Wavefront shaping for deep tissue imaging

Transcript

• 00:00I'm gonna introduce the next
• 00:01speaker, Li Ying Guan, so
• 00:03who graduated from Tsinghua University
• 00:05and got her PhD in
• 00:06statistics from Stanford.
• 00:08And, right after that, she
• 00:09joined, the department of biostatistics
• 00:12here at the Yale School
• 00:13of Public Health.
• 00:15She,
• 00:16her research focuses on
• 00:18robust predictive modeling and
• 00:20uncertainty quantification,
• 00:22model development for complex data,
• 00:24including multiomics data, single cell
• 00:26data,
• 00:27and to apply modern statistical
• 00:29ideas and machine learning approaches
• 00:31for improved data driven immunological
• 00:33discoveries.
• 00:34Welcome.
• 00:37Thank you, John. Thank you,
• 00:39everyone.
• 00:40I'm really excited to share
• 00:42my, one of my recent
• 00:44work, with the audience here
• 00:46and which is also a
• 00:48good example of how I
• 00:49apply the combined, model machine
• 00:51learning,
• 00:52in data driven in logical
• 00:54research.
• 00:58Thanks.
• 01:01So, today, I'm gonna talk
• 01:02about how we utilize longitudinal
• 01:05multi arms data to conduct
• 01:07unsupervised,
• 01:08disease
• 01:09subtype
• 01:10discovery. The data we're using
• 01:12is from this impact cohort.
• 01:14So, in this cohort, like,
• 01:16we generalize the data for
• 01:18more than a thousand hospitalized
• 01:20COVID nineteen patients and the
• 01:21collected longitudinal samples,
• 01:23during their hospitalization
• 01:25on,
• 01:26PBMC and the insert transomics,
• 01:28serum protein, plasma protein tablets,
• 01:31large, body of data, as
• 01:32well as antibody, viral loads,
• 01:34and others.
• 01:36Along with this, deep immunofield
• 01:38typing, we also collected a
• 01:39lot of clinical measures,
• 01:41both from a prior to,
• 01:43infection or during acute infection,
• 01:45as well as a one
• 01:46year follow-up after discharge,
• 01:49on their post acute recovery
• 01:51based on, self reported patient
• 01:53survey.
• 01:55So this rich, data resource
• 01:57actually, is very valuable and
• 01:59allows us to reliably
• 02:01identify or confirm immunosignatures
• 02:04associated with the primary clinical
• 02:06endpoints such as acute disease
• 02:07severity,
• 02:08measuring primarily based on risk
• 02:10for status. For example, like,
• 02:12previously we'll have done a
• 02:13per assay,
• 02:15immunosertial
• 02:15identification
• 02:16as well as an integrated
• 02:18manner saying what are the
• 02:20multi omics programs or cascade
• 02:22associated with, mortality or their
• 02:24severity.
• 02:27One thing is, although this
• 02:29is very meaningful, like, to
• 02:30identify the immune correlates with
• 02:32a particular,
• 02:33clinical endpoint. We know that
• 02:34patient are very heterogeneous.
• 02:36If we look at their
• 02:37clinical characterization,
• 02:40in panel a, it is
• 02:41like the primary clinical endpoint
• 02:42will have been considered in
• 02:44a lot of our previous,
• 02:45analysis
• 02:46is called something called a
• 02:47clinical tragedy group, which is,
• 02:51disease group,
• 02:52defined by our clinical team
• 02:54based on longitudinal response data
• 02:56still in hospitalization
• 02:57from t g one to
• 02:58t g five in indicate
• 02:59increased severity.
• 03:01This is very convenient for
• 03:02us to do our analysis,
• 03:03but as you can see,
• 03:05a patient can be characterized
• 03:06by a lot of different
• 03:07aspects.
• 03:08If we you look at
• 03:09the panel say, here is
• 03:10a very simple example where
• 03:11we show the prior, like
• 03:14probabilities
• 03:15and other, demographic information.
• 03:17I highlight here, like, there's
• 03:18a chronic,
• 03:19lung disease role here. You
• 03:21can say even, the most
• 03:22moderate group, there is some
• 03:24portion of,
• 03:26chronic lung disease. And among
• 03:28the two very severe group
• 03:29like t g five and
• 03:30t g four, t g
• 03:30five is mortality, t g
• 03:32four is very, very poor
• 03:33recovery during acute infection. It
• 03:35is not necessarily true that,
• 03:36like, all the two most
• 03:38severe group actually have less,
• 03:40common ability compared to others.
• 03:42Although the overall,
• 03:43there is association between t
• 03:45g group and the chronic
• 03:46lung disease.
• 03:48The same thing is true
• 03:48for complications, which I didn't
• 03:50show here. And also we
• 03:52can if we want to
• 03:53look beyond the acute infection,
• 03:55one interesting thing, like, we
• 03:56you know, I want the
• 03:57previous work,
• 03:59observed is actually there's a
• 04:00very little,
• 04:01almost no association between the
• 04:03acute infection
• 04:05severity and the, post acute
• 04:07recovery, which is kind of,
• 04:09very, very surprising thing we
• 04:11we saw.
• 04:14The same thing happens for
• 04:15the immune response.
• 04:17Not only the clinical characterization
• 04:20can be quite high hygienous
• 04:21and it cannot be not
• 04:22be easily captured by one
• 04:24single clinical parameter.
• 04:26The same thing happens for
• 04:27their immune response. Here is
• 04:29a severity factor, which is
• 04:30a one highlighted factor in,
• 04:32the integrated, integrated analysis manual
• 04:34script.
• 04:35As you can see, both
• 04:36at the base of one,
• 04:37which is panel a baseline
• 04:39visit and And over time,
• 04:40there is it is indeed
• 04:42true that this severity factor
• 04:43captures acute severity
• 04:45very well. There is a
• 04:46very clear increase in trend
• 04:48from t g one to
• 04:49t g five. And over
• 04:50time, the we also see
• 04:52a divergence between t g
• 04:53five and t g five.
• 04:55But again, you can say
• 04:57there is a lot of
• 04:58variability
• 04:59here. As John mentioned, maybe
• 05:00variability is a hallmark
• 05:02of in your response.
• 05:04So,
• 05:04motivated by this observation, so
• 05:06we decided to take a
• 05:08different approach. Can we just
• 05:08directly understand the heterogeneity among
• 05:09the host immune response characterized
• 05:10by this super high dimension
• 05:10of our comprehensive molecular isis
• 05:12that is not self reported
• 05:14from the,
• 05:16lab measures?
• 05:21And this can be very
• 05:22useful because this if we
• 05:24can done this correct, like,
• 05:26if the result is, good,
• 05:28then we can actually link
• 05:30the web the immune response
• 05:32heterogeneity
• 05:33to all kinds
• 05:34clinical measures in a very
• 05:35unified framework.
• 05:37And also it also offers,
• 05:38like, not only offers more
• 05:40comprehensive
• 05:41understanding on the host immune,
• 05:43heterogeneity, but also potentially offer
• 05:45more insights into personal and
• 05:47the treatment because different, immune
• 05:48response may,
• 05:50indicate a different type of
• 05:51treatment.
• 05:56When we go so the
• 05:57first step we're going to
• 05:58do is the major analysis
• 06:00step is we want to
• 06:01identify,
• 06:02immunophenone that,
• 06:04these are subtypes based on
• 06:05this longitudinal molecular high dimensional
• 06:08profiles.
• 06:08There are two challenges here.
• 06:10First,
• 06:11there is a lot of
• 06:13we have,
• 06:14several high dimensional assays from
• 06:16a transtomics, proteomics,
• 06:18and each of them has
• 06:19many, many features. So it
• 06:21is very important for us
• 06:22to be able to, identify
• 06:25the most you most meaningful
• 06:26and coherent
• 06:28information from the high dimensional
• 06:29assays.
• 06:30The second challenge is this
• 06:31longitudinal data. Of course, like
• 06:33we can do a subtyping
• 06:35for with one sample. Of
• 06:37course, we can do that,
• 06:38but that will be a
• 06:39miss a lost information when
• 06:40we're trying to characterize each
• 06:42participant and their similarity.
• 06:44We want to use all
• 06:45the visits when determining the
• 06:47similarity between two,
• 06:49two participants here.
• 06:51And, to do this, like,
• 06:52with with some dimension reduction
• 06:54technique to identify the multi
• 06:56omics factors capturing the co
• 06:58varying patterns across multiple omics.
• 07:01So this help us to
• 07:02reduce the dimension and then
• 07:03focus on the most important,
• 07:05multi omics factors.
• 07:07The second issue is the
• 07:09time series aspect.
• 07:11And this is not only
• 07:12a time series data. This
• 07:13is a time series data
• 07:14with very strong messiness, very
• 07:16high messiness. As you can
• 07:18see here, wide vessel one
• 07:20among a multiple, like,
• 07:23one thousand one hundred forty
• 07:25eight participants out of the
• 07:26in total, one hundred
• 07:27one thousand one hundred fifty
• 07:29two participants have vessel one
• 07:30samples. As a way to
• 07:32increase, like, the number of
• 07:33samples available just sharply.
• 07:36And, what's more troublesome here
• 07:38is this
• 07:40this missusness is not just
• 07:41random.
• 07:42This is actually severely,
• 07:44confounded by their, actually, patient
• 07:48status. If you,
• 07:50will recur if the patient
• 07:51recovered very quickly and then
• 07:53discharged, then you may not
• 07:54have sample measure. If the
• 07:56patient, actually were have very
• 07:58poor recovery in diet. Right?
• 07:59So we won't have any
• 08:00measurements.
• 08:02Due to this kind of
• 08:03bias and missingness and it
• 08:04is a high proportion missingness,
• 08:07we
• 08:08we don't really want to
• 08:09do, say, like, very,
• 08:11intensive data imputation
• 08:13because who knows what the
• 08:14impute quality will be like.
• 08:16So to do this, we
• 08:17conduct we we actually did
• 08:19some, imputation free, longitudinal,
• 08:22subtyping.
• 08:23We are we actually,
• 08:26get some kind of pairwise
• 08:27distance between two patients based
• 08:29on their available samples only
• 08:31across the Bay and then
• 08:33conducted some,
• 08:34class run based on the
• 08:35pairwise distance.
• 08:38And this enabled us to
• 08:40identify,
• 08:41six subtypes
• 08:42based on using some automated
• 08:45decision criteria.
• 08:46And the subtype one to
• 08:48subtype f.
• 08:50And, panel a shows,
• 08:52like, how does each point
• 08:54means a participant
• 08:55and how the participant looks
• 08:57like when we project it,
• 08:58each each participant into this
• 09:00two dimensional space using,
• 09:02multidimensional,
• 09:03scaling.
• 09:05And,
• 09:06and so we can say,
• 09:08although there's some overlapping, but
• 09:10it is very clear that
• 09:11the different subtypes,
• 09:13they they occupy very different
• 09:14space, in this two dimensional,
• 09:17space here.
• 09:18And then the next thing
• 09:19we want to check is,
• 09:22sure. So this participant, they
• 09:23have a very different, host
• 09:25response.
• 09:26But
• 09:27this host response difference actually
• 09:29meaningful and being reflected in
• 09:31various clinical measures.
• 09:33So the first thing we
• 09:34check is our primary clinical
• 09:36endpoints,
• 09:37used in previous study, which
• 09:39is a clinical trial group,
• 09:41defined,
• 09:42from the mixed mixed modeling
• 09:45longitudinal mixed modeling using the
• 09:46respiratory status over time during
• 09:48hospitalization.
• 09:49And as I mentioned, from
• 09:50t g one to t
• 09:51g five, it's a increase
• 09:53in severity. With the t
• 09:54g one two three, they
• 09:55tend to have, like, better,
• 09:57like, quick fast recovery. T
• 09:58g five, like, people all
• 10:00die here. And then t
• 10:01g four is a group
• 10:03where,
• 10:04they didn't die in the
• 10:05acute infection, but maybe that
• 10:07later.
• 10:08And and the recoveries are
• 10:09quite poor. Like, they usually
• 10:11tend to have prolonged hospital
• 10:13stay.
• 10:14And, panel in panel b,
• 10:16you can say when we
• 10:17plot the distribution
• 10:19of different TG group here,
• 10:21we can say it's very
• 10:22clear that there is,
• 10:24different subtypes that are enriched,
• 10:27have differential enrichment in the,
• 10:29t g group.
• 10:31And you can say that
• 10:32the ABC subtype ABC, they
• 10:33tend to be more enriched.
• 10:35They have, like, very few
• 10:37mortality
• 10:38and then primarily consists of,
• 10:40t g one, two, three.
• 10:41And t g one and
• 10:42five, they are very, very,
• 10:44very severe. They have a
• 10:45lot of mortality and primarily
• 10:47consider t g four and
• 10:48t g five.
• 10:49And t, and then some
• 10:51type d.
• 10:52It it's kind of more
• 10:53on the one hand, in
• 10:54the modeling towards the, less
• 10:56severe side. On the other
• 10:57hand, it also has a
• 10:58decent amount of mortality. So
• 11:00we consider this one be
• 11:01a mixed group.
• 11:03And later,
• 11:04so so the same message
• 11:06can be confirmed if we
• 11:07look at their,
• 11:08survival curve, beyond the acute
• 11:11acute infection,
• 11:13we can say that,
• 11:14subtype f is very enriched
• 11:16in the mortality, like the
• 11:18head ratio is, very high
• 11:20and then followed by subtype
• 11:22e. Subtype d, although it
• 11:23has a lot of,
• 11:25less severe, participants,
• 11:27it also has,
• 11:30appear to have some kind
• 11:31of enrichment in the mortality,
• 11:32and ABC is also
• 11:34they also have less mortality
• 11:35here.
• 11:37And, although subtype d can
• 11:39be very interesting to to
• 11:40to be investigated, it's a
• 11:42mix and it's kind of
• 11:43interesting. But we decided to
• 11:44focus on ABC and EF
• 11:46because they have more participants.
• 11:48And it's easier for us
• 11:49to do a later analysis
• 11:51of when we have going
• 11:52to have more recent data
• 11:53later.
• 11:56So, just to summary, we
• 11:57decided to we we identified
• 12:00four
• 12:01five major,
• 12:03molecular subtypes
• 12:05with three being severe, sub
• 12:06a, sub b, sub c,
• 12:08and two being very critical,
• 12:10sub e
• 12:12and
• 12:13sub
• 12:14f.
• 12:15And,
• 12:16these subtypes not only,
• 12:18have strong associations with the
• 12:19primary clinical points as we
• 12:21mentioned. They also showed a
• 12:22very strong association patterns between,
• 12:26with, other clinical characterizations including
• 12:29demographics, comorbidities,
• 12:30and the complications.
• 12:32For example, you can say,
• 12:36sub e is actually the
• 12:37group has the oldest age
• 12:38even though sub f is
• 12:40more severe seems to be.
• 12:42And then sub a and
• 12:43b, they they are tend
• 12:44to be younger.
• 12:45They also show some difference
• 12:47in the ethnic ethnicity distribution.
• 12:50And regarding the sex, so
• 12:51we can say sub c
• 12:52has a very high enrichment
• 12:54in female and then sub
• 12:56a and sub c has
• 12:57less females.
• 12:58There are many comorbidities and
• 12:59any complications. Just overall, I
• 13:01do not dig into one
• 13:03by one. But you can
• 13:04say, although sub c is,
• 13:06on a more a more,
• 13:07like, severe side, not critical
• 13:09side, sub c and sub
• 13:10e and f, they both
• 13:11have more prior comorbidities,
• 13:14because it's in their, population.
• 13:16And when we look at,
• 13:17sub a and sub b,
• 13:18they are, they tend to
• 13:19be more healthy regarding the
• 13:21prior conditions. When we look
• 13:23at the complications,
• 13:24like, beyond risk for status,
• 13:26status, we can say sub
• 13:27e and sub f especially
• 13:28sub f has strong enrichment
• 13:30in different complications.
• 13:32And, we,
• 13:34we decide to from now
• 13:36on, let's focus more
• 13:38on the comparison between among
• 13:40APC and among EF because,
• 13:42there's a tons of work
• 13:43separating, like, severe versus the
• 13:45critical. But the characterization
• 13:47between within, participants,
• 13:50expecting similar severity levels is
• 13:52actually less, available out there.
• 13:55So when we look at
• 13:56the for example, when we
• 13:57look at the complications
• 13:58and let's say comparing the
• 14:00comparison between EF and the
• 14:01comparison between among ABC.
• 14:04We we can say that
• 14:05it is indeed like there's
• 14:06a lot of statistical significance,
• 14:08regarding the elevated complication in
• 14:11subtype f compared to subtype
• 14:12e. And even among ABC,
• 14:15all of the pattern is
• 14:16less,
• 14:18less obvious because they tend
• 14:19to have less complication overall.
• 14:21We still see something very
• 14:22interesting. For example, we say,
• 14:24overall, like, subtype c seem
• 14:26to, have
• 14:28a slightly higher cardiac con
• 14:30complications, especially the CHF.
• 14:32And it also has, higher
• 14:34renal complications.
• 14:36Although subtype c is also
• 14:38the one that has the
• 14:39least amount of, pulmonary, the
• 14:41lung related complications. So there
• 14:42is some difference between other
• 14:44organs and the lung here,
• 14:46for the subtype c.
• 14:52What's more interesting is we
• 14:53mentioned when we compare the
• 14:55acute infection severity and the
• 14:57PAS, our previous work actually
• 14:59identified almost none association at
• 15:02all, not even, like, significant.
• 15:05Weak in effect size.
• 15:07What's interesting is when we,
• 15:09check the task,
• 15:11self reported task and the
• 15:13subtype we define, there is
• 15:14actually
• 15:15quite a strong enrichment in
• 15:17certain sense. As you can
• 15:18see for, when we're comparing,
• 15:20like, each subtype against others,
• 15:22we overall,
• 15:24there is a strong, there's
• 15:25this this is statistical significance,
• 15:29regarding, like, subtype f, subtype
• 15:31c, and subtype a,
• 15:32which have, like, very different
• 15:34distribution compared to the,
• 15:36global distribution of the past.
• 15:39So the,
• 15:41let's
• 15:42what are the past category
• 15:43here? So minimum past here
• 15:45can be viewed as,
• 15:47the recovered convalescent.
• 15:49And then the, physical cognitive
• 15:51multiple, just three kinds of
• 15:52different past characterization by clinical
• 15:54team. And the multiple means
• 15:56that it has all kinds
• 15:57of deficits.
• 15:59When we look at the
• 16:00distribution, we can say that
• 16:01the the subtype a is
• 16:02actually a subtype that has,
• 16:05much more minimal deficits and
• 16:07the less other,
• 16:08task groups.
• 16:09And then, subtype c is
• 16:11the one that is more
• 16:13enriched in,
• 16:14like, long COVID and the
• 16:16less has less, like, comes
• 16:17convalescent,
• 16:19but
• 16:20and driven mostly by the
• 16:21physical and a little bit
• 16:22by the multiple deficit.
• 16:25Subtype f is also a
• 16:26a subtype that is, has
• 16:28more past compared to others
• 16:30and particularly compared to subtype
• 16:31e. You can we can
• 16:33say it has, like, much
• 16:34fewer, minimal deficit, mainly driven
• 16:36by the enrichment of cognitive
• 16:38deficit.
• 16:41And when we,
• 16:43so we also check whether
• 16:45utilizing the molecular profile
• 16:48can improve over a pure,
• 16:50clinical model that utilize
• 16:52age, sex, which, can be
• 16:53potentially be related to past,
• 16:55prediction,
• 16:56and comability probabilities
• 16:59as well as a viral
• 17:00load and antibody because in
• 17:02the previous,
• 17:03study, like, we we identified
• 17:05this as an early correlative
• 17:06PASC. So you can say
• 17:08we're a little bit, not
• 17:09one hundred percent fair to
• 17:11other models. The clinical model
• 17:13has been selecting the most
• 17:15important predictors from from our
• 17:16previous study. But, anyway, so,
• 17:19that that's okay because if
• 17:21we can improve this model,
• 17:22it's the improvement is real.
• 17:24So, we we we consider
• 17:27two models. One is a
• 17:28clinical model
• 17:29plus only one additional feature,
• 17:31just our subtype.
• 17:33The other is a clinical
• 17:34model plus subtype and plus
• 17:35the multi omics features
• 17:37because our subtype is unsupervised
• 17:39and
• 17:40okay,
• 17:41and then we want to
• 17:42say whether we miss anything.
• 17:43So the result is very,
• 17:45that that is a very
• 17:46challenging task. So but we
• 17:47can see that there's improvement
• 17:49comparing clinical plus subtype versus
• 17:52clinical, clinical plus subtype, and
• 17:54the plus means additional,
• 17:55other factors. They both improve
• 17:57clinical and the, the statistical
• 17:59significance of the real. So
• 18:01this is the evaluative test
• 18:02data that haven't been touched
• 18:04on the training.
• 18:05And,
• 18:07so the model we use
• 18:08is,
• 18:09is the kernel SVM, and
• 18:11then we can also use
• 18:12in the recently, like, popular
• 18:14measure, like, the shape value
• 18:15to identify what are the
• 18:17important features for our model
• 18:18prediction. We can say there
• 18:20are some clinical measures, but,
• 18:21like, it's the subtype is
• 18:22one of the top features
• 18:23here that is,
• 18:26very important for our model
• 18:27prediction in the last model
• 18:29including everything. And when we
• 18:31plot the percent of,
• 18:34shift value on the test
• 18:35sample only, and we can
• 18:37say that indeed,
• 18:39their improvement their contribution from
• 18:41this the subtype and some
• 18:42other top top features,
• 18:44they actually,
• 18:46strongly associated with the task,
• 18:49on all the sample evaluation.
• 18:52Okay.
• 18:53So
• 18:54to summarize, like, we identify
• 18:56the molecular subtypes that are
• 18:57not only separate severe and
• 18:59critical, but also
• 19:00associated with a bunch of
• 19:02different, clinical calculations
• 19:04as well as long as
• 19:05long COVID.
• 19:08I want I don't have
• 19:09time to tell you the
• 19:10details about what the immune
• 19:11implication is, but there are
• 19:13actually many interesting things we
• 19:14found. And we and, again,
• 19:16I you said the last
• 19:18two minutes. I won't talk
• 19:19about difference between APs and
• 19:20EF because it's very clear.
• 19:22The only systematic inflate inflammation,
• 19:25is very hallmark for separating
• 19:27critical against severe. But there
• 19:29are also a lot of
• 19:30interesting things, separating ABC and
• 19:33the second EF. Particularly,
• 19:34in ABC, when we look
• 19:36at the serology
• 19:37and the set off and
• 19:38the, like, different, cytokine chemokines,
• 19:41we we saw, like, subtype
• 19:43ABC, they actually have strong
• 19:44shape to the in their
• 19:45antiviral immune response, particularly sub
• 19:48a has much stronger humoral
• 19:50response compared to sub c
• 19:51where sub c has very
• 19:52early strong T cell response
• 19:54as T cell cytotoxic response,
• 19:56but it's a hormone response
• 19:57just somehow do not work
• 19:59very well. And it also
• 20:00has a very slow var
• 20:01varial clearance.
• 20:02And the difference between between
• 20:04sub e and sub f
• 20:06is also is, you know,
• 20:07is very different story. They
• 20:08are both in a very
• 20:09hyperinflammation
• 20:10state, but the sub e
• 20:11somehow tend to be adapting
• 20:13to it better.
• 20:14For example, where they are
• 20:16both, in a hyperinflammation state
• 20:17with a lot of calculation,
• 20:20regulation,
• 20:21calculation complement,
• 20:23protein.
• 20:24Sub e is actually the
• 20:26one that also has, associating
• 20:28alongside with, anti coagulation components.
• 20:31But the sub eigen just
• 20:32given up very strong coagulation
• 20:34complement, but the anti coagulation
• 20:36is just silent.
• 20:37I don't know why that
• 20:38that's what happening. And, also,
• 20:40sub although, you know, in
• 20:41during this hyper inflammation state,
• 20:43sub e is, has,
• 20:45less ox oxidative stress,
• 20:47indicated by both its complications,
• 20:49its metabolites,
• 20:51and protein levels. And the
• 20:52sub f is has, like,
• 20:54higher oxidative stress and with
• 20:56a higher n,
• 20:58n six to n three,
• 20:59like,
• 21:01on certain fatty acid ratios,
• 21:03anemia, and all kinds of
• 21:05things. So,
• 21:07this this is very essentially,
• 21:08very we found this very
• 21:10interesting. Just this kind of
• 21:11different in immune responses that
• 21:13can potentially
• 21:14be explaining why they're so
• 21:15different in many different aspects
• 21:17and even in long COVID.
• 21:19So lastly, so this is
• 21:19the work. It's a teamwork.
• 21:21It's part of the impact
• 21:22project, and we also have,
• 21:23many PIs from the EL
• 21:25sites as I listed here
• 21:26and other people. And also,
• 21:27like, this, here are the
• 21:29six lead authors, student authors,
• 21:31and the collaborators
• 21:32that, without whose help this
• 21:34project won't be possible. Thank
• 21:41you. Thank you, Liying.
• 21:43Maybe time for one
• 21:45any pressing question?
• 21:49Not actually one quick one.
• 21:51I see that the percent
• 21:53of PACE,
• 21:55it's
• 21:56it's highest in CNF. Right?
• 21:58So and then you have
• 21:59a gradient sort of going
• 22:00up from a, b, c
• 22:01to e, f. And then
• 22:03within each one of those
• 22:04two sub subgroups, you know,
• 22:06is it the same gradient,
• 22:07but just more intense in
• 22:08the e, f group?
• 22:10I didn't check. That's a
• 22:10good question. I didn't exactly
• 22:12check
• 22:12the ratio in
• 22:14the decrease.
• 22:17Yeah. But we can see.
• 22:18But there is a decrease.
• 22:19Yeah. Okay. Thank you. Thank
• 22:21you. Okay.
• 22:22Thanks again.