Player is loading...
Genomics, AI and Privacy
Julien Duc, Co-Founder and Co-CEO of Nexco Analytics
Wednesday, Oct. 11, 2023 · 2:01 p.m. · 16m 41s
Julien Duc is a leading bioengineer and data scientist working at the frontier between Computer Sciences and Biology. After obtaining his master's degree in Life Sciences at École Polytechnique Fédérale de Lausanne (EPFL), he worked as a data analyst for 10 years at EPFL in the laboratory of Virology and Genetics led by Prof. Didier Trono. There, he developed several bioinformatics and Artificial Intelligence pipelines that played a crucial role in unveiling groundbreaking scientific findings published in prestigious peer-reviewed journals. These methodologies served as the foundation for the establishment of Nexco Analytics in 2022, a Swiss company specializing in delivering Artificial Intelligence solutions for Life Sciences to both private and public sectors.
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
exactly so thank you very much for the invitation and have a nice introduction
00:00:04
so i'm i'm not gonna be too long 'cause it's uh almost time for the up
00:00:08
arrow but yet to them gonna talk about jenna makes uh yeah and and prices so
00:00:12
just before anything else just start with a little introduction about our company so it's called
00:00:17
next going at it takes uh we are artificial intelligence and uh that dan ellis is uh
00:00:22
service providers so we work with academia i would work with
00:00:25
the fire mind history and also um i get in and
00:00:29
uh we're focused on i've signs so we are experts in a and g. s. which is next generation sequencing
00:00:35
so on genome sequencing um transcript domains which sequencing of the expression of jeans i'm
00:00:41
always we have i. p. but we developed together for for more than ten years
00:00:46
and that's actually where we we specialist um so before going into
00:00:51
um into gentle makes let me just start with a situation that
00:00:54
i i'm sure all of you um encoded once you you you wake up you want to go for work or uh or run
00:01:00
and your your shoes look like that so they are company broken so you go on your feet websites to actually buy
00:01:06
some shows and this seems completely harmless but actually we all
00:01:11
everyone is room i think knows that we uh actually feeling
00:01:14
quite a lot of data to the website where we have visiting the and buying a our next gift
00:01:19
and it's it it seems that's we are with anonymous in all of this data because we
00:01:23
know we we give data but with the amount of data that we actually getting to to everyone
00:01:28
and the amount of people doing the same it's maybe it's difficult for us to be traced back right even if we're not loading
00:01:35
but actually it's not sure you just need to have a couple of data
00:01:37
points to be able to infer precisely which can be italy's is who actually
00:01:42
and it's it's not on meeting you i mean it's it's been proven also but uh this mit researchers that
00:01:47
show that with a just a few data points um the from your credit cards uh actually for data point
00:01:53
uh from one on one uh one that one million people um they could actually choice ninety percent of
00:01:58
the people unique so they don't need to know what you buy the just the time and date and
00:02:02
that's just enough to uniquely identify people so if you have have it's it's very easy to find and
00:02:07
you know we record behaviour in out of the so how does this translate to channel make well actually
00:02:13
as you may know that we are all a form by a genie right they need something to instruction to construct a human
00:02:19
being from beginning to end so it's it's quite a lot of
00:02:22
information but how much information actually is is in there so um
00:02:28
it's not moving i guess the computers reachable because the pictures speak but uh so it's all encoded
00:02:33
into well it's what encoded into a four letters right it's this eight a. g. c. that you may
00:02:38
know that founded in it so it's it's just the name of the different money 'cause that are
00:02:42
uh as a link to each other to make this very long string of character defines a our genome
00:02:48
and how long is it exactly for human well it's three billion of these letters one after the other
00:02:54
but if you want to print you would actually end up with a thousand
00:02:58
books of a hundred a thousand pages on the books of a thousand pages
00:03:01
which is a hundred thousand pages um of little letters that
00:03:05
defines just who you are actually and to discriminate between two individual
00:03:10
uh from this phoebe and uh bases you just need to look at seventy places
00:03:14
statistically to distinguish one of the eight billion a person meeting on this planet so it's
00:03:19
it looks like a lot of information but actually to to discriminate individual it's it's quite easy
00:03:25
and it's not completely lost in the in the amount of information that you how so what actually can you read uh
00:03:30
in the genome um and what is what is there to protect their um so jen energy obviously i mean we all
00:03:37
did uh just moments which we know that the jury energies obviously something you can infer from engine on the just looking
00:03:43
annotations you can infer you parents or siblings and your cousins um which is uh some people actually do that for a living
00:03:51
can you actually track a genetic diseases and cancer risk uh just looking at fermentation you can have a new d.
00:03:56
n. a. well the answer is obvious yes because we know that you have this position maybe seemed happy that's booming now
00:04:03
and e. does you know did mine they um they well after beating
00:04:06
the chess game uh the go game the protein folding again without couple
00:04:12
the games paper like a month ago now that is called a and family sense which is a a to do
00:04:18
predicts mutation and to know the pathogen e. of each mutation
00:04:22
on the job so you take the four dimensional disproven bases
00:04:25
and for each of the different mediation you can have compared the it's a a normal individual
00:04:30
you can measure the pathogen you to visit risk for this this is yes no
00:04:34
and you have a risk factor so it's um it's it's just came out now
00:04:38
um can you look at the hair colour and i colour yes obviously this we all know it's it's just come from the normal uh courses
00:04:44
that we don't also but what about heights is is it possible actually to tune all the hype of unusual just looking at his d. n. a.
00:04:51
well actually yes you have study that shows that um you have a big
00:04:54
factor that is coming from the genetic and you can have a pretty good estimate
00:04:58
like ten centimetre and it's not like it's gonna be extremely precise like for the age you
00:05:03
gonna be riverside was you have a very good idea and why wait actually is the same
00:05:08
because you have a factor which is genetic from the um the metabolic
00:05:11
that you have the metabolism of course you can estimate then have a
00:05:16
everything will depend on your diet in yours the sport you do in that i study huh but in general uh you can have a very good estimate
00:05:22
the h. from the genie from the faces apparently it's easy uh uh from
00:05:27
the dean is it also possible uh yes because you have the gender genie
00:05:31
that changes across what what age it's like the end of the call was
00:05:34
only at the shrinking becomes once more if you can have a pretty good estimate
00:05:38
and finally is it even possible just system interface of someone just looking at that ginny and actually
00:05:44
yes you have uh the trend is the the neural network trying to recognise and linking the in imitation
00:05:50
and what just chit jenny in general and the the the shape of faces and it's actually doing a very good um you know this
00:05:56
like this portraits the two for the the police it looks very similar it's not it's not the precise picture but it it's close enough
00:06:03
so with all of this and i'm not saying you know obviously but it's very clear that
00:06:07
generic data is very sensitive and needs to be very uh secure 'cause it's a privacy uh concerned obviously
00:06:15
so how long ah yes last thing uh just the treatment efficacy which is now the the field is moving
00:06:22
because you can sure get in invitations and you can actually pretty
00:06:24
if the treatments for instance let's say for cancer and can have stratification
00:06:28
after different diseases and you can have uh the best treatment which of course is a big uh is a big win win for the
00:06:34
uh for the firemen is if they can predict if the drugs working that one with ones and and of course but the for the patients
00:06:41
so how long exactly does it take to have all of this information uh uh well
00:06:45
with the new technology that are developed now this is another aspect uh from human now
00:06:49
one of the sequence machine that is the most the common now it just takes one hour to have your
00:06:54
functional so the time for you it takes two to eat lunch that they can actually sequence or full uni
00:07:01
so it's it's very quick and this goes of course with the um
00:07:04
a drop in the price in the general price of uh of the sequencing
00:07:07
of william some i i guess you can agree but uh when they
00:07:11
started sequencing the genome so you the human genome project in two thousand one
00:07:15
uh the cost was a hundred million dollar and into doesn't twenty one it's less than one thousand or
00:07:20
and they are talking now at last no hundred dollar to sequence of for a human so it's it's getting cheaper and cheaper
00:07:27
and obviously with the the the um this is price dropping you have
00:07:30
an adoption in general of um of precision diagnostics as i was saying
00:07:35
you can just have a good idea for drugs can work or not uh or we can better better characterise the diseases
00:07:40
it's much better for the for the um let's see the the full house system so they're trying to push this and you have a
00:07:47
this this increase that is predicted from the um the domestic market
00:07:50
actually present agnostic market would increase a lot during the following year
00:07:54
and this is something that timber isn't and it's from up to five and this this uh this thing on fourteen business inside
00:08:00
so um yeah so you have this increase of data of generally data in the public domain in austin
00:08:06
in general i have more more of the state and as i was doing at the beginning more data i have the less privacy you have
00:08:12
um so you need to have solutions for this so the different glitzy attacked
00:08:19
that can happen uh engine regarded the same that were discussed before in peace talks
00:08:23
so you have this identification in this fan type inference so even um what
00:08:28
everyone is doing obviously to comply with a different laws is to have anonymous data
00:08:33
or the identification of the big data sets so you you have all the generic data
00:08:37
you have no information on the um on the the names of the people in the study obviously
00:08:44
and it's just action doesn't work as was said interest talk because you have this paper from i
00:08:48
think it doesn't thirteen that game that's just they identify the participant in a big study my name
00:08:54
just crossing different data sets looking at the different they could infer ceilings
00:08:58
in genealogy interested because you have some of the data that is probably
00:09:01
like the h. some demographics and it's enough crossing with the public databases
00:09:06
to identify people in the study by name you can have the full genome
00:09:10
just like uh some some information um and of course this is very useful to sort a to sort of court
00:09:16
cases is if you have to have this this also i think i i think was five or six years ago
00:09:22
they had this uh this cold case of uh for criminal in the us it was not fun for very long with the hat
00:09:27
the genie so they could trace back looking at probably data sets and they could actually trace
00:09:32
back uh siblings ask the siblings and they could trace the the who it was and counted
00:09:37
i'm hiding information from production homes like masking
00:09:41
some information that would be very useful to recognise
00:09:45
what the corrected aspiration it's a uh for instance you mask everything that is related to diseases
00:09:50
um this is difficult because you can actually inferred information from all the genes
00:09:55
so as soon as you have a germs that match under close enough by siblings
00:09:59
the missing pieces you could replace and it's it's very difficult and also since
00:10:04
a lot of data has been in the public domain for why you completely lose control it is the the data
00:10:10
is is actually there um and you have you markers i bring everyday because people are doing research on this is very
00:10:16
important for new markers of uh of diseases and if you mascot genome then everyday you would have to mask and
00:10:22
if people download the data locally you have no control on what to do with the data so that's this doesn't work
00:10:27
and finally doing aggregated statistics this is what just on a um and works quite
00:10:32
well um but you have to be careful because if some people have and say
00:10:38
partial information on the gene jenny to go for your individual they can actually interfere
00:10:43
if this person was in the study or not so you can know if they are
00:10:46
in uh the case group a or b. knowing if they are with a disease or not
00:10:50
it's a bit like uh the um yes with your networks when you try to infer if that they that was
00:10:55
used to train you can do the same scene of things looking at the the some genetic markers if you want
00:11:01
of course it's not a it's a it's probability that it's it's actually bigger such
00:11:04
that ends in a lot of people are just giving a their genetic data for free
00:11:10
or for very cheap uh two different companies as you may know twenty three and
00:11:13
mia and street you need all this uh i make an a company that are just
00:11:18
for very cheap to take your genome this you can see from being twente uh
00:11:22
and they do some analysis and then you have access to the results we can look
00:11:26
what is your ancestry if you have a risk of disease uh and you have all of this information is available for you and for very
00:11:32
cheap but because actually they are just guessing data that they can use
00:11:35
to do big study that twenty three and we actually did a big um
00:11:39
cell of the data to happen didn't send the data itself the salt some of the results they found
00:11:45
to develop new truck compound so they're working now with a with a farmer company to develop destruct
00:11:50
i'm from the data that it just acquired basically for free and as you
00:11:54
know when you you defeated to to any company you have to hope that
00:11:58
the to the data safe and this is just bringing a lot of work
00:12:02
to my mean because the um actually it's october seven's wasn't in the news
00:12:06
the some crackers were selling the data of millions of um of relief that things from twenty three
00:12:11
in me because they they had access i think it was in the chinese the oration that the base
00:12:15
and they have access to the looking so they could just log in and just
00:12:18
downloads basically the um well just look at the information that any user would have
00:12:23
access to which is all the mutations you have and you can download the data
00:12:27
and are now selling it on the on the dark weapon according to this article
00:12:31
uh and it's a vision i mean that twenty three me acknowledged um and they are very sorry and
00:12:39
the problem is that you don't have a lot of newness of uh action to do anything about that because you have the um last
00:12:46
obviously which is the the the most important thing that should be in place you already have the um different laws according john makes which is
00:12:52
very strong uh you cannot of course have names like with the the agenda date that you have to comply with many things
00:12:59
um but still i think the low it's not strong enough
00:13:02
um because of all the information you can extract from genome
00:13:06
uh i think it should be very clear definition of cool and
00:13:10
what can be done with human genome a study because it's a
00:13:14
it can be it can it very to be probably don't want to have any you know the screen is commissions a of a
00:13:20
of people from hell says them or whatever i mean it needs
00:13:24
to be very protective something that is already in place is control
00:13:28
access so now more and more trucks they down line which is
00:13:32
it's becoming tricky to do analysis but the they are blocking data
00:13:35
meaning that you need to apply need to feel um the full document explaining why you want to have access of the data
00:13:41
what you are going to do with it and that you will destroy the top you after you have been doing you analysis
00:13:47
and this is very important that this is in place to just control who can access to the
00:13:50
data and what people do with uh finally you have storage security of course in france for security
00:13:56
uh this rule we always at the things we tried our best to not keep any day that or
00:14:01
because we want to have to deal with all of this which is to then i says and then we remove
00:14:05
it and same for the transfer uh if we can avoid to transfer data because it's very sensitive data then we can
00:14:11
just run it on site or anything but for that you need to have also strong encryption when you exchange data
00:14:16
and it's quite as you can imagine is quite a lot of data which
00:14:19
is not easy to include the crypt kind just to stand over anyway doesn't work
00:14:24
uh and one thing that really helps but is not
00:14:27
used um at almost i think john excuse the cripple graphic
00:14:32
'cause you cool it's typically do this a home or fig encryption and try to do
00:14:37
computation on the encrypted genome on the clouds and you don't really know what you are looking
00:14:42
at that you could do some analysis the promise that with jenna makes it very often
00:14:46
you need to really understand exactly what is happening so you need to know exactly which mediation
00:14:50
related to put the data sets no it's this one's goes to this chains with this this and and you have
00:14:54
some biology to to to do not only just machine learning you know so it's you have a mix of two
00:15:00
and we knew you just give a result to the badges they want us to to know the full detail of what you
00:15:05
have been doing it needs to be in you need to be transparent it's not gonna be a full machine or anything for you
00:15:11
put it in the black box and have an output and you need to spend a
00:15:14
biology somehow so it's it's tricky but there are some ways and i i mean i'm
00:15:18
thirty and you can plug off eleven live science engine or so i have no proof
00:15:22
to directly but i hope that there will be a clever way to actually do this uh
00:15:27
this uh this analysis or query information without having to decrypt the genome so then
00:15:31
they are they're sick let's see and obviously keep the data encrypted and probably data sets
00:15:37
so yes i'm not gonna be much longer than that um but some of the
00:15:41
take home message i want to uh to transmit to today's that should be good
00:15:44
ice on the right the rice um so you have more margin device it's really
00:15:49
crazy the uh the size of the study that are being put their uh online
00:15:53
nowadays and some of them are just in public space you can just download that you know the more uh and we we need to have a
00:15:59
better like a framework uh to to define and and have a do and
00:16:03
don'ts of the genie basically like what you can do what you gonna do
00:16:06
uh and we didn't name the low d. v. d. if they want the it should exist i'm also just
00:16:13
always work with trusted partners and sign privacy agreements asking for the removal
00:16:17
because i think it's it's very important not to keep the status somewhere
00:16:21
uh you know not on the hard drive just to hang in there
00:16:23
and taking a dust because it's uh it's it's very sensitive data and uh
00:16:28
yes just a racing at what awareness that that sharing your genome is
00:16:31
not only a just a like a given your email address is a good
Conference Program
(Keynote Talk) Privacy-Preserving Machine Learning : theoretical and practical considerations
Prof. Slava Voloshynovskiy, University of Geneva, professor with the Department of Computer Science and head of the Stochastic Information Processing group
Oct. 11, 2023 · 7:55 a.m.
2509 views
5 minutes Q&A - Privacy-Preserving Machine Learning : theoretical and practical considerations
Prof. Slava Voloshynovskiy, University of Geneva, professor with the Department of Computer Science and head of the Stochastic Information Processing group
Oct. 11, 2023 · 8:40 a.m.
Enabling Digital Sovereignity With ML
Vladimir Vujovic, Senior Digital Innovation Manager, SICPA
Oct. 11, 2023 · 8:44 a.m.
5 minutes Q&A - Enabling Digital Sovereignity With ML
Vladimir Vujovic, Senior Digital Innovation Manager, SICPA
Oct. 11, 2023 · 8:58 a.m.
Privacy-Enhanced Computation in the Age of AI
Dr. Dimitar Jechev, Co-founder and CTO of Inpher
Oct. 11, 2023 · 9:01 a.m.
138 views
5 minutes Q&A - Privacy-Enhanced Computation in the Age of AI
Dr. Dimitar Jechev, Co-founder and CTO of Inpher
Oct. 11, 2023 · 9:20 a.m.
Privacy by Design Age Verification & Online Child Safety
Dr. Onur Yürüten, Head of Age Assurance Solutions and Senior ML Engineer in Privately
Oct. 11, 2023 · 9:26 a.m.
5 minutes Q&A - Privacy by Design Age Verification & Online Child Safety
Dr. Onur Yürüten, Head of Age Assurance Solutions and Senior ML Engineer in Privately
Oct. 11, 2023 · 9:41 a.m.
(Keynote Talk) Biometrics in the era of AI: From utopia to dystopia?
Dr. Catherine Jasserand, KU Leuven (Belgium), Marie Skłodowska-Curie fellow at Biometric Law Lab
Oct. 11, 2023 · 11:06 a.m.
5 minutes Q&A - Biometrics in the era of AI: From utopia to dystopia?
Dr. Catherine Jasserand, KU Leuven (Belgium), Marie Skłodowska-Curie fellow at Biometric Law Lab
Oct. 11, 2023 · 11:42 a.m.
AI and Privacy
Alexandre Jotterand, CIPP/E, CIPM, attorney-at-law, partner at id est avocats
Oct. 11, 2023 · 11:48 a.m.
5 minutes Q&A - AI and Privacy
Alexandre Jotterand, CIPP/E, CIPM, attorney-at-law, partner at id est avocats
Oct. 11, 2023 · 12:06 p.m.
Preliminary Pperspectives on the Ethical Implications of GenAI
Julien Pache, A Partner at Ethix and Venture Partner at Verve Ventures
Oct. 11, 2023 · 12:12 p.m.
5 minutes Q&A - Preliminary Pperspectives on the Ethical Implications of GenAI
Julien Pache, A Partner at Ethix and Venture Partner at Verve Ventures
Oct. 11, 2023 · 12:30 p.m.
AI & Media: Can You Still Trust Information
Mounir Krichane, Director of the EPFL Media Center
Oct. 11, 2023 · 12:32 p.m.
5 minutes Q&A - AI & Media: Can You Still Trust Information
Mounir Krichane, Director of the EPFL Media Center
Oct. 11, 2023 · 12:54 p.m.
(Keynote Talk) Unlocking the Power of Artificial Intelligence for Precision Medicine with Privacy-Enhancing Technologies
Prof. Jean Louis Raisaro, CHUV-UNIL, assistant professor of Biomedical Informatics and Data Science at the Faculty of Biology and Medicine and the head of the Clinical Data Science Group at the Biomedical Data Science Center
Oct. 11, 2023 · 1:22 p.m.
5 minutes Q&A - Unlocking the Power of Artificial Intelligence for Precision Medicine with Privacy-Enhancing Technologies
Prof. Jean Louis Raisaro, CHUV-UNIL, assistant professor of Biomedical Informatics and Data Science at the Faculty of Biology and Medicine and the head of the Clinical Data Science Group at the Biomedical Data Science Center
Oct. 11, 2023 · 1:50 p.m.
Genomics, AI and Privacy
Julien Duc, Co-Founder and Co-CEO of Nexco Analytics
Oct. 11, 2023 · 2:01 p.m.
5 minutes Q&A - Genomics, AI and Privacy
Julien Duc, Co-Founder and Co-CEO of Nexco Analytics
Oct. 11, 2023 · 2:18 p.m.
How trust & transparency lead the success of an Idiap student's Master's project in fraud detection
Raphaël Lüthi, Machine Learning and AI Lead at Groupe Mutuel
Oct. 11, 2023 · 2:22 p.m.
5 minutes Q&A - How trust & transparency lead the success of an Idiap student's Master's project in fraud detection
Raphaël Lüthi, Machine Learning and AI Lead at Groupe Mutuel
Oct. 11, 2023 · 2:38 p.m.
Recommended talks
Management of health data in an industrial company
V. Srikanth Nallanthighal
Sept. 5, 2019 · 1:57 p.m.
146 views
