YT YouTube2Text

YouTube2Text → Transcripts → MPG Primer: Genomic Privacy: Key Issues and Emerging Solutions (2024)

MPG Primer: Genomic Privacy: Key Issues and Emerging Solutions (2024) — Transcript

by Broad Institute · 8,064 words · 1,251 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00great thank you all for being here today
0:02for our MPG primer how could you
0:06forget we are very excited to have Dr
0:09hun Cho here who's an assistant
0:10professor of biomedical information and
0:12data science at Yale University he was
0:15previously here at bro as a Schmid
0:17fellow and uh principal investigator he
0:19received his PhD in electrical
0:21engineering and computer science at MIT
0:23in 2009 um and before that had been at
0:25Stanford where he got both a bachelor's
0:27and master's degree uh in computer
0:29science with honors he's been named a
0:31recipient of the NIH director's early
0:33Independence award
0:35congratulations um and uh the Cho lab
0:37now works to create algorithmic
0:39solutions to tackle a variety of
0:40computational challenges uh introduced
0:42by the presence now of these large scale
0:45heterogeneous um data sets that have uh
0:48important privacy concerns and so Dr Cho
0:51is very happy to take your questions
0:53during the talk today um and so please
0:55feel free to come up to the microphone
0:57uh or type in the Q&A on Zoom um and
1:00thank you again for joining us all the
1:01way from you yeah thank you Sarah thanks
1:03everyone for making here in person
1:05despite the rain I appreciate people on
1:06Zoom as well um as Sarah mentioned I was
1:10a fellow here for four years before I
1:12moved to Yale last September um so I'm
1:14very familiar with this space it's
1:15always good to be
1:17back so yeah in this primary I'll be
1:19talking about genomic privacy and um
1:22given my background in computer science
1:24I'll be mostly sharing my perspectives
1:26on algorithmic and computational side of
1:28things so just wanted to mention that up
1:29front
1:30uh in the first part of this talk I'll
1:33be giving you an overview of privacy
1:34risks that are known to be associated
1:36with genomic data and then in the second
1:38part I'll talk about some of these uh
1:40emerging Solutions based on what are
1:42called privacy enhancing Technologies
1:44for addressing these risks and
1:46throughout the talk I'll try to both uh
1:48you know Define the the key Concepts in
1:50these areas and also give you some
1:53vignettes of our recent work on these
1:55areas as well so I want to first start
1:58by talking about what we mean when we
2:00say privacy and distinguish it from a
2:03another closely related term security so
2:06privacy in general is a right to be let
2:08alone that's a common way to define it
2:11and it also means that an individual has
2:12control over how information about them
2:16uh is collected shared or used in
2:19different settings on the other hand
2:21security is generally about a property
2:23of a system that provides protection
2:25against unauthorized attempts at
2:27accessing the data or tampering with it
2:30and also tampering with the uh the
2:31functionality of these systems so these
2:33are closely related as you can tell um
2:37but one does not necessarily imply the
2:39other and we're generally interested in
2:41developing systems that are both secure
2:43and provide protection uh for the
2:44privacy of the underlying
2:47individuals and as you know privacy is
2:49an important topic in biom medicine and
2:51genomics uh in general and this is
2:54mostly due to the fact that biomedical
2:57information it contains something that
2:59very that's very
3:00personal to us uh about our biology and
3:04health and the thing that makes this
3:06problem so complex is that there are all
3:08these different types of biomedical data
3:10types that uh reveal different things
3:13about our biology so including things
3:15like electronic health records medical
3:17images personal genomes and more
3:19recently data from wearable devices
3:21there's also these variety of omix data
3:24types transcriptomics proteomics and
3:26also metagenomic data from microbiomes
3:29so all these different types tell us
3:31different things about um human health
3:33but at the same time they increase the
3:34stes uh the stakes for keeping all this
3:37information safe and protect it from
3:39potential
3:41misuse and this is not just hypothetical
3:44concerns there um many real world
3:46examples of privacy threats you may have
3:48seen in the news uh a recent privacy
3:50breach data breach at 23 andme last year
3:53which affected millions of users the
3:55attackers were able to access raw
3:57genotype data as well as health related
3:59information like disease predisposition
4:01status as well as genetic relative
4:03information for these users is quite
4:06concerning and it's also even more
4:08concerning that the attackers were able
4:10to create these curated lists targeting
4:12specific groups of individuals and they
4:15were s selling these data on the dark
4:18web here's another more recent example
4:20from just a few weeks ago I think the
4:22situation is still unfolding there was a
4:24large uh ransomware attack at change
4:27Healthcare which is a prescription
4:29processing arm of this insurance company
4:32called United
4:34Health it was reported that around 94%
4:38of hospitals in the United States were
4:39affected financially due to this attack
4:42and there's still an ongoing
4:43investigation into what kinds of data
4:46had been accessed uh through this
4:49attack so there are real threats and I
4:52want to make the point that this is not
4:53just a concern in uh the in the
4:55commercial
4:56domain those of us uh who are involved
4:59in academic research and genomics
5:00genomics should also care about privacy
5:02too and here are some reasons so behind
5:04every data set there are real people who
5:06could be affected by a potential privacy
5:08breach so we have a certain amount of
5:11responsibility to uh look for safer ways
5:13to conduct research and second these
5:16privacy concerns and related regulations
5:18they make it difficult to for
5:21researchers across different
5:23institutions to collaborate and share
5:24data so having better ways to protect
5:26privacy can help us uh break down these
5:29barriers
5:30and acceler
5:32Science and lastly if we think about
5:35converting biomedical insights from from
5:37uh Research into public health impact
5:39for example for supporting health
5:41related decision- making for private
5:42individuals we need reliable and
5:45trustworthy systems that can process
5:47private information so having better
5:49safeguards can help us in that regard as
5:52well so now let's talk about what's
5:55known about genomic privacy in more
5:57detail um so there are Key Properties of
5:59genetic data uh that make it unique and
6:03different from other types of personal
6:05information like passwords or online
6:07search history uh and what have you
6:10these are six of these properties that
6:12were summarized by uh this review
6:14article by naita at
6:16all I'll briefly go through them so uh
6:19genetic data it contains information
6:21about our health and behavior as I
6:22mentioned this is static kind of
6:24information it's mostly determined at
6:26Birth and it's hard to change it's also
6:28very unique to each IND idual so only
6:30tens of snips let's say would be enough
6:32to single out an individual from the
6:34global human population Mystique is
6:36referring to the fact that there's still
6:38a lot that we don't know about what can
6:39be inferred from genetic information
6:41these data have intrinsic value because
6:44they can be used for biomedical research
6:45or clinical purposes and lastly kinship
6:48means that when a privacy breach happens
6:50it reveals information not just about
6:52the affected individuals but also about
6:54their biological relatives so when we're
6:56talking about risks and safeguards it's
6:58important that we consider these uh
7:00properties of the
7:01data and there's been a lot of
7:03literature on different ways of
7:05breaching genomic privacy and this is a
7:08categorization that was provided in this
7:09other review article in nature genetics
7:11from a few years ago and they broadly
7:14classify the the attacks into two
7:16categories identification and phenotype
7:18inference uh I think this is fairly
7:20self-explanatory but just to Define it
7:22uh for clarity identification means that
7:24you start with a supposedly Anonymous
7:26data instance and we're uh uncovering
7:29the real IND idual behind that data
7:30instance pype inference means we start
7:33with some piece of data about an IND
7:35individual and we infer additional
7:37attributes that are not included in the
7:38data set so including their health
7:39status disease predisposition things
7:41like that so in the next few slides I'll
7:44go over some of the seminal works that
7:46explore different types of privacy
7:48attacks in this
7:49domain one of the earliest examples was
7:52from 2018 uh when next Generation
7:55sequencing first came about this attack
7:57demonstrated the difficulties of asking
7:59phenotypic information and genetic data
8:02so uh the first personal genome that was
8:05released was uh from Dr James Watson's
8:08um sample and in this release data set
8:12variance in the apoe gene which is known
8:14to be linked to alzheimer's disease were
8:16masked due to personal requests from Dr
8:19Watson but just a few months later
8:21another group of researchers show that
8:23these were actually not really Mass uh
8:24you could infer these variants using
8:27nearby variants using linkage dis
8:29equilibrium patterns in other words we
8:31could impute these
8:32genotypes so this led to a larger window
8:35around this Gene of two uh meab base
8:38pairs uh to be excluded from the data
8:41set and note that this is just for
8:43hiding one Gene and so it's going to be
8:46really difficult if you want to hide
8:47information about much larger array of
8:50disease Rel genes that we know uh that
8:52are scattered across the
8:55genome and from the same year there was
8:57also this seminal work by Homer at
9:00where they showed that even aggregate
9:02level information like minor Al
9:04frequencies and gwos summary statistics
9:07can reveal private information as well
9:09and what they were able to show is that
9:10given a Target subject's genome you
9:12could compare it against these released
9:14statistics to figure out whether the
9:16person was in the underlying cohort or
9:19not and this is also called membership
9:21inference attack and while this
9:23information about membership might not
9:25be too concerning by itself uh sometimes
9:27you study course or associate with
9:29sensitive attributes like rare medical
9:32conditions so that could lead to a
9:33potential privacy breach and uh which is
9:36a major concern
9:38here and uh so the intuition here behind
9:41this attack is that if if the subject
9:44has some rare variants that are uh not
9:48really found in the general population
9:49at high frequency then if these variants
9:51are represented highly in the released
9:54data then it makes it more likely that
9:55the subject wasn't the in the original
9:57data set that's the main idea
10:00in a more recent work people have shown
10:03that it's also possible to demonstrate
10:05this type of leakage uh with genomic
10:07Beacon Services where uh we answer
10:09questions like uh is is there a person
10:12with this particular mutation in the
10:14underlying data set and you can think of
10:16this as a different type of aggregate
10:19level information that's focusing on
10:21individual mutations and the idea here
10:24is the Same by combining information
10:26across different queries we can try to
10:27single out a uh an
10:31individual another type of attack that's
10:34been um receiving a lot of attention in
10:37the media uh a few years ago is
10:40identification attack based on public
10:42genealogy databases uh like GED match so
10:45these Services include genetic profiles
10:47from millions of users and their
10:50genealogical
10:51relationships and this is a great
10:53resource for recreational purposes for
10:55finding one's Roots but they can also be
10:56used for other purposes uh like police
10:58officers to catch um criminals based on
11:01the sample that they get from the crime
11:03scenes and the the main idea here is
11:05that you again you start with a genetic
11:07profile from a Target subject and then
11:09we create against these databases to
11:11find out their uh close or distant
11:13relatives in these databases and we know
11:15who they are so now we can try to
11:17triangulate and figure out who the
11:18person was uh sometimes additionally
11:20using information like their um their
11:23demographic information like age or
11:26rough geographic area where the person
11:28is
11:29and there's been several papers that
11:32explore different aspects of this uh
11:34threat Factor so this first paper by
11:36jimar at all um I think Melissa came
11:39here to give it talk a few months ago
11:42last year and uh so this work showed
11:45that just using short tandem repeats on
11:47chromosome y um You can predict the
11:50surname of the target individual later
11:54early Cal show that uh the majority of
11:57individuals of European descent are
11:59susceptible to an identity infer attack
12:02like the one that I explained and this
12:03is indeed how um the Golden State killer
12:05in San Francisco was caught if you're
12:07familiar with that
12:08story and more recently NL showed that
12:12uh it's actually not just the the target
12:14subject whose ident identity could be
12:16revealed the individuals in the database
12:19itself could also have their genomes
12:21leaked U through what's being revealed
12:22by these services and uh they also
12:25demonstrated that one can insert these
12:28uh artificial samples into these
12:30databases to Mis direct uh police
12:32investigators who trying to find the the
12:35the
12:37perer so going Beyond genotype profiles
12:41there's also leakage in other functional
12:43genomic data that we need to think about
12:45uh so I'm referring to gene expression
12:47data protein expression things like that
12:50the idea here is that even though these
12:52other types of molecular measurements
12:54tend to be a lot noisier they contain
12:56enough identifying information so that
12:58we can link uh samples from the same
13:01individual across different data sources
13:04and this is called a linkage
13:05attack and going beyond that we can also
13:08try to predict genotypes from these data
13:11through what are called quantitative
13:12trait low size or
13:14qtls these are genetic variants that are
13:16correlated with other molecular
13:19measurements and by predicting genotypes
13:22of the person based on their omix data
13:23we can also create a linkage between
13:25their genetic profile data in one data
13:27set against another omix data in a
13:30different data
13:31set and as I mentioned uh these
13:34possibilities have been demonstrated for
13:36different types of uh data so including
13:38transatomic proteomics and microbiome
13:41data and next I want to describe a
13:43little bit about our recent work from
13:44last year in genome research where we
13:46try to uh study this question of um this
13:51topic of transatomic data privacy a
13:52little
13:54better so again we're interested in
13:56looking at these quantitative tra low
13:58size uh but for gene expression so these
14:00are called
14:01eqtls and the concern is that a an
14:05attacker a potential attacker can start
14:07with the gene expression level that
14:09they've obtained and try to guess the
14:11genotypes of that person and there are
14:13many of these eils in the genome uh so I
14:16think it's known that the majority of
14:17human genes have some utl associated
14:20with them so you can imagine extracting
14:23a substantial portion of one's uh DNA
14:27through uh this approach
14:30the main issue is that the existing
14:32Works um on this
14:34idea were focusing on demonstrating this
14:37possibility and not really fully
14:39characterizing the extent of leakage so
14:42there were limitations in their
14:43statistical models that allow them to
14:45only analyze a smaller subset of eqls
14:48that are statistically independent so
14:50this is the problem that we try to
14:51address and we introduce this model
14:53that's working at the sequence level uh
14:55so we're able to look at all the DLS
14:58accounting for their corations as well
15:00for those of you who are familiar with
15:01this we're starting with this Le and
15:02Stevens hden Markoff model which um
15:05basically puts a distribution over an
15:07entire genetic sequence accounting for
15:09Link Link this equilibrium and we're
15:12adding another layer on top of this
15:14model or on the bottom here uh that's
15:17corresponding to gene expression
15:18variables and we create these links that
15:20capture statistical correlation between
15:22genotypes and gene expression and note
15:24that this is again including all theuts
15:26that are known and we train this model
15:29in an end to-end fashion so that the
15:31probabilities are calibrated to account
15:33for uh correlation between nearby
15:37genotypes and the idea is that this
15:39using this model we can produce a match
15:41score for a given pair of gene
15:43expression and genotype profile and this
15:46could be used to try to link individuals
15:48across data sets uh which can
15:50potentially lead to lead to
15:52reidentification
15:54so in terms of evaluating this model we
15:57took a data set we had pairs of these
15:59samples gen expression genotype samples
16:02uh from a set of individuals around 300
16:05individuals here in this case we
16:06included additional genotype profiles
16:08from a larger uh um data set in this
16:12case it was hype reference Consortium
16:14data and this was a much larger set so
16:1622,000 individuals in the um in the
16:19additional cohort we're trying to figure
16:21out for a given gen expression profile
16:24uh which genotype profile it corresponds
16:26to and what I'm showing you here is the
16:28fraction of these individuals who were
16:30able to correctly link between these two
16:32samples as we add more of the additional
16:35candidate um genotype profiles to the uh
16:39to the data
16:40set and I'm comparing our method DSM to
16:44two other Baseline methods ebl and gmbb
16:46these are the the previous approaches
16:48that I briefly mentioned that uses an
16:50independent set of
16:53etls the main thing to note here is that
16:55even after including this large number
16:58of individuals so uh 22,000 individuals
17:01we're still able to link around 90% of
17:03more of the original individuals based
17:05on gen expression information
17:08only and uh our method is able to link a
17:12greater fraction compared to the
17:13Baseline and this Gap becomes more
17:16pronounced if we consider a a setting
17:20that's using a smaller set of eils as
17:22well so just uh based on Chism 20 um the
17:26overall numbers go down a lot but the
17:28gap our method and the previous
17:29approaches uh becomes
17:33clearer the other thing that we can do
17:35here is that we can measure the
17:37confidence of our prediction of the
17:39links and this is a metric that we
17:42developed uh it's basically a P value in
17:45some sense that measures the gap between
17:46the best match and the second best
17:48match and if we compare this between our
17:51method and the previous methods you can
17:52also see here that uh our method is on
17:55the y- axis the previous ones are on the
17:57xaxis the the the confidence that we
17:59have over the links that we produce tend
18:01to be
18:03greater and what this leads to is that
18:06if we're working in a setting where we
18:08don't know if the matching individual is
18:10in the candidate pool or not we have to
18:12draw a threshold somewhere right and
18:14have some uh number of false positives
18:17and uh because of this higher confidence
18:20uh we also get predictions that are um
18:24that minimizes the number of mistakes
18:27compared to the the through mattress
18:29that we're able to find as a trade-off
18:30fing Precision
18:33recall and now I want to describe
18:35another recent work which is
18:37illustrating a different type of um uh
18:40privacy risk and I'm bringing this up
18:42because I think this is a th Vector that
18:44not a lot of us uh you know have been
18:48aware of so it's kind of a neat idea I
18:51think that we should all be aware of and
18:54so here we're looking at the security of
18:56genotype imputation servers I think many
18:59of you know about this and so this is
19:01referring to services like topmed and
19:03Michigan imputation servers what they
19:05generally do is we provide a uh a
19:08partially observed genotype profile from
19:11array uh microarray genotyping platforms
19:14for example and then we obtain an
19:17imputed genotypes that fills in all
19:18these missing positions and the way
19:21we're able to do this is by using a
19:23large reference panel of high quality
19:26genomes and using linkage dis
19:28equilibrium pattern uh between nearby
19:30Snips and the question that we asked
19:33was this process is using some
19:36information from the reference panel to
19:37be able to fill in the missing data but
19:39how much information is actually being
19:41revealed uh throughout this process and
19:43included in the output and in fact can
19:45we actually try to reconstruct some of
19:46these sequences in the reference panel
19:48by interacting with these servers in
19:50certain
19:52ways and just to avoid any doubt we were
19:55we took great care to do this work um
19:58respon II L and ethically so none of
20:00this work was performed on these servers
20:01it was all done in a local simulation
20:03environment we also informed the
20:05stakeholders well in advance actually
20:07almost a year in advance and uh we're
20:10aware that additional security measures
20:12have been introduced after we published
20:14our work and all the information that
20:15I'm sharing today is public and it's
20:17published last
20:19year okay so the the key idea that led
20:22to this work is actually the fact that
20:24if you plug in an input sequence that
20:27uniquely matches with a single reference
20:28sequence it will extract the the
20:31matching person's genome um almost
20:34exactly so this is this uh this figure
20:38is showing that idea so uh in this
20:40bottom left corner you're seeing an
20:42example output from the imputation tool
20:45um we're showing it as predict the
20:47dosages which is a value between zero
20:49and one and you can see this value
20:51across the the genomic window that's
20:52being imputed on the
20:54xaxis if there's a single match the
20:56predictions are very confident it's near
20:59zero near one and this corresponds to
21:01the actual genotypes of the matching
21:03person what's even more surprising is
21:05that if you have a match um have a query
21:08that matches a small number of reference
21:10samples you can actually see this unique
21:14pattern that corresponds to that too so
21:15if it matches three samples then you'll
21:17see these predicted dosages that are
21:19aggregated around multiples of 1/3s so
21:22what this means is that you can look at
21:23the imputation output and figure out how
21:24many mattress there were in the
21:26reference panel and what does these to
21:28is this type of um potential attack
21:31pipeline where we start with a lot of
21:33random queries impute them and see what
21:36we get in terms of these dosage outputs
21:39and figure out which one of them uh was
21:41unique which ones had multiple matches
21:43and for those with multiple matches we
21:45can try to extend the query to make it
21:46more unique and in the end we pull
21:49together all of the unique matches and
21:50this corresponds to a set of um path
21:53type sequences from the reference
21:57panel and using 1,000 genomes data as
22:01the reference panel uh as an example we
22:03were able to demonstrate that this type
22:04of attack can actually extract a fair
22:06amount of sequences from the reference
22:08panel so for example if we impute 1
22:12million um input queries that we
22:14constructed in some way we're able to
22:17extract almost 80% of the uh the
22:20sequences in the reference panel there's
22:22some amount of error that's happening
22:24here because of the uh Mis
22:26classifications um in the pipeline but
22:29this error rate tends to be fairly
22:32low and here's a slightly different
22:35setting where you know to try to prevent
22:38this attack what we could do is instead
22:40of releasing the predicted probabilities
22:44we can uh release just the discrete
22:46genotype predictions so either zero or
22:49one nothing in between but even in the
22:51setting there is a workr that we that
22:53allows us to extract unique matches so
22:56the general idea is that you can take
22:59the output um sequence introduce
23:02additional
23:04changes or or keep a subset of positions
23:07uh mask again some of the other
23:09positions imputed again if the results
23:12stays the same then it's more likely
23:14that the original sequence that was
23:15output was corresponding to a unique
23:17individual so using this type of idea we
23:19can still extract sequences um at
23:21slightly lower Effectiveness as you can
23:25see one big caveat here is that
23:28sequences that we're getting here are
23:30chunked from the private uh genomes it's
23:33not the entire genome and this is
23:35because imputation pipeline is run
23:36independently on uh smaller genomic
23:39Windows typically around 20 megabase
23:42pairs but it turns out that you can
23:44start from these fragments of the genome
23:47and Link them across different Windows
23:49using relatedness patterns and uh this
23:51is just a visualization I won't go into
23:53details here that if you compute the
23:55kinship coefficient between these
23:57individual pieces with um relative of a
24:00certain degree you can see that the
24:02kinship is uh separated from is far
24:06enough from zero that we can distinguish
24:08it from unrelated
24:10samples and what this means is that we
24:12can imagine a setting where we have
24:15access to a secondary data set which
24:17might potentially include relatives of
24:20the individuals who are in the reference
24:22panel and then we could compute these
24:25kinship coefficients uh it's written as
24:27semi kinship because it's a modified
24:28version that's working with half types
24:30instead of typo types so we can compute
24:33these K coefficients between Pairs of
24:35samples between t between these two
24:38sources and then we feed this into our
24:40prisic linking algorithm that's taking
24:42this Matrix as input and then outputs uh
24:44groups of sequences that are likely to
24:46have come from the same individual so
24:48that's the overall outline of the attack
24:53factor and uh we were able to show that
24:57um given a data set with um real
24:59relatives that are known to be included
25:01in the relatives that we were able to
25:03link a substantial portion of the parget
25:06genome um up to third degree relative so
25:10beyond that the signal becomes too weak
25:12so we can't like link too many uh
25:16samples and based on this data we can
25:19extrapolate to get a sense of what
25:21fraction of the genome could be linked
25:23for different proportions of the
25:24reference panel individuals given
25:27different sizes of this uh relative set
25:29that I was referring to um relative to
25:32the overall size of the population that
25:33the reference panel came from so for
25:35example if we're given a relative set
25:37that includes 0.5% of the underlying
25:40population uh we
25:42could expect to link 17% of the genome
25:46for at least 5% of the individuals in
25:48the reference panel so that's the one of
25:50the estimates that we were able to get
25:52and this is a realistic number given
25:54that 05% is comparable to size of the UK
25:57biank for example
26:00so just to summarize this portion
26:03um I've shown you different types of
26:05genomic pry risks we looked at
26:08reidentification attacks membership
26:09phenotype inference data linkage as well
26:12as data
26:13reconstruction and the overall takeaway
26:16is that our understanding is still
26:19rapidly shifting and this needs to
26:21continue to change as we gain access to
26:23different types of data different models
26:25uh more advanced Ai and ml tools as well
26:28as well as other types of data sharing
26:30systems that are emerging and given all
26:33these changes it's important to have
26:34these rigorous approaches to analyze
26:36these risks and assess them uh using new
26:40models and come up with new ways to
26:42address these
26:43risks so on that point in the second
26:46part I'll give you some ideas of some of
26:48the emerging Technical Solutions for
26:49addressing use
26:54risks I think you're all familiar with
26:56these existing policy Frameworks like
26:58common role privacy uh Hippa privacy
27:01role gdpr they Define these Notions that
27:04we all know and love like IRB review
27:06processes informed consent the the idea
27:09of protected health
27:12information and uh there's no question
27:15that these Frameworks provide an
27:16important uh tool for us to safeguard
27:19the use of biomedical data but it's also
27:21important to note some of the inherent
27:23limitations to these regulatory
27:25approaches so the first thing is that
27:27designing and updating these uh
27:29regulations is generally a slow and
27:31long-term process for a good reason um
27:33one example is that genetic data was
27:35classified as protected health
27:37information fora only in 2013 several
27:41years after genetic data have become
27:43more widely available there are also
27:45these key ambiguities in some of the
27:47terms uh that we use in these
27:49regulations right so when we say
27:50deidentified or anonymized data what
27:53does that actually mean and how does it
27:54apply to the data sets that we're
27:56actually dealing with and
27:58it you know as I've demonstrated
28:00hopefully a lot of the bical data types
28:03it's not possible to completely de
28:05identify it there's always going to be
28:06some amount of ident identifying
28:08information sorry uh so we need to draw
28:11a line somewhere when we think about um
28:13laws and regulations and perhaps the
28:16most important uh Pitfall is that these
28:19approaches don't resolve the core
28:21conflict between the need to share data
28:23and the need to uh protect privacy and
28:25this is because uh they're typically
28:27about limiting the scope scope of um
28:29data sharing and usage depending on the
28:32context uh in order to provide PR
28:36privacy so it turns out that uh there's
28:39a collection of techniques from the
28:41computer science literature which are
28:43referred to as privacy enhancing
28:44technologies that can help us resolve
28:46this conflict and at a high level these
28:49provide us with mathematical techniques
28:52uh that can help us use and share
28:54private biomedical data while at the
28:56same time protecting privacy so this is
28:58what I'll tell you about briefly
29:01next um I'm going to highlight five of
29:04these technologies that belong to this
29:05category I'm going to start with homor
29:07for encryption which is a type of
29:09encryption
29:11technique and it's a special form where
29:14the uh the structure of the encryption
29:16allows us to perform operations directly
29:19on the encrypted data without having to
29:21decrypted
29:23First initially the uh these Frameworks
29:28were limited to performing uh limited
29:30types or limited numbers of operations
29:32on the private data but there's been
29:34remarkable progress over the years and
29:35now we have these practical systems that
29:37can help us Implement uh fairly complex
29:40analytic
29:41tests but they do come with a a
29:43computational overhead because now we're
29:45operating with encrypted data and that's
29:47a lot more expensive than working with
29:49non-encrypted
29:51data a slightly different approach as
29:53SEC multiparty computation it has two
29:55acronyms NPC or SMC depending on who you
29:58ask and the core of this framework is
30:03not encryption but we're going to use
30:05this technique called secret sharing and
30:06what it means is for each private number
30:08we um divide it into a set of random
30:11numbers that somehow all add up to the
30:14private number and we're going to split
30:16them up and then distribute it to
30:18multiple parties so they all together
30:20know uh well the shares collectively
30:23encode information about the private
30:25number but if you look at each party's
30:28information they don't know what the
30:29private number is and the idea is that
30:31we can uh design these interactive
30:33protocols where these parties work
30:35together to perform computation on the
30:37underlying secret without revealing
30:39anything throughout this process so at
30:42the end of a protocol like this we get
30:43the the um output of the analysis
30:45results as secret shares as well which
30:47we can combine to reveal the
30:50output so this is not using encryption
30:52so the computational operations tend to
30:55be cheaper than home or encryption but
30:59at the cost of uh greater communication
31:02burden so that's the
31:06tradeoff trusted execution environment
31:08or te it's a hardware approach to secure
31:12computation uh the most well-known
31:13examples are Intel sgx and amdv
31:16Technologies and what they provide as a
31:18way to uh create these isolated
31:20environments in a computer that protects
31:24the data inside it confidential uh keeps
31:27it uh confidential even from the person
31:29who's operating on these
31:32machines and the idea is that we can use
31:34this to send private data into this
31:36environment isolated environment of
31:38secure Enclave do some processing on
31:40them and return the results to the user
31:43uh through encrypted channels so that no
31:45private information is leaked during the
31:48process the key feature of these
31:50Technologies is what's called remote
31:52adastation and this is the use of
31:54cryptographic signatures that actually
31:56allows us to verify that um the The
32:00Enclave environment has been created
32:02faithfully and also the fact that the
32:04program that's running inside it is what
32:06we initially uh assigned to it to
32:11run so in this environment we're working
32:14with non- encrypted data so there's no
32:16computational overhead here or very
32:17minimal computational overhead but this
32:20comes at the cost of uh relaxed privacy
32:22notion since we're relying on a secure
32:24Hardware
32:26component
32:28you've probably heard of Federated
32:30learning this is popularized by Google
32:33several years ago initially designed for
32:36uh this use case of training machine
32:38learning models across millions of
32:40mobile devices the idea is fairly simple
32:42and that you know we have a model and
32:44then we compute these local model
32:46updates on each device aggregate them
32:49aggregate them across devices and apply
32:52the uh the global update to the the
32:54shared model and redistribute them and
32:57we're repeating this process to improve
32:58the accuracy of the overall
33:00model this has um you know immediate
33:04extension to the biomedical setting
33:06where we have a smaller group of
33:07collaborators we can also try to jointly
33:09train machine learning models this
33:12way the focus so far has been mostly on
33:15uh machine learning models or deep
33:16learning models that can be trained with
33:18gradient based optimization so the
33:19relevance for uh you know some of the
33:23genomic analysis tests may be limited
33:27because we're using different types of
33:28statistical models but it has found
33:31really useful application in in medical
33:33image analysis domain where we do need
33:36these deep learning models uh that are
33:38trained with
33:40gradients and this aggregation of local
33:42Updates this can reveal some private
33:45information from each party that's
33:47participating in this workflow so this
33:49needs to be carefully considered and
33:51typically these methods are combined
33:53with other techniques like encryption or
33:56differential privacy which is what I'll
33:57mention in the next slide to increase
33:59the security level of this
34:02solution so uh differential privacy it's
34:05addressing a slightly a different
34:06problem where we're concerned with the
34:08leakage of private information and the
34:10data that is released at the end of the
34:12computation so we can think about chwa
34:14statistics for
34:15example the key concept here is this
34:18notion of neighboring databases so these
34:20are two hypothetical uh databases that
34:23differ in exactly single individual and
34:26the idea is that we want to introduce
34:28some noise into the system or the into
34:30the computation so that only based on
34:33the release data we can't really
34:35distinguish between these two
34:36neighboring databases for any two
34:38neighboring
34:40databases and this privacy guarantee is
34:43typically controlled by this privacy
34:44parameter
34:46Epsilon and in this case values of
34:48Epsilon that are closer to zero means
34:50higher privacy and greater values
34:53um mean lower privacy but at the same
34:56time it allows us to roduce a smaller
34:58amount of noise so the results tend to
35:00be more accurate so this framework has
35:03been applied to Jo statistics release
35:05and other types of biomedical database
35:07queries like looking for corts that
35:09match certain
35:11criteria um but for many of these
35:15existing studies uh they were limited to
35:19uh specific simpler analysis settings uh
35:22for for example for the the case of
35:25chios we're instead of releasing entire
35:28genome wide Vector we can only release
35:30say like the top K most significant hits
35:32or something like that that's been the
35:34typical setting that it's been study and
35:37the the main reason is that the the
35:39actual noise that we need to introduce
35:41to achieve differential privacy
35:42guarantee tends to be very high uh in
35:45high dimensional
35:48cases so these techniques are useful in
35:52you know a range of different settings
35:53in biomedical research I've listed out
35:55some examples here that we've worked on
35:56over the years uh just to mention at a
35:58high level uh we can use these tools to
36:01build um uh softwares that can help
36:05researchers across institutions run
36:08collaborative studies in a secure way
36:10can also build analytic services that
36:12take users private data process them um
36:16and returns statistical insights back to
36:18them all in a privacy preserving way
36:21there's also this private data release
36:22setting similar to what I just mentioned
36:24with uh jwas where we want to release
36:27some privatized version of a bi medical
36:29data or analysis results with some
36:32meaningful guarantees of privacy for the
36:33underlying individuals in the data
36:35set and we're not the only ones working
36:38on this there's a growing community of
36:39researchers developing these tools I
36:41just wanted to mention a couple of
36:43examples of these uh real world
36:45competitions where we bring people uh
36:48together from different communities to
36:49create solutions for these tasks so idh
36:51is a popular one in this domain uh last
36:54year there was also a governmental
36:56collaboration between us and UK on
36:58developing tools based on these
37:00Technologies uh one of the tasks was
37:02creating a model for pandemic
37:05forecasting and uh some of us at bro
37:08entered these competitions and had
37:10winning solutions for for both of
37:13them so now I'll briefly go through some
37:16of our recent work in this domain as
37:19well uh so in our recent work we
37:22introduced this tool called secure
37:23Federated chws or sfj this is a
37:25cryptographic approach to performing Jos
37:28across multiple institutions without
37:30sharing any private information between
37:32them the key technical Insight that we
37:34introduce in this work is that as
37:36opposed to using a single technology um
37:39one of the ones that I mentioned earlier
37:40we can actually use a combination of
37:42them to gain computational speedups
37:44which tend to be quite important in
37:46practice so in this case we're combining
37:48NPC and homeric encryption which allowed
37:52us to design these Federated uh secure
37:55computation systems and what that means
37:58is that instead of encrypting the entire
38:00input data set and sharing it between uh
38:02the parties we can keep all the local
38:04input data sets local and only share
38:06encrypted versions of intermediate
38:08analysis results to carry out the global
38:10uh computation this reduces
38:13communication and also speeds up
38:14computation at the site level because
38:17now we're able to use unencrypted
38:19data and we were able to show that this
38:22um results in an order of magnitude
38:25improvement over the prior art that was
38:27only using MPC
38:30technology across these different uh
38:32data
38:34bases and this scalability Improvement
38:37allowed us to take these tools and then
38:40apply to really large biank scale data
38:42sets here I'm showing you two examples
38:43emerg in UK byy bank so the latter in
38:46our experiment included around 276,000
38:49individuals we're splitting these
38:51resources across uh six or seven
38:53simulated Federated units to demonstrate
38:57the ability to do um perform joint
38:59analysis without sharing private data
39:01and I'm showing you the results of chos
39:03that matches accurately with centralized
39:05analysis um where all the data is pulled
39:08together into a single
39:10location and the overall runtime for
39:13these pipelines uh tend to be around a
39:16few days for uh databases of this scale
39:20so this is not a you know Che cheap
39:23pipeline to run but uh the pract the run
39:26times are still practically feasible and
39:28these can be brought down further by
39:30introducing additional Computing
39:33resources and I want to point out that
39:36the um the experiment that we've run
39:39that we ran on UK bank is actually 2,000
39:41times larger than the previous Benchmark
39:43that we test tested in a previous
39:46study and also want to highlight that uh
39:50in this initial pipeline that we
39:51demonstrated we're looking at PCA based
39:54workflow where we run principal
39:55component analysis to look for these
39:57ancestry karious include as fixed effect
40:00term in the gwas model we're but it's
40:03also common to replace this with a
40:04random effect model uh just known as
40:07linear mix model approach to gas and
40:11this is generally regarded as being more
40:14accurate for jwas I I don't need to tell
40:16you this but it also comes at a greater
40:19computational cost even without any
40:21encryption so we weren't able to do this
40:23using secure computation techniques for
40:25a long time but recently there was this
40:27method that was introduced called
40:29regini and it's doing an efficient
40:32approximation of this random effect term
40:34using a whole gen regression model and
40:36what this allowed us to do is uh
40:38distribute this computational workflow
40:41across the parties using our
40:42cryptographic tools and also obtain a
40:44practical pipeline for uh running LM
40:47based
40:48Jos and um the distributed technique
40:51that I alluded to turned out to be
40:54fairly significant so we were able to uh
40:57remove dependence on the size of the
40:59cohort that goes into these stall
41:01studies as you can see in these runtime
41:04plots that stay nearly um constant
41:07across different datas as sizes we're
41:09able to show that the output of our lmm
41:13analysis workflow closely matches with a
41:15centralized execution of this regini
41:18tool on the pool data set so now we're
41:21able to run these different types of Jos
41:23workflows in a secure and Federated
41:26manner
41:27we had this related work uh more
41:30recently which was about finding genetic
41:33relatives across different data sets
41:35that cannot be pulled together due to
41:37privacy concerns and this is a key
41:40pre-processing pre-processing step in
41:42many studies or Federated databases like
41:45Nomad where we um where the presence of
41:48these relatives could skew the
41:50distribution of the statistics that we
41:52want to compute and release to the
41:54public so we have the cryptographic
41:56tools to be able to comp compute this um
41:58this relatedness coefficient privately
42:01but the the main challenge here is the
42:03fact that we need to perform all pawise
42:06comparisons between these two data
42:08sources and that turns out to be very
42:10expensive um so the key idea that we had
42:12in this work is that we
42:15could hash and bucket the individuals so
42:18that we can perform comparisons only
42:21between individuals that belong to the
42:23same
42:23bucket and the idea is that due to
42:26identity by this
42:27uh sharing between close relatives uh
42:32the close relatives have a higher
42:34probability of getting assign to the
42:35same bucket so this still retains high
42:37accuracy we're able to demonstrate that
42:40on UK by bank and all of us data you can
42:42see that uh this overall
42:45recall for detecting up to third degree
42:48relatives uh is very high it's um 97% or
42:53higher for in all three cases the
42:55Precision is also High higher than 98 %
42:58and some of these false positive
43:00findings are um arguably not false
43:03positives what they are is that they
43:05have a relatedness coefficient that's
43:07really close to threshold and due to
43:08numerical Precision it's Crossing that
43:12boundary and the overall runtime was
43:15also practical less than a day for all
43:16these data sets so where we head it next
43:19I'll try to conclude the talk in the
43:21next few minutes um so I think these
43:24Technologies are maturing enough that
43:27now we're able to see these uh inklings
43:29of practical tools for different genomic
43:32analysis tasks and I think over the next
43:34few years we'll see a lot more of these
43:36uh emerging for different analysis
43:39workflows including things like fine
43:41mapping or predicting disease risks
43:43based on personal genome so we're
43:45actively working on some of these ideas
43:47and I think these would be valuable
43:49tools for people doing genomic analysis
43:51across different sites uh to help with
43:55collaboration we are currently working
43:57on a deployment study where we're trying
43:59to run some of these joint analyses
44:01between million V million veteran
44:04program and all of us research program
44:05so as you know these are two of the
44:07largest fire banks in the US they're
44:09also known for um their strict security
44:12guidelines that prevent the data from
44:14being shared externally so we're trying
44:16to use our tools to create a link
44:18between them to um enable joint
44:21studies last year we launched our SF kit
44:24web server this is Joint work with grow
44:27data Sciences platform and this is a
44:30webbased service that allows people to
44:31run secure Federated tools the
44:34cryptography based ones that I mentioned
44:36uh in a push button kind of way uh and
44:40they can bring their own data set run a
44:42collaborative analysis with their
44:43collaborators uh using the service all
44:45the components that we have here are
44:47open source and this is actually the
44:49service that we're using for the ongoing
44:50pilot
44:51study and currently we're also working
44:54on integrating this uh set of tools into
44:58existing cloud-based analysis platforms
45:00like Tera which has also developed at
45:02Road as well as all of us researcher
45:04workbench uh which is based on Tera so
45:07these are some of the things that are
45:08coming up down the road and lastly I'll
45:10just mention that there's an upcoming
45:11conference that would be of interest to
45:13all of you so recom is happening it's
45:15being organized at MIT how
45:18convenient and there's actually this
45:20inaugural satellite conference on
45:22biomedical data privacy and equity which
45:24is happening one day before the main
45:26conference and we have several talks in
45:28both of these venues if you're
45:30interested please check out the programs
45:32and you can also reach out to me if you
45:33want to discuss more and with that I
45:35want to thank all of the uh former and
45:38Cur current labp members of my group and
45:41my collaborators and mentors uh in
45:44different institutions for all their
45:46input and support without them we
45:48couldn't have done you know a lot of
45:50these
45:50works thank you for your attention I'll
45:53take any questions great thank you so
45:55much Dr Cho I think we have time for one
45:58or two questions I might start them off
46:02um so if we think about how like each of
46:05us we're all required at publication to
46:07deposit all of our transcript omix right
46:09in a public data frame and I try to
46:12think of how we implement this does it
46:15mean that like the gene on our gene
46:17expression
46:18Omnibus data like should it be at the
46:22repository level that these things are
46:23implemented so researchers could still
46:25upload their kind of
46:27naive data frames as we do currently and
46:30then the the implementation of privacy
46:33happens at the repository or do you
46:35think this is actually something that
46:36where even that step is not secure
46:38enough and it should be uh the Privacy
46:40should be implemented prior to upload so
46:43I I can imagine a repository like Goo uh
46:48being run like something like DB Gap
46:50with the genotype data with stricter
46:52access uh control mechanisms I'm not
46:56necessarily proposing that needs to
46:58happen but we do have some concerns
47:00about the way currently transic data is
47:03being shared because they do Le
47:04genotypic information and it's not just
47:06our concern there are other people um
47:09who's been pushing on this as well and I
47:11know that there's an ongoing
47:12conversation with NIH on this topic so
47:15uh yeah there might be some changes in
47:17the future but we're not sure yet I I do
47:20want to make a note that the the risks
47:23that I showed about that data is kind of
47:25in a controlled environment right where
47:28we have this matching ancestry between
47:30like the the target individuals and like
47:32the other individuals in the candidate
47:34pool so there are like other
47:35considerations like that that needs to
47:37go in when we actually think about the
47:39the risk and the Practical
47:42setting Mak
47:44sense great well thank you so much for a
47:48broad sweeping talk that really
47:50highlighted I think so many different
47:51ways in which privacy impacts the
47:53research that we all do every day um and
47:55then please feel free to go grab second
47:57breakfast and then we'll start uh the
47:59next session at
48:079:30

About this transcript

This page contains the full transcript of MPG Primer: Genomic Privacy: Key Issues and Emerging Solutions (2024) by Broad Institute, generated from the public captions YouTube serves with the video. The transcript has 8,064 words across 1,251 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.