Player is loading...
5 minutes questions: Interpretable artificial intelligence for cancer personalized medicine
Dr. Maria Rodriguez Martinez, Group Leader, IBM
Wednesday, Jan. 25, 2023 · 10:10 a.m. · 09m 51s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
but ah yeah i i
00:00:07
uh_huh points one of condition one question on that but no no no no system where you pretty your c. v. c.
00:00:14
well but but what you could have you been using full training system to see only small molecule
00:00:19
pinpoint automated tools any soul who does w. productivity
00:00:23
depending on the type of the body to you using
00:00:26
so for the that that said that we use we use a lattice clean
00:00:29
is i would have waited to see this yeah and they see a lady
00:00:33
so most of then and this will that doc screen in they use if they
00:00:37
approve molecules so today that small compounds i don't think that that ties in our database
00:00:43
it's obvious it's more compounds to thanks uh_huh
00:00:49
thank you great great worked the legal department as well um you mention of a
00:00:53
folder to be you need to do use you will to introduce a strip protein
00:00:57
structure information a source to come precisely guy what what part of the molecule could
00:01:03
be important point to specific target didn't you get into the beauty that it's one
00:01:10
and so i thought all the centennial bunny many directions how to put this it i've seen it but i thought well
00:01:15
why not focus on propane slice it more contributing to the second project to date is that it that it with a project
00:01:21
so then i want to do but i just didn't have the chance to delete this is just use
00:01:26
in alpha for to play train i'm more the focus on predicting that idea sent them from separate things
00:01:32
but what is it sounds awful but looks very well said entertains buttons on it i
00:01:36
thought what is not very good with pretty that have this sort ensure that like flexible
00:01:40
looks like this that are set talk or even the certain proteins cause i thought yes
00:01:44
give your canonical structure you have you been different for the celebrities multiple would not predict that
00:01:50
so first new proteins for this and i thought what is not enough but what i said is that even though when you but
00:01:56
this is not the not you can use these to pertain more
00:01:59
bit and inflate unit with these that visitors at the specific data
00:02:03
and this is what i would like to try and i think there's a lot of potential there that like the high need to have the right collaboration police
00:02:10
but about whether it's and
00:02:14
have the microphones so i have a question about the the kind of do that you would
00:02:21
expect full last example for for instance you have some images that seem to be standards screen
00:02:27
yeah and then in terms of them legally legally movie dawn ah
00:02:34
but one one on the whole thing so you have no special information yeah that's right right yeah that's yeah i don't have to say it right
00:02:40
but i do not make it clear exactly do have this this light and
00:02:44
you only have one label that's like what is the average unit special value
00:02:48
i knew then that they they as i mentioned so at the beginning the and a sequence is done
00:02:52
on the me that section of the biopsies not even on the say in the on the same nice light
00:02:58
of course you hope that it hadn't inhalation and this like this big bit of the bad but it's
00:03:02
also a kind of a special gap between what dennis it comes from and they they like to have
00:03:08
but it's an inch of these uh decided right so here we
00:03:10
have no special information whatsoever you just take the average i donate value
00:03:15
and then you hope that in one that is gonna be smart enough to face special weakens
00:03:19
and this is what if this was trying to show here that we get so hot spots
00:03:23
i would hope that the associated with high special values you probably conclude for twenty thousand genes but you you can do
00:03:29
some gene living room few released spatially simply look to confirm
00:03:34
the predictions that you have that you reach such projects are are
00:03:40
from detention mall actually correspond to these were genes would actually expose
00:03:46
so so exactly that thing there's no hope and trying to buy the finale special
00:03:49
level of twenty thousand is with with accuracy julie when it's and this nomadic here
00:03:54
i think what you can tie that i think the t. is that respect to happening through it from the morphology and so that's what we
00:03:59
focus all for the day that had been select the fastest based upon
00:04:03
a lofty that that used to classify concert concert bacon correct that cancer patients
00:04:07
so we hope that these yes thing happening back to more for you so they can be predicted we set up an acute se from the images
00:04:13
and of course we saw that's one unit so you said is that the product i know
00:04:16
that's what you suspect it's all the stronger the impact the morphology the stronger then back to predict
00:04:21
but also what i show that the accuracy is not perfect of course you know set you
00:04:25
know how they say level of information and and a sickly sequences like that you would have
00:04:29
but you can buy the excel pen with reasonable accuracy and the the thing that i
00:04:34
think that i'm looking at you get this information on their regions that that potentially untouched
00:04:42
hi i'm due for group talk um i have a question about the
00:04:47
teaser circular production arm i think it's quite easy to work to most cities
00:04:53
really to omit be true within you lose the selective e. t. h.
00:04:59
i am in terms of working um my question is did you also include negative information
00:05:07
your two two were enforced selectively to use in use oh
00:05:14
in this case how can you remote you interpret ability for negative information
00:05:21
so this isn't it questioned suggest typically what we discussed probably about like database it's called the the ca
00:05:27
which basically you have reported pairs of these deceptive device to dissipate top but it'll happen at the samples
00:05:33
so what we have done is what people typically though the staff felt that beta i assume
00:05:38
that any posse random combination obviously not in your positive say it to be a negative sit
00:05:44
this is not perfect of course there is more likely cool that one of these negative shuffling is still
00:05:48
bias but we hope that is the likelihood small this is what people typically the when in the literature
00:05:54
i was just i don't listen up and in the beginning of the samples so
00:05:58
i did not show it here but we have don't follow up work in which
00:06:01
we have in the model i mean that but i will more that pipeline specifically
00:06:04
focus on and running the rules be kind of this or that the predictive more that
00:06:09
i'm that actually you can that interpretation bottle deposit examples on
00:06:12
on the negative samples and you can generate rule say okay
00:06:15
these spare bias because these i mean once it is in this position on this one is here here here
00:06:21
and despair that are not bind because you don't have the selling artists here here so you can only get it beep and get really
00:06:26
more humanity but double rules rather than the fashionable of interpret ability
00:06:30
that they showed here but this is the right is an important point
00:06:36
f. f. t. uh_huh uh_huh up and
00:06:44
and thank you my question and is regarding that
00:06:47
person that had project to use our any second data
00:06:51
you face any shoot at them between a model it on the
00:06:55
second as it depends a lot on the technology protocols used and everything
00:06:59
then we want to test and a new data set you call like the all the space that only
00:07:04
sake is we conclude that it's very noisy does state i want to know how you deal with that
00:07:10
well this is not a specific but in a secluded price i think this is their advice if in biology when you
00:07:15
have a small and make it very active assets and it's huge back to fix for the for the technology some platforms
00:07:21
this is a problem this nomadic about it and in the
00:07:24
end you there certain techniques for much correctional for normalisation and
00:07:29
you have to try and it's only the from and what are to decide what is that right pipeline to correct for it
00:07:35
so this is specific problem in this in the first grade and we used to be for latin the drug screening and
00:07:40
we already have survived effects so it's the look like something so with the proper normalisation you can still get full signal
00:07:46
that's a value of surveys have different in technology and interested in what
00:07:51
and for the second project for the last project the image space now as
00:07:54
i said we are working with uh they police it at the ballot university
00:07:58
and we will try this not different data sets so we also expect that
00:08:02
they would be batch effect on a meeting it's got different different staining different scanners
00:08:07
so again it's gonna be another creative effort to try to correct as much as possible for this
00:08:11
but effects i'm sure cancer there's no magic solution you just have to collect as much as you can
00:08:19
i would say that for you describe the the what you can use your
00:08:22
tension maps to interpret what what is doing this or so to to improve them
00:08:27
the aim was h. developing here and i wonder much of the mature shown seems is
00:08:31
that it's it first order in the sense that they will tell you how much um
00:08:35
at the cheese please just disagrees during much you describing the c.
00:08:39
fifty or could you take these two second order no we for example
00:08:42
uh i'm standing right there's attention paid to specific recede you are specific module because there's expression result in
00:08:48
gene in in your your profound for to distract you see fifty and that would be interested in it
00:08:55
we have yeah that's an interesting question i haven't thought about it we have not
00:08:58
going to this that action what we haven't gone maybe that personally answered your question
00:09:03
is that something what happened when he does get marks for individual
00:09:06
features but we know that they should be correlations with the features you
00:09:10
know especially for them protein binding is not the one i mean i see is based on the senate together that contribute to the binding
00:09:16
so why yeah we haven't we have kind of a trying to test is as to how this
00:09:20
hierarchical interpret ability but you first uh kind of credit which features in the middle features an important
00:09:26
to reduce the number of features to explore and then just don't combinations of them so with this smaller set of feature to
00:09:31
try to see what it paid is of these features get you high yeah in but but that really got me that interpretation
00:09:37
and then we're kind of eternity plane you can maybe be able to
00:09:39
kind of up a combined feature gonna get more predicted that single features
00:09:45
but yeah i will have to discuss
Conference Program
The Idiap Research Institute in Martigny is launching a new public series of symposiums
Lonneke, Idiap Research Institute
Jan. 25, 2023 · 9:39 a.m.
420 views
Interpretable artificial intelligence for cancer personalized medicine
Dr. Maria Rodriguez Martinez, Group Leader, IBM
Jan. 25, 2023 · 9:48 a.m.
5 minutes questions: Interpretable artificial intelligence for cancer personalized medicine
Dr. Maria Rodriguez Martinez, Group Leader, IBM
Jan. 25, 2023 · 10:10 a.m.
NIPMAP: Niche Phenotype Mapping of Multiplex Histology Data by Community Ecology.
Dr. Jean Hausser, Assistant Professor, Karolinska Institute, Sweden
Jan. 25, 2023 · 10:21 a.m.
133 views
