Player is loading...
Applications of AI in oncology drug discovery
Dr. Slavica Dimitrieva, Associate Director & Senior Principal Scientist, Oncology Data Science, Novartis Institutes for BioMedical Research
Wednesday, Jan. 25, 2023 · 12:07 p.m. · 23m 24s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:05
and but also it's a bit call that okay so um
00:00:11
i work in the ontology the design department in and about this in bars not
00:00:16
and today i'm going to give you a high level all that real of how
00:00:19
the up line to beta size across the what you tried to start a pipeline
00:00:24
a specific i i'm going to talk about field equations of a our in our portfolio
00:00:29
so just to give you a brief uh introduction that oh the
00:00:33
the size is very well integrated maybe steps of the drug discovery pipeline
00:00:38
so in early target identification face there and allies
00:00:41
think about two miles from patients but also the at
00:00:44
analysing a data from kansas allies in the text models
00:00:48
so we do lots of experiments on this preclinical models
00:00:52
we did a large scales cleans up to patients screens battery no doubt
00:00:57
specific genes inconsistent lines and then we see how a disaffected the cell proliferation
00:01:02
and this latest identification of your part is to discuss the project done in collaboration
00:01:07
with the buildings that don't now we're also doing these perturbations cleans all single cell level
00:01:13
and uh when we can five compound uh we're
00:01:17
also doing combinations tons of different components to see um
00:01:21
well uh better by combining different components we have a better effect and then for the
00:01:26
components in the optimisation face we're also using imaging and then we're also blanket yeah i hear
00:01:32
on the cycle then uh imaging so for example we
00:01:36
are uh operator being cell lines using different perturbations and
00:01:40
then they are using a uh trying to learn them
00:01:44
all the factions of these the of the different uh perturbations
00:01:47
and then yeah internally that also leveraging all the full to in order to
00:01:52
learn more about a protein protein interaction so we're trying to
00:01:55
predict protein protein interactions using out of all these certain specific cases
00:01:59
and then you also want to identify tractable will get speed in
00:02:03
in in the approach things that are in our targets before you
00:02:08
and then they are using a high to integrate data across the portfolio that
00:02:12
we can still hear the clinical data we get lots of clinical data spell
00:02:16
and we're using k. i. to integrate all these data
00:02:19
we are developing clark internal consisting the cell are close by
00:02:25
and here we also leveraged a i made us in order to remove the page effects between the different data sets
00:02:30
so in these outlets we're uh integrating proprietary data says that they can't
00:02:34
but also probably status is from the literature and we're also using digital technology and
00:02:40
uh uh lately specialised people because they don't in order to inform on uh oh
00:02:46
i own organisational different cell types within the transportation we're also my material what evidence they thought
00:02:52
and we're combining all these data in order to do data data base installation for our clinical trials
00:03:02
so yeah i can i really have an enormous input been accelerating that
00:03:06
pass the research and i will give you just a very brief overview
00:03:09
where we are applying a high in our our drug discovery pipeline so today i'm
00:03:15
mainly going to talk a about how we're using a high to do big integration
00:03:20
in preclinical model selection because we're doing lots of experiments on preclinical models
00:03:25
a lady briefly describe one made that that we develop to remove ambient identical domination
00:03:31
of from a a single so in droplet based and also experiments like the next
00:03:36
and then i will briefly talk about how they had applied a
00:03:39
i don't hide complete imaging to do what i supervise modify action discovery
00:03:44
we are also blanking right to do with our combination production and we're using a
00:03:49
high to mind the real world data is the time the purpose of this is
00:03:54
begin divide by marcus response but also by marcus for it is the stuff
00:03:58
so that begin identifying new targets for the resistant patients in the
00:04:03
goal is to really be able to start a fight the patient
00:04:06
so that we can really and all the patients uh that i patience for the ride clinical trials
00:04:12
and as i mentioned we are also leveraging awful to in our ah
00:04:17
attack discovery pipeline so maybe it'll be i'm going to talk about these the project and uh in particular of obvious one
00:04:26
so internally we developed i made that which we called
00:04:29
that more but still stands for multi origin backed effected
00:04:33
him well and this is done yeah i to uh that to be developed region and it'll do preclinical model selection
00:04:40
but not we're using get violent to data integration across the part for you
00:04:47
so as we all know the clinical cancer research politically lives in preclinical models
00:04:52
let's say kansas delightful these are two more that that the one in the
00:04:56
beach and then a patient to merge universal does that patient years that i think
00:05:00
about that in the middle upper my slices and with a lot of experiments
00:05:04
in these preclinical model so we we to read them i says with different rocks
00:05:09
we do lots of perturbations coming in some classes the lines and then are there is a lot of knowledge
00:05:14
that we derive from this preclinical models but now the question is how much of this is translated well to patients
00:05:22
because we know that there is a difference in the drawer when conditions all these two more stars
00:05:26
and then they also are propagated within the different blocks
00:05:30
and then they change over time so basically that that
00:05:34
at least was that we see in preclinical models is
00:05:37
not always a representative of what we see in patients
00:05:43
so now the question here is how do we know which models that look and we should use
00:05:48
them for preclinical cancer research and which one we should avoid because we have about paulson preclinical models
00:05:54
and for this weekend and a line on the data sets that we can't so for example on patients we can't about
00:06:00
ten ten thousand patients who was from the d. c. j.
00:06:03
consort deal and we can't complete medical characterisation of these to mars
00:06:09
and then we also have a bought mormon thousand cell lines a recap these advances the line is the copied yep
00:06:16
which we developing uh uh we have complete molecular characterisation
00:06:20
but also complete characterisation of darkest was to these cell lines to different
00:06:24
treatments and this list a large project in collaboration with the buildings that don't
00:06:28
internally we also came about tiles and critics models very being crafted patient
00:06:34
to marcy mice is and then we also give complete characterisation of these models
00:06:39
and uh all the smallest little characterisation but also the dark response
00:06:44
and uh uh the problem with the preclinical gas that is that
00:06:47
is that because of the differences in the drawing conditions as i mentioned
00:06:51
and the absence of the two what michael environment which we don't have in these models
00:06:56
the track response that we see in models is not always translate able to patients
00:07:01
and also the data is a mixture of technical artifacts in biological signals
00:07:06
but also there's a lot of page if it were a bit in fact
00:07:09
even when the data come from the same region for example even between the patient
00:07:14
because these patients your basic mess the different centres different hospitals
00:07:18
then the the steve differently bundles just they have different history and so on
00:07:23
so i'll uh now the question is how we can combine all this
00:07:26
preclinical and clinical data to him for a moment lee preclinical model selection
00:07:31
so just to give you on our idea of what the problem is
00:07:36
so when we combine here are all these two miles from preclinical models
00:07:40
a cancer cell lines and upbeat x. models with patient data we see the the
00:07:46
data sega gets based on the origin so on this block digital does that too much
00:07:50
and a cancer cell line starting being pretty exciting to be interested you adding blue
00:07:55
and is that of the the two mercer group based on the two like that which is what we expect
00:08:00
the group based on the origin so now you really want to start a plastic particular patient population
00:08:05
it is very difficult to say what would be our uh what is the best model for that population
00:08:12
so for that to be developed um model which we call to what is so eyes i skipped introductory slides on i'll
00:08:19
think what our budget very briefly explain here what it is
00:08:22
so it's a it's an eight foot here here gene expression data
00:08:27
which comes either from the models or from the patients and
00:08:30
says we're using topic what we're projecting these two little dimensional space
00:08:35
and then we try to reconstruct the data the transcript don't because they don't patience and um
00:08:40
models from these low dimensional space in our we also get
00:08:45
a uh this other set a network here so this is that
00:08:49
source discriminating at at which we are training in other started a fashion so
00:08:53
that we create in bed is here that the three from the source effect
00:08:58
so this a neural network tries to determine what is the origin of the sample
00:09:03
but then it's because it is trained in art history fashion we we make a them begins here so
00:09:09
that this network cannot recognise anymore what is better
00:09:13
as sample time service that line or from a patient
00:09:17
and then we also do all the reconstruction in the conditional manager
00:09:20
so that for example when we want to project everything to patient data
00:09:25
we say that for example the cell line or the pretext bobble uh we if we get
00:09:30
that information to the decor that ended the call the data is if it was coming from patients
00:09:36
so first we change this network and then we do this with a switch
00:09:40
here with the bits so that we can project one data type into another
00:09:46
and uh these are the results so practically here yet aligning five different data set
00:09:51
so this to ga is a dark blue well it's it comes on patient liabilities made five hundred so
00:09:56
does that um with this topic to mars also directing patients and assist in these and pretty exciting that
00:10:03
orange and red and then we share this company in data set so this
00:10:08
is a ah if this is thinking and it comes from a breast cancer
00:10:13
so we think that even men these these patient data it's still class that is on its own but when
00:10:19
we applied this page effect removal method that we develop we see that now the data is a really well aligned
00:10:25
so the cancer models are aligned with the patients but also for example at
00:10:29
the top me in data set is uh aligned well within the breast cancer uh
00:10:34
a. b. s. plastic and yeah so this is one
00:10:38
all material again here every don't discuss i really don't use it too much and
00:10:43
it's coloured uh based on the too much back so this is the best class that
00:10:47
and uh the patients are are uh the models are uh uh with the uh the cell lines
00:10:53
that with circles and they are in bold and indian was we see the p. t. a. symbols
00:10:58
and we see that in most of the cases they are lined up in a well with the patients that was
00:11:05
but this allow was living to find willow that kept the
00:11:08
birch or i've not looked anymore to be used in preclinical research
00:11:12
and now we can evaluate let's say which are still lines demote class
00:11:16
that together with the respective plastic so if everything was perfect it would
00:11:20
have been in that they are cannot but for example receipt here the
00:11:23
the best still lines are not a good representative of breast patient to mars
00:11:28
which is not the case with the obvious it takes more uh being sorry we'd been picnics models
00:11:34
so the brain to take small so good representative
00:11:36
but the brain um consistent lines they're not good representative
00:11:40
but for example it will here so this is uh i'm sorry the small but this is key here and we see that most
00:11:48
of them are cell lines that we cared about that i've derived
00:11:51
from scheme can set the really well they're good representative of melanoma patients
00:11:59
and uh i also wanted to mention that uh the alignment that we do it so not completely unsupervised well but
00:12:05
a way but we can preserve the stop types of for example in restaurants that we have uh these different sub types
00:12:12
i and with these integration of the mob boss of subtype from the lyman i'll start times and so on
00:12:18
and then says we're using i'll think well that we can actually
00:12:22
investigate what happens when we aligned that might preclinical models to patients
00:12:27
so here for example this is that we cannot lot of the g. is that we find is differentially expressed envy
00:12:32
outline of breast cancer cell lines to breast cancer patients and anything that lots of the genes that we
00:12:39
see as articulated are related to the to micro environment
00:12:42
because based is highly fibre optic to show a concept
00:12:46
and then with a lot so far uh things that related to flip roles that we
00:12:51
in the two micro environment that we think patients but we don't see in the cell lines
00:12:57
so how we apply a little bit across the course volume
00:13:01
so i will give one example of far how we achieve bet that
00:13:06
patients with ability using this kind of uh our uh using a i
00:13:11
uh so i can't values inc one data set that was published two years ago by that that mark was or you know
00:13:16
so they inject that us allies into the heart of mice and then they
00:13:21
wanted to stay back there where these cell lines will my task is that
00:13:25
in which the lines though matt asked the sites faster and then after five weeks they they collected
00:13:32
five different organs from the mice in the back of the the cell lines and then they counted
00:13:36
how many of of each cell line obviously in each of the different organs and then
00:13:41
they they don't have a sort of file mutation mark for each of these the lines
00:13:46
and we wanted to use this data to understand why some primary to myself for making
00:13:51
it that's this is a faster than others and then what are the driving jeez forgot
00:13:58
so what we did here with the transcript bombs all the of the cell lines
00:14:03
and the matters that the potential scores that week yes from the thing that mark and then between the she learned models
00:14:08
to see which uh genes are productive for chi or lonely to stop discourse
00:14:14
and it is we wanted to know how this is the relevant for patients because this experiment was done on cell lines
00:14:20
and then we took the patients transcript comes that we kept from t. c. g. a. and he was
00:14:25
the only the president models all set lines and then we predict that met the static scores on patients
00:14:32
but we don't know what are the only to start discourse in patients that we
00:14:35
could look at the stage of the patients and then all at the survival of
00:14:39
these patients and we think that it really for the patients were which we predicted
00:14:43
time at the starting score they cared for survival so this is the line into
00:14:48
but that is the little correlation with the stage of the stations
00:14:52
so next about with it we used our big defect removal tool so we project that the cell lines
00:14:58
to patients so b. s. if they but patience and we'll that project the data to train these models
00:15:05
so practically we change the middle those now we uh
00:15:08
this project that as they are with the project that uh
00:15:12
they tell still lies to patients and then when they arrived another set of genes care
00:15:18
and notice a hell of a there this is more relevant for patients begin that to the patient beta
00:15:24
now with these strange models to be predicted them at the static scores and now when we do this
00:15:30
the decimation here we see that now we can back that about
00:15:34
predictions actually we could identify jeez that that more relevant for patients
00:15:38
so here is the better the separation of survival course the prevent you without
00:15:44
is more significant enlisting the tendency for correlation with the stage so it
00:15:48
we we predict a higher matters that discourse for the patients that that it later stage
00:15:54
so this is just one of the examples that we it how we're using this method but internally we're also use it a lot
00:16:00
to remove better face between data's it's coming from different origin and
00:16:05
next i want it to briefly go to other examples of a i
00:16:10
so i'm going to talk next very briefly i bought a machine learning math on that we develop
00:16:15
to remove ambient are in a uh from droplet base single cell experiments
00:16:21
so when it into a single cell experiments uh using let's say ten x. technology
00:16:27
being capsule laid the cell in a droplet and then is to cleanse the compared to what is in the droplet
00:16:33
and not the problem is that uh are sometimes uh and it often happens to be cancer tissues
00:16:39
some cells that are more sensitive and then they break down more easily than others and when that happens
00:16:45
a base that they are and aren't they into the suspension here so
00:16:50
they had in mind and then when it was get into the drop let
00:16:53
and then they get incorporated here and then we cannot recognise the more
00:16:58
whether it's contamination no it's really neat transcript come from the cell
00:17:03
and uh uh for that we can look at the into droplets o. l. droplets that
00:17:08
do not contain assemblies that but they would still have some more because the site
00:17:14
and um and just to give you an evil station up all the program so let's say we can't this thing will sell data sets yet
00:17:21
and uh we say that for example the gene is expressed in b. cells
00:17:26
and it should be expressed only in b. cells but then with the expiration of it
00:17:30
kind of in a they are the cell and this is just because of the contamination
00:17:34
so we see kind level exploration b. cells but also the other labels that not zero
00:17:39
they are uh the c. b. s. is still high levels of expression of these gene
00:17:45
a a in dallas all types as well and this is typically
00:17:48
not a problem is single so experiments in many single so experiments because
00:17:54
we can still plaster the cell site and we can determine the cell type but it just a problem for us because
00:18:00
we wanted to do an experiment where we could in fact the cell lines with different ah a. g. r. an ace
00:18:07
so that we can so these are what what it is gerard and his battery wants to knock out specific genes
00:18:13
and then we wanted to do sequencing on the single cell level to see what is the infected all these knockouts
00:18:21
soul tactically after the in fact that the cells that we want that after a certain period we wanted to cancel
00:18:28
how many cells the remaining played with led let's say but there are these block the
00:18:33
proliferation of the cells like ben week known doused with the g. overeating crazed and so on
00:18:39
and he had this kind of this is that really a problem because we wanted
00:18:43
to know later with which specific injury i don't maybe in fact the the cells
00:18:48
and then now and then that's why we
00:18:51
develop our solution how to remove uh this contamination
00:18:55
it was here so i'm not going to going to do it was because of our time results
00:19:00
but we call this metal testing the so i'm going to do while it is available on the top
00:19:04
and we're using the empty droplets to estimate the contamination profile
00:19:09
and then we're using the neural network to estimate the levels of contamination
00:19:14
and to do this uh this thing tandem and here off the two council members of the noise
00:19:21
and very briefly i will describe that uh uh one i made out of a real sky compared images
00:19:28
to uh in phantom all the functions of uh uh the different compounds that
00:19:33
week yet so what time the experiments that we're doing in this space are
00:19:38
are we are placing 'cause cell lines or are on where else and then we are
00:19:43
applying different chemical perturbations in the different uh when i was with different dosages and so on
00:19:49
and then we are using high contact indigent and then we want
00:19:53
to identify what is the mood of options of these chemical perturbations here
00:19:58
and for that we developed an internal diplomatic method that has been polishing by from artist to your circle
00:20:04
n. in completely unsupervised manner via clustering these they don't uh
00:20:09
oh we're protesting this data and then get plastic and then uh
00:20:13
uh this way we can do for them all the functions for
00:20:16
the compounds that we don't really know what they're doing because in
00:20:19
in our before your we also hear this type of compounds and
00:20:23
then this also inform stuff all on on the dark um curves
00:20:27
so for example the apply different concentrations and then we're measuring the
00:20:31
f. x. and then we can see how this curve looks like
00:20:34
and what will be the optimal concentration of that specific track will
00:20:38
also using cell segmentation for example when we apply specific activation we want
00:20:44
to count on the two more cells how many of them survive
00:20:46
or not and then we're doing that in automatic manner using a i
00:20:52
and now very briefly i'll just want to mention that uh
00:20:55
there will so freely applications on t. i. that we're doing
00:20:59
so we are also looking get a jar combinations using single cell data
00:21:04
so typically in the past i'll uh the patients were uh and uh uh over to that let's say so
00:21:11
we will build bach artistic thing from our patients to
00:21:14
mark and then uh we compute with that on average cannot
00:21:18
and then we will to given on drug to that patient uh that corresponds to
00:21:23
this average long and typically what we think leaning says that um some patients will respond
00:21:30
but then the sample programs because the the track may not cool with the two more calls
00:21:36
so now we're moving stalwarts more precision uh how they're x. so that we want to identify the
00:21:41
different close it single cell level and to do a uh to use uh to give different tracks
00:21:48
and for that we're leveraging k. eyes val and this is my still work in
00:21:52
progress but uh for each tomorrow because the two met our target the ridge areas
00:21:56
we compute the different calls and then we try to predict which interact aquatic a lot on the least possible
00:22:03
so for example here with the the drag two we is affecting uh the blender yellow long and directly is
00:22:09
affecting the uh the green one so these two should be given to combination in someone so just to be
00:22:18
i want to finish here so they are really lots of applications of a i and that
00:22:23
uh there are many different and uh they are affecting the
00:22:26
drug discovery pipeline in many different phases in our core for your
00:22:31
and uh in the future ideally school that he i will uh oh speedup the darkness got a pipeline
00:22:38
it won't do discovery on it all but it has the
00:22:42
potential to really make a doing it would be to increase
00:22:47
the button to speed up all the steps within the there
00:22:50
are discarded process so with that they just want to acknowledge the
00:22:54
the team specific to the team impossible so we're not so many data
00:22:58
so this is just local would you they designs the men but i thought
00:23:02
and that yeah so we share main all dressed in different uh in favour of colours and then we class there
00:23:08
uh and yeah these are the people that they really want to thank and
00:23:13
yeah so uh this is just the oncology they designs in by the way
00:23:16
so i can see where not many more but uh we also have a
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
5 minutes questions: NIPMAP: Niche Phenotype Mapping of Multiplex Histology Data by Community Ecology.
Dr. Jean Hausser, Assistant Professor, Karolinska Institute, Sweden
Jan. 25, 2023 · 10:43 a.m.
Data Science for Precision Oncology
Prof. Olivier Michielin, Head of the Center of Precision Oncology, CHUV, and Group Leader at the Swiss Institute of Bioinformatics
Jan. 25, 2023 · 11:24 a.m.
5 minutest questions: Data Science for Precision Oncology
Prof. Olivier Michielin, Head of the Center of Precision Oncology, CHUV, and Group Leader at the Swiss Institute of Bioinformatics
Jan. 25, 2023 · 11:54 a.m.
Applications of AI in oncology drug discovery
Dr. Slavica Dimitrieva, Associate Director & Senior Principal Scientist, Oncology Data Science, Novartis Institutes for BioMedical Research
Jan. 25, 2023 · 12:07 p.m.
5 minutes questions: Applications of AI in oncology drug discovery
Dr. Slavica Dimitrieva, Associate Director & Senior Principal Scientist, Oncology Data Science, Novartis Institutes for BioMedical Research
Jan. 25, 2023 · 12:32 p.m.
Recommended talks
Big data: possibilities and threats for complex disease genetics
Zoltan Kutalik
Feb. 1, 2019 · 12:04 p.m.
1347 views