathways or sequences
0:02:42	of
0:02:43	uh
0:02:44	activation
0:02:46	of different
0:02:46	genes or protein products
0:02:49	that need
0:02:50	from one state to another
0:02:52	right so
0:02:53	uh you know a pathway
0:02:55	uh four
0:02:56	the inflammatory response
0:02:58	leads
0:02:59	uh starts with
0:03:00	a a uh uh uh some uh
0:03:03	infectious agent or some
0:03:04	some in salt to the uh and you system and ends up with production of sight of kinds that uh
0:03:10	basically a induce sinful information there's a very complicated sequence of gene expression
0:03:16	that uh is associated with that process
0:03:19	so um one of the principal problems is the discovery of
0:03:23	how these pathways become perturbed were deregulated
0:03:27	uh under uh for example disease states
0:03:30	and uh
0:03:32	uh the
0:03:33	principal
0:03:34	uh fact uh of uh the matter is
0:03:36	that
0:03:37	these functions are not just expressed by a particular gene expression
0:03:41	uh uh uh uh uh a factor but they're expressed over time
0:03:45	and over space
0:03:47	and
0:03:47	that's what we're going to talk about in particular
0:03:50	in the context of
0:03:51	the uh
0:03:52	uh
0:03:54	so your response to infection "'em" addition shows some data later around for like be so techniques as fusion
0:03:59	of of
0:04:00	uh expression data
0:04:02	and uh ontological
0:04:03	a gene not gene ontology data
0:04:06	so
0:04:06	uh this just shows a a uh i i a typical picture that you'll find in this pretty impressed at
0:04:12	a particular case and nature of use in two thousand five
0:04:15	which describes the uh you know molecular biologist understanding
0:04:19	of how it's cell responds to infection faction in terms of protein production productions all of the user or protein
0:04:25	uh in the uh
0:04:26	uh the the the nucleus you have
0:04:29	the uh uh a process of a
0:04:31	the an a transcription and replication
0:04:33	and that generate proteins that the
0:04:35	uh that are located at particular regions within the cell so close to binding sites
0:04:41	receptor sites
0:04:42	production of uh
0:04:43	uh of sight of combines in fear on so for
0:04:46	a so there is a very complicated
0:04:49	uh diagram here
0:04:51	had ways
0:04:52	can be i characterised as these sequences of
0:04:55	events that leads for example to sell that program cell that
0:04:59	i thought that was this
0:05:00	uh
0:05:00	which are is is an immune response
0:05:04	so the the point though
0:05:05	uh is that we want to somehow
0:05:08	uh compress all of this complicated and and relatively vague information this picture
0:05:13	to some kind of
0:05:14	topological
0:05:16	uh a constraint
0:05:18	on a how
0:05:19	uh say
0:05:20	two genes
0:05:21	can be related or not
0:05:23	so this shows
0:05:24	uh what's called a gene ontology semantic graph and this captures
0:05:29	function
0:05:30	of of of different genes
0:05:32	in particular captures
0:05:34	one of three
0:05:35	a gene ontological uh uh uh classifications
0:05:39	uh which is
0:05:40	the uh cellular location
0:05:42	of the protein that's produced by a particular G
0:05:45	so you have here for example
0:05:47	uh in the membrane us sell your membrane versus and uh
0:05:51	in the a protein complex
0:05:53	oh or and the nucleus
0:05:54	uh down here you'll have different genes that are associated
0:05:58	with
0:05:59	this particular location
0:06:01	in terms of proteins that they produce the larger the circle
0:06:04	the more genes are in that particular uh uh
0:06:07	functional
0:06:08	uh body for this particular uh
0:06:12	uh a process this particular pathway which is the of one
0:06:17	"'kay" so
0:06:18	this is the diagram
0:06:19	that basically were gonna you is to merge
0:06:22	with the the the expression data
0:06:25	the raw expression data this is this comes from literature
0:06:29	gene ontology is a
0:06:30	a database which collects
0:06:32	uh from different data
0:06:34	uh a a a database is uh
0:06:36	that uh represent experimental uh and validate results on
0:06:41	uh in this case location
0:06:43	a cellular location of protein production
0:06:45	we're gonna we're gonna take this semantic description
0:06:48	oh
0:06:49	uh relations between
0:06:51	between genes that are there attached to a particular component
0:06:54	and we're gonna use that to sort of precondition the clustering
0:06:57	oh the gene expression
0:06:59	that's and not nutshell
0:07:00	oh what we're doing
0:07:02	so this just shows
0:07:03	that that's the this shows the how it's sort of a
0:07:06	uh put together you know more graphic uh
0:07:09	uh a context
0:07:10	we have a gene microarray array here
0:07:12	with uh uh a a genes are expressed say over different treatments in class one which might be help the
0:07:19	in the class two or the trip or you all is uh
0:07:21	a subject
0:07:22	uh and uh
0:07:23	uh and then uh these would be different genes along the rows
0:07:27	and we take uh uh these ontological
0:07:30	uh uh
0:07:32	speakers they shorten the previous slide
0:07:34	so this might be the nucleus
0:07:36	uh this one might be the side of plasma of this might be the for use all
0:07:40	and that then
0:07:41	i gives a a like a
0:07:43	prior on how closely related these genes are in terms of uh function
0:07:51	right so
0:07:52	going from
0:07:53	clusters the functional pathways is is
0:07:56	a very difficult problem
0:07:57	and uh
0:07:58	uh the the problem is that genes with similar to have
0:08:01	russian
0:08:02	uh do not necessarily have similar function
0:08:05	right so we if use correlation
0:08:08	the correlate late
0:08:08	gene expression from two different genes and say that they but they the same function just simply because
0:08:13	they seem to have the same shape
0:08:15	over their temporal
0:08:17	uh uh expression profile
0:08:19	uh that
0:08:20	maybe completely spur it's they may not have share function
0:08:23	how do you incorporate function in
0:08:26	the uh as additional information
0:08:28	you use
0:08:29	G on top
0:08:32	so uh in order to uh to capture this ontological
0:08:36	uh uh uh function uh relationship between two genes
0:08:40	uh we're gonna use basically a a a a a a manifold learning uh in a uh a a a
0:08:45	type of a uh approach
0:08:46	a lost uh eigen maps approach which is going to basically bed
0:08:51	the genes into a lower dimensional map a manifold
0:08:55	where
0:08:55	distance is
0:08:57	in that manifold are gonna be directly proportional
0:09:00	two
0:09:01	the ontological similarity between those two G so of the genes live in
0:09:05	one of the common
0:09:07	a locations within the cell in terms of the protein production
0:09:10	uh then uh the similarity W Y G between the two genes i J
0:09:15	uh will be more
0:09:17	right and uh
0:09:18	and that will be used as a weighting in this uh a plus in eigen maps
0:09:22	a a clustering procedure which will give us a lower dimensional uh
0:09:26	uh
0:09:28	the plus raffle plus N
0:09:30	induced
0:09:31	in of the date
0:09:33	okay
0:09:33	no where does the gene expression come
0:09:36	it comes in in a very weak sense in this uh in
0:09:40	this embedding you but look at it is being driven by the ontology
0:09:43	is been driven by function
0:09:45	so similar functions in this case being
0:09:47	similar locations within the cell
0:09:49	that these genes jeans and
0:09:52	uh
0:09:53	uh that's R W Y G K
0:09:55	but the gene um
0:09:57	expression
0:09:58	uh controls the neighbour
0:10:01	uh so we're going to
0:10:03	uh basically a zero wait
0:10:05	if the expression profiles are to dissimilar
0:10:08	"'kay" so it's it's a way of
0:10:10	of conditioning
0:10:11	uh the um
0:10:13	uh the embedding which would just be based on pure
0:10:16	uh on ontology
0:10:18	uh
0:10:18	uh based on the uh
0:10:20	a similarity of that of the gene expression profile
0:10:23	i'm not gonna go through a this look life the eigen mass spell can and i U V
0:10:26	in two thousand to two publish very nice paper on and sites five is the say that are gonna fine
0:10:31	yeah clear that's why i and Y J and some lower dimensional space
0:10:36	uh maybe two dimensions the visualise
0:10:38	uh such that
0:10:39	you basically preserve distance
0:10:42	in the ontological space
0:10:43	now of force
0:10:45	uh well we prune neighbours
0:10:47	if there expression profiles are to the simple
0:10:51	okay
0:10:52	so
0:10:53	this uh embedding i has been applied to uh the uh a particular dataset that's out there um called the
0:11:00	young data set
0:11:01	uh
0:11:02	by um
0:11:03	uh
0:11:04	looking at the
0:11:06	in this case
0:11:07	uh in each vitro
0:11:08	uh to kill and uh tuberculosis uh
0:11:11	uh
0:11:12	infection
0:11:13	of mac fe just cells
0:11:15	uh that uh
0:11:17	i then are pass say using mike raise an a and R T P C R
0:11:21	to produce that gene map that he map actually four
0:11:25	uh and a there are eight time points they wanna basically look at
0:11:29	how the uh this the these uh a dendritic a mac the phase cells
0:11:34	respond
0:11:35	uh after that trip kill when uh uh
0:11:38	uh has been introduced
0:11:40	and so uh here's the data for the control group
0:11:44	here's the data for the tuberculosis uh a group again over time
0:11:48	and so we're gonna be trying to do was determined changes
0:11:51	uh uh uh a a that and and associate those changes from control to the uh
0:11:56	but a brick you and uh group
0:11:58	uh with
0:11:58	functional uh uh
0:12:00	uh
0:12:01	protein of production
0:12:04	okay so
0:12:05	we're gonna take a first of all
0:12:07	uh a difference between the control and tuberculosis so that we can have a baseline which is
0:12:12	control
0:12:13	and then we're going to and bed
0:12:16	those expression profiles as deferential expression profiles
0:12:20	into a two dimensional space using this what plus the eigen map
0:12:24	um so uh
0:12:26	i'm first gonna show use all uh the standard
0:12:29	functional pca and batting
0:12:31	which uh simply applies
0:12:34	uh a
0:12:35	singular value decomposition
0:12:38	on
0:12:39	a a a a a a spline basis
0:12:41	interpolation over time of those uh the an expression profiles i should be the previous slide
0:12:46	and uh
0:12:48	uh then afterwards it applies a a a gaussian mixture model clustering
0:12:52	to try an associate these different uh
0:12:55	uh uh these different genes
0:12:57	uh as uh hopefully a associated with different pathways or different the
0:13:02	uh a function
0:13:04	so uh what you see is the sound of classic
0:13:07	this classic uh a a uh
0:13:09	uh
0:13:09	concentration
0:13:11	of uh
0:13:12	uh of
0:13:13	of measure here on
0:13:15	on B
0:13:16	uh where where the the mit the
0:13:18	a gaussian mixture models
0:13:20	have basically the trying to match with a two dimensional
0:13:23	a domain they're trying to simultaneously capture clusters
0:13:27	which might
0:13:28	actually be two dimensions
0:13:30	but uh there are also clusters are probably just one dimension so it's a very
0:13:34	uh a a a a a a a very heterogeneous speech
0:13:38	if you go and look at after you clustered
0:13:40	you would expect these plus to do very good right "'cause" this clusters uh obviously
0:13:44	the centroid doesn't even near
0:13:46	uh
0:13:48	i near the any any particular uh a a G
0:13:51	uh
0:13:51	uh
0:13:52	the uh
0:13:53	uh that that we can classify how good this clustering is simply by looking at the percentages
0:13:59	oh of uh
0:14:00	uh a a of each one of these genes in a given cluster that in this in the same location
0:14:04	in the set
0:14:05	and
0:14:06	is over fifty percent
0:14:08	uh do not uh are not call like it locate over fifty percent of all the genes in any cluster
0:14:13	are not co-located co located the cell which indicates that's again
0:14:16	that uh gene expression over time
0:14:19	other profiles do not
0:14:21	i discriminate accurately between uh a genes that with different function
0:14:26	oh yeah hand if you use this manifold colour betting that i described
0:14:30	you get a much nicer
0:14:31	us spread
0:14:32	oh of the uh uh a of these clusters
0:14:35	into uh well defined groups
0:14:38	these are for different clusters
0:14:40	uh the uh you drop but to a much
0:14:43	much lower
0:14:44	percentage of uh
0:14:46	a of uh impurity right some money more of the genes within the cluster groups label green blue and
0:14:52	uh turquoise and so forth
0:14:54	a a i close to each other within the cell
0:14:57	which is which is the sign of course
0:14:59	uh that uh
0:15:01	the the method that we implemented
0:15:03	is actually capturing this co location ontology
0:15:06	um
0:15:07	and does so would this just i was some by clustering and in the C is i'm not gonna bother
0:15:11	dwelling on that i'm running out of time
0:15:14	uh but in the indicating that the we have improved performance not surprisingly
0:15:18	because we're using ontological date gene ontology data
0:15:22	to condition
0:15:24	uh the uh
0:15:25	these cluster
0:15:26	if you don't if you just use a functional pca
0:15:29	uh you get uh a a cluster indices
0:15:32	a for these various pathways these uh that uh
0:15:35	or
0:15:36	uh i have much lower or a quality and if you use our method that uh which you can see
0:15:41	just pairing these numbers
0:15:43	this is the this is the quality index between zero one one
0:15:46	that uh again indicate the purity of each shot
0:15:49	uh of each one these clusters
0:15:51	in terms of the number of that
0:15:53	a genes that i within
0:15:55	the same class
0:15:58	okay so in conclusion
0:15:59	i describe this this method
0:16:01	uh which deals with calm embedding
0:16:04	of both
0:16:05	uh expression data and functional
0:16:07	uh and
0:16:09	uh of genes
0:16:10	uh uh in terms of their work uh expression under in this particular case that tuberculosis uh
0:16:16	uh infection
0:16:17	uh i i've uh uh
0:16:19	basically uh describe how we do this using a plus an eigen map
0:16:23	uh it allows us
0:16:24	to uh
0:16:26	uh to to if you like couple
0:16:28	the
0:16:29	uh uh the variable selection
0:16:31	clustering
0:16:33	and
0:16:34	functional annotation in one package
0:16:36	and as a result
0:16:37	we can improve uh
0:16:39	uh
0:16:40	pathway way analysis
0:16:42	uh by by doing
0:16:44	that's a say
0:16:57	on what you
0:16:58	elaborate a little bit on the test in
0:17:01	a pitch T
0:17:02	do you terms
0:17:03	yeah on case is use that you now using the hierarchical structure at that you it aims
0:17:09	we also yeah so those distance
0:17:13	uh we are to go back to
0:17:17	this
0:17:18	equation so
0:17:19	yeah i i skipped over this
0:17:22	uh partially because there's a type
0:17:24	but
0:17:24	S
0:17:26	a G should
0:17:28	uh but uh uh so this distance is defined in terms of
0:17:32	the
0:17:33	number of
0:17:34	go terms
0:17:36	which are common to the two gene
0:17:40	and the number go terms that
0:17:42	uh you know a are margin
0:17:44	so that what that's actually doing if you look at the
0:17:47	uh this graph here
0:17:49	is it saying if i if i have to genes
0:17:52	and i look at where they'll a ice let's say the you have to genes that lie within
0:17:57	uh you know this uh this particular we're alan membrane co location
0:18:01	well then the car you go back and look at the pair
0:18:04	as the parent
0:18:06	uh statistics
0:18:07	that tell you they give you this topological mapping
0:18:10	now granted
0:18:12	it's only the parents we don't look up the grandparents and great grandparents and so forth
0:18:17	uh but we are taking account of the pollen G and that at least first sense
0:18:21	yeah
0:18:22	sure
0:18:24	a question
0:18:33	thank you what can you do is you have only a partial knowledge of the ontology
0:18:37	which is a
0:18:38	actually what we have because
0:18:41	you know one can't believe
0:18:44	the that all published results and so
0:18:46	uh
0:18:48	we don't have when we but we we would like to have
0:18:51	uh measures the cough
0:18:53	uh
0:18:54	in terms of you know
0:18:55	the degree to which the ontology can be relied upon
0:18:59	that just doesn't exist yet
0:19:01	right
0:19:02	uh
0:19:03	i think that uh
0:19:05	it's one of the main deficiencies
0:19:08	of
0:19:09	the functional annotations
0:19:11	we have today that there is no
0:19:13	figure of merit figure of confidence finance that one can you
0:19:17	that allows you to
0:19:18	you know you know systematic way know what kind of waiting you need to apply to that on ontological information
0:19:25	in order to balance
0:19:28	the uncertainty of ontology versus the uncertainty
0:19:31	uh of gene expression right
0:19:34	so
0:19:35	the answer is not a satisfactory one and fortunately i can't
0:19:38	can give you
0:19:39	uh
0:19:40	and not be answer there
0:19:47	if it to a question that its um you have
0:19:50	the thing impulse impossible to use or disease use
0:19:53	the
0:19:54	gene expression location be change
0:19:56	yes
0:19:57	yes absolutely
0:19:59	in then how do you i mean which means that you you know way you you gone the by use
0:20:03	or completely a
0:20:05	you locate
0:20:07	a gene
0:20:08	yeah them the ontology which is maybe not the one corresponding to the disease
0:20:13	so you
0:20:13	yeah that that's an excellent question and and it's related
0:20:16	of course uh
0:20:18	it to the fact that the ontology really should be a temporal
0:20:22	database
0:20:23	it's not
0:20:23	right a it collapses
0:20:25	the entire time course
0:20:27	of you know functional activation
0:20:29	into a summer
0:20:31	the only the only way that we account for that
0:20:34	is by the fact that
0:20:36	the ontological uh i notation each one of these genes
0:20:41	uh is not unique it's not just a one dimensional quantity
0:20:45	so
0:20:45	a gene may be long simultaneously
0:20:49	two
0:20:49	several locations to in of it if the protein production
0:20:53	uh that it
0:20:54	that it's responsible for
0:20:56	in a two different phases
0:20:58	uh is say in the nucleus under one phase and in uh you know that
0:21:05	over over the
0:21:06	the the the membrane and in another fit
0:21:09	but again the ontology is not rich enough yet to be able to capture that temporal information
0:21:14	if it was
0:21:16	uh we could obviously do much better and we could really start talking about pathways which are temporally modulated
0:21:22	excellent question
0:21:30	basically
0:21:31	you know with slide
0:21:33	saw
0:21:33	for or change if the data
0:21:35	or
0:21:36	right
0:21:37	yeah just one at this time approach change to the
0:21:40	yeah station
0:21:42	right
0:21:43	right i like to agree with you there that we've done that but we have
0:21:46	i
0:21:47	uh the because it we have we could do that if we computed distance
0:21:52	sort of short time distance over a window
0:21:55	between gene expression
0:21:56	but we actually print
0:21:58	compute the distance over the entire temporal
0:22:00	uh
0:22:02	period that's collected
0:22:03	so time again is collapsed
0:22:06	but if we truly had ontological data that was
0:22:09	temporally
0:22:10	uh specific
0:22:12	we could develop a much more sophisticated model
0:22:15	B
0:22:16	frankly frankly more used
0:22:18	but this is a is the beginning
0:22:20	right exactly
0:22:22	so yeah that's that uh al again
0:22:25	i

BIOLOGICAL PATHWAY INFERENCE USING MANIFOLD EMBEDDING

Systems Biology

Presented by: Alfred Hero, Author(s): Arvind Rao, Carnegie Mellon University, United States; Alfred O. Hero III, University of Michigan Ann Arbor, United States