>>Hi, I’m Whitney Espich, the CEO of MIT
Alumni Association. And I hope you enjoy this digital production created for alumni and
friends like you. Aviva: Hi, everybody.
Let’s get started. Welcome to the MIT faculty
forum. My name is Aviva Hope Rutkin,
and I serve as the moderator for today’s presentation, Machine Learning Insights into
Antibiotic Lethality. As a reminder, we welcome your
questions during this chat. If you are an alumni joining via
A Zoom, you can use the toolbar.
If you are watching it on YouTube, you can use the
comments stream, and we encourage you to tweak using the
#MITBetterWorld. We will do our best to get to as
many questions as we can. I am delighted to introduce our
featured presenters today. Jason Yang, a research scientist
in the lab of James Collins, and Allison Lopatkin , the Infectious
Disease and Microbiome group by the Institute as a member of Dr.
Collins’ lab. Jason received his degree in a
double major biomedical
engineering from the University of Virginia.
As a post-doctorate research scientist at the Collins lab, Gay has been developing systems
approaches for understanding mechanisms of Antibiotic
Lethality. He is the recipient of the K99/R00 Pathway to Independence
Award from NIGMS and is interested in developing network
modeling and machine-learning-based
approaches for advancing the practice of precision medicine.
Allison completed her PhD training in biomedical
engineering at Duke University. Her bachelor’s degree in applied
mathematics at the University of Rochester.
Allison’s work uses quantitative methods to study the spread of
antibiotic resistance within bacterial populations.
Her current research interests include modeling the
evolutionary dynamics of bacterial communities,
leveraging horizontal gene transfer to target pathogens,
and examining mechanisms of antibiotic lethality with a
focus on the role of metabolism. She will be joining Barnard
College of Columbia University next year as an assistant
professor of computational biology. And with that, I will turn
things over to Allison and Jason. Jason: All right, so, thanks,
everybody, for turning in. I am Jason, and this is Allison,
and we’re both postdoctoral researchers as the Collins
Leavitt MIT. — lab at MIT. Allison: OK, great.
So I am going to start with some slides to make sure we’re all on
the same page. The discovery and implementation, really the
revolution of modern-day medicine.
After the introduction we started to see the first
documented cases of Helen — penicillin-resistant organisms, and
unfortunately, that trend HAS has
seemed to retain, where every single antibiotic we have
brought to market has seen a resistance occurs shortly
thereafter. And so the rise in
antibiotic-resistant pathogens is a global growing threat.
The way we can bring it to market and significantly slower.
It has actually been predicted that if we do not do something
drastic about it, mortality from infectious diseases will
skyrocket, potentially reaching 10 million people by the year
2050, becoming the leading global cause of death.
So researchers around the world are researching and really
trying hard to find some sort of solution to this.
In order to do that, we really need to take a few steps
backwards and address or learn more about exactly how
antibiotics function so that we can better implement and design
things in the future that look better. So what do we know about
antibiotics? We know that antibiotics target
key bacterial processes that are critical for growth, so things
like inhibiting a cell wall, inhibiting the services,
prevents them from occurring, and eventually, the cell itself
dies. It is knows price, then, that
for a very long time, it has been assumed and shown that the
faster event interior cell shows, the better the antibiotic
works. So there is a correlation
between the growth rate and the killing rate here that has been
shown to be true, but more recently, this has been work now
validated by many other groups, which is shown this is not
necessarily the whole story, and thus antibiotics hit their
primary target, and that is a huge part of how they function.
But following that, the bacteria also tries to fight back,
because it has a toxic unknown organism that introduces
itself, and it is a long, metabolic, demanding process
that results in toxic broad problems — toxic byproducts.
This note raises able to open questions that our group has
been exploring in-depth, because
we were not necessarily aware of all this until recently.
One of the first questions that becomes quite apparent is
because growth and metabolism are so correlated, and perhaps
this widely established convention of growth rate as a
function of antibiotic filling, it is a proxy perhaps of the
underlying metabolic rate. This is one example of something
we have been investigating, so this question makes metabolism a
better predictor for antibiotic quality. It is something that is
currently impressed right now. What we did as we were able to
establish a very wide parameter space of all these different for
mental conditions, which ultimately amounted to many
different growth rates and some of a couple, some were not, so
we could tease apart growth rate and metabolic state and measure how effective the antibiotic was
individually. So you can see on the left-hand
side across all of these many provisions, when we have the growth rate, the top hand
showing you the level, you can see very clearly on the
left-hand side that there is effectively no correlation
between growth rate and survival winter coupled, and yet on — w hen decoupled, and yet on the
right-hand side, there is a clear transition between low
metabolic state , so the high
metabolic state interlinear — and linear correlation.
I am going to give this to Jason. Jason: Thanks, Allison.
One of the things Allison and I have been interested in is to
better understand how does metabolism specifically
contribute to how antibiotics kill. As many of you guys may
appreciate from thanks we may have been learning in high
school chemistry or high school biology, metabolism is very
complex.For every cell in every organism,
there are several different processes necessary for any sales to replicate, and in most
cases, we talk about metabolism, you oftentimes are talking about
essential metabolism, so the TC A cycle or the Krebs cycle, and
that is shown in blue, the giant circle in the middle.
As you can see from this diagram, many processes are
important for cell division, and these all share and cross up
with each other, with from a subsistence perspective, makes
it difficult to tackle and investigate.
So we are challenged with this
question am asking ourselves — how can we better understand
biology in this complicated context?
And I would say today there is really predominantly there has
been two main approaches biologists have used
recently, one of them the more informatics-driven approach.
We have technology, large experiments, and these include
things like chemical screens or genetic strings — scree ns, and
affected on the left-hand side commoditize we perform a screen,
and then subjectively select the
criteria and will determine which of those were important.
Some of those criteria may be things like a certain P value.
What we hope to do in the approach is to take those hits,
those pieces that we think are important for biology, to drive
some sort of downstream analyses that we hope can test the
hypothesis. One of the challenges to the
approach is often times, the number hits you might get from
any sort of screening for in the 1% to 3% range.
Consequence really, that is often too few to release to
physically the downstream analysis you want to do later. So often what comes out is not
actionable. Most of our viewers are aware
today there is excitement around the class of approaches that are
more data-driven, so things like Machine Learning-based
approaches like deep learning, and another sort of — and other
sort of new techniques. We may begin with some set of biological data, maybe, for
instance, biological features that represent the data from our
initial screens. We have with those corresponding
data, and what we want to do is use the computer to determine
some sort of transfer function linking the initial biological
feature data to our data, and
while computers are really good at now defining really strong associations, one of our final
barriers here is many times those associations, those
transfer functions are OK. We often call them black-box
approaches, because the members or the test functions themselves
cannot be mapped in biology, so
we cannot necessarily interpret how those associations are being
made. To address these challenges into
kind of further, we recently developed what we call white box
Machine Learning, whereas in certain of — instead of using
it to train a predictive model based solely on the raw data, what we are going to do first is
use network models to interpret the larger biological feature
data into a representative state that already has some sort of
mechanistic understanding, and then what we do is we use those
interpreted and simulated network states as the computer
is selecting different features that have already been
interpreted to explain to us how it is that, in this case,
antibiotics are working. And more specifically, what we
do in this project come up was also recently published, what we
did is we first performed on the left-hand side a biochemical
screen, where we took E. coli cells and treated them with antibiotics, and this gave us in
a variable input/output information between the system
and the antibody. Then we used experimental to
simulate and give us a large matrix of simulated network
space, and we took the combination of this empirical
antibiotic efficacy data and our simulated metabolic data network
space together as our Machine Learning of what ask the
computer to tell us what reactions can prescribe or
predict antibiotic killing, what pathways or what processes are
they involved in? And this to us became — we view
this as an agent that enables us to very rapidly identify
mechanisms that can be actionable and validated.
A high level of this approach is that overall addresses the two
fundamental presentations addressed on the previous site.
The informatics-driven approach, because we are not flesh holding
by some sort of subjective criteria, we actually get to use
all the data that we measure, including our negative data.
That allows us to enrich and expand the content available in
our raw data, our raw s creen. What we are doing here is using the network models as a
representation of our existing knowledge bases and basically
giving our analysis additional prior knowledge by which we can
interpret the more broad data we have.
So together, we recited that this can help our approach, to
give you a sense of how this actually look like in our study,
we performed a screen reduction with different metabolites with
three different antibiotics — antibodies on the left-hand side
of target very different actual self Philly — cell physiology. We can elicit different types of
response from the different metabolites, you can see from
the different colors and the different columns, that the
colors are different from each other.
Hopefully one of the things you appreciate of the
similar metabolites are able to have similar reactions come as
you can see that by looking across rows, how some rows have
similar colors, although in a different magnitude of shading.
This informs us that perhaps in addition to the very canonical
or conventional understanding about and by Terry up — about
antibacterials kidding their targets, what makes it —
hitting their targets, what makes an antibiotic successful
or not. We first took these measures, to
simulate the metabolic’s corresponding, and we used
Machine Learning to explain the data that we have here.
What emerged from this analysis was several different pathways,
you know, computers are predicted to be commonly
involved in how antibiotics kill across the three different
classes of antibiotics. One of the things really
exciting as if you look in the top cluster, the approach is
completely unbiased perspective, it was able to return to us what
we are writing about, several aspects of metabolism in the
cycle, but what we were really excited about was the
opportunities for the approaches to lead us a new virology —
lead us in new biology. You can see the synthesis that
blocks the pathways. Two things interesting about
this is number one, these were not previously implicated, and
number two, you can see from the shading that because we have the
simulations available to us, we were able to use some of the
data to compute qualitative pathway scores, and you can see
from the shading that two of our
antivirus in this case had a different shading of anti-myosin, so we had a
different resolution to the biological phenomenon.
So the data for these analysis, we started to understand maybe
what could explain the biology happening in the middle block of
pathways, and one of the things we found commonly when we opened
up the boxes they looked at the
different reactions and reaction coefficients in Machine Learning
was that it appears the computer — many of these affects were
explained by biosynthesis or differences in a very specific
synthesis. So this led us to experiment
back in the lead to see if indeed biosynthesis was all
relative to antibiotic healing. What we can see is that if you
look at the two panels on the left, in comparison to
gentamicin on the left, that is a control case, and in many of
these, we saw protection or decreases in ample myosin but an
increase in gentamicin. We were very competent that the
Machine Learning was leading us to new biology that we could
test indirectly violate. — and directly. The new metabolism is very much well studied and in the two
classes, these two classes are directly with each other,
sharing cell shape and possessing data feedback.
In a previous study, characterizing several of these responses to antibiotic stress,
what we can see is in red that they have observed it is
depleted, so this led us to hypothesis that may be the
reason that purine biosynthesis is important is for this to deal
with cell nuclear type stress, when we supplement these
nucleotides back, we can supplement your we did an
experiment, and that is exactly what we see, you can see in the
red lion, that these equalized cells across antibiotics, in
contrast, we use the other type of nucleotide, pyrimidines, this
actually increase the killing, bio tools that allow us to reach
opposite effects. So this alert us to ask
questions about why it might be at all important to understand
how antibiotic killing is occurring, and we were showing
you earlier in relation to achieving different metabolic
state’s as important. If you go back and look at the
literature from the 1990’s, one of things interesting is when we
do and inventory, all of the metabolic cross it takes to make
these cells, what you find is a new type of metabolism that is
very inexpensive. In fact, it costs just as much
energy for a cell as it does for new amino acids, which if you
think about it, it is important for almost all cell functions.
So we challenge ourselves to wonder if perhaps maybe the
reason why it had something to do with energy.
So because we have the model simulations, we could go back to
the simulations and see what the computer might be predicting
could be happening in cases where we have it as a lever that
protected themselves against antibiotic killing, and what
could be happening as a pyramid Dean — Pyrimidine lever. If we supplement, we get totally
opposite biosynthesis activity, for example, supplemented with
the Purine. The next model shows it uses less ATP.
In the third panel, the model further predicts that the sales
decreased the amount of central activity to make ATP. And finally come in the panel on
the right, we can actually measure this by looking at
changes in oxygen consumption and respiration, so this
directly let us to a biological hypothesis we can test, that we
added at a mean back to cells, and it could decrease what is
necessary to make energy. The fact when we did these
experience, that is what we seek him as a when we look at the
black line, all of these cases, on the left-hand side is
control, you see the Gentamicin, antibiotic stress, it is
increasing, metabolic activity following antibody treatment.
When we add Adenine in red, it is inhibited and decreased,
meaning that supplementation with Purine is able to block
words of the cells of the need to create more energy, because
the overall demand for energy is less.
And conversely, the opposite is true in blue when we add
Pyrimidine. When we do other experiments,
for instance, measuring different metabolites, different
energies can use it, but what we see on the left-hand panel is
when we use metrics to ascribe overall ATP energy balance, the
addition of Adenine decreases cells to make ATP, and other metrics, and the addition of
Adenine decreases the need of the cell to synthesize new parts
, but the third panel does not change the ability of
the Celtic taken nutrients, so what you see is that moves away
from the control condition to a metabolic state that is more
dormant because the cells have to both make less ATP and also
are decreased in their activity in synthesizing new bio
components. Now I will turn it to Allison.
Allison: So just to wrap this up, I am going to discuss some
recent data we are currently generating that is pretty
exciting for us, and that is looking at this question, in response to a lot of this new
information, which is why your metabolism is so important.
Why is this not more commonly seen as a mechanism, why is interfering in his metabolic
processes not one of the main mechanisms by which bacteria if
aevolve? And in fact, modifying the targets specifically or actually
modifying the antibiotic self, none of those are directly related to the metabolic
processes that Jason just spoke about.
And so one of the answers to this that we figured is that we
were not really looking for these, and that is because the
way these experiments are typically done are dependent on
the growth of new cells in response to the drug.
And so in a lab, what we do is we evolve cells over incrementing the increasing drug
concentrations, so that means that the end of the evolution
when we collect the cells and we sequence them, what we have had
our cells that have had to grow
and grow over time, which means they competed with the rest of
the population, ultimately dominated by the one that is the
most resistant, while everything else becomes noise.
If instead of adopting cells to growth over time, we get a more
metabolic-adaptive approach assessing the metabolic state
overtime, then perhaps it could
fuse a part of mutations, which would give us somewhere to look
now so that we can start to think about them as metabolic types.
And stead of — instead of incrementally increasing the
drug concentration, we increase the metabolic itself, a short
burst of Atlantic followed by drug-free growth to remove the
competitive aspect of selecting for the most resistant cells,
and as you can see her on the left-hand side, on the classical
evolution, approach, you can see the most dominant and rich by
logic process that emerges are those in response to the
antibiotic, those canonical mutations of the three I just
described, but when you use the metabolic approach, the first
thing that comes up is the generation of precursor
metabolits and energy, followed by ATP puricne metabolic processes.
Neither of these approaches are either more or less clinically
relevant than the other. Both of these are artificially
implemented in the lab to try to tease out mutations that emerge
so that we can then go look at these clinical isolates and see
if anything is relevant, but we know that it is actually
occurring. And so we looked at these
mutations, these new metabolic mutations, compared to the
phosphates, using the published E. coli genomes, about half that
were clinical, the other half were not, and what we saw is
that these mutations, the metabolic mutations were
significantly overrepresented compared to the canonical
representations, to a similar degree or in many cases even
more so than the ones we were expecting to see, so that is
really exciting, and that is where we are going to end off
today. With that, we want to thank
every collaborator we had working on all of these projects
, many of which come here right here at the broad at MIT and
Harvard, others that are farther away, along with funding support
and all of the additional grantors that have been
instrumental in his work, and of course everybody here who tuned
in to listen, and finally, if anyone is interested in learning
about anything we spoke about, Jason’s paper is already
published, and mine will be eminently available, and we both
welcome any questions. Jason: Thank you for your
attention. Aviva: all right, thanks,
Allison and Jason. We will open it up to questions.
I have had a few come in. If you want to add yours to the
list, I will just remind everyone that if you are on
Zoom, you can add them via the Q&A feature in your toolbar, and
if you are on YouTube, you can add your questions and comments
next to the stream. Let’s kick it off with a
question from Morgan, first together question in.
Did you ever explore problem is sick graphic mail — problematic graphic models?
Jason: That is a great question. In this specific instance, we
did not. Part of the reason for that is
there already exists some very high fidelity and fairly
comprehensive models of equine metabolism based on chemistry.
A more sophisticated approach using problem is stick modeling
could be very helpful, and that is using additional new biology
that we did not yet uncovered, but one of the things we did
here is to begin, see if we could use some tools that were
much more easily accessible, and one of the things we were able
to bring to you guys here is using this more simple approach,
the network models, we were able to lead us to biology that we
could actually test and validate. Aviva: OK. There is another question here
from an anonymous list appeared — listener.
What is it compared to the — mechanisms of antibiotics?
Allison: Another really good question.
Both of these that store all of these — are widely used already
and provide different insights, different types of information,
all of which eventually will be integrated together to give us a
more holistic understanding of those happening and all that is
involved here, metabolites for. Jason: Yeah, and to add, that is
a fantastic question, obviously all of those data, ideally, all
of those different measurements ought to be compatible with each
other. We are all still measuring the
same biology. I kind of think France’s of the
three blind men touching the elephant and trying to figure
out what an elephant looks like. We have different tools with
different perspectives on the fundamental biology, and Al as Allison is striving, when the
tools are available so we can do those, right now, we are examining phenomena such as
metabolism, something very a proximal like metabolites is
likely to be more sensitive. Aviva: OK. Let me ask a question, I was
curious as to beginning, when you talked about the decision to
switch from the kind of typical black box machinery, and who
knows what is happening inside, to this white box method, do you
see when you look at other potential uses of Machine
Learning and other biochemical research, other roadblocks that
might need to be tweaked or figured out before you can do
everything you want to do? Jason: yeah, that is a
fantastic question. One of the motivations for
designing and developing the study as we have been really
challenged by some of the fundamental challenges
associated with the purely data driven Machine Learning, mainly
that again it is often times very difficult to understand or
interpret what it is that Machine Learning is telling you
when it arrives at a solution. In this case, we decided to
develop this approach prior to antivirus, was a something we all care about, but this could
be the approach to any type of general biology.
And I think at a high level, here again, we are using more
simpler approaches, a simple type of metabolic modeling.
If you read our paper, you will
find that we use a simple type of linear regression-based type
of Sheen learning. But each of these — of Machine
Learning. But each of these steps are
often generalizable to improve these kinds of predictions.
So we think that this is timely. We think this is something that
may be generalized to many different kinds of questions,
many different kinds of networks.
More probably, if you were interested in metabolism and in
bacteria, maybe we could extend the exact same approach to study
cancer or heart disease or other human disease that are really
important and look at other types of networks such as the
signaling networks. So as a framework, we think
network modeling can be very powerful for augmenting and
enhancing Machine Learning, and there are many ways to do these
different steps. Aviva: that is great, it is
fixed to the other question here about how applicable this kind
of work is to cancer research. You mentioned that in the list.
We have another question here from Ann DeWitt.
What are the plans to translate this work to critical impact?
Allison: That is a really exciting questions are certainly
some that we found in the works, we are currently in the process
of validating them all
come in the lab, in vitr, some of them have already turned out
to be positive. Hopefully we will already have a
whole list of new genes or new targets that have emerged from
this, and that is what it seems like we will have.
Following that, many of these are in coding sequences that
presumably have biochemical elements to the.
I should say that following what we are already doing is next
doing an analysis to see what the actual functional effect of
these mutations are. Many of which will have a
corresponding biochemical type, you don’t, actual — know know, molecules we will use.
Following that, looking at how we can implement this, and there
is a lot of validation we can do in between.
But that is something many are very interested in doing.
Aviva: OK. I am going to ask about, and I
know you mentioned this in the presentation, the antibiotics,
how they all brought something different.
Tell me about the decision to
choose these as what you focus on in the study. Jason: sure. As many of you probably know,
there are many different types of an aquatics, and they were by
targeting different — of antibiotics, and they work by
targeting different aspects of physiology.
Here in the Collins lab , we
have been doing most of our work with the three types of
antibodies, so we have one that target the ability to synthesize
new cell wall. Another is an antibiotic that
targets processes involved in replication. And Gentamicin.
And certainly there is a lot of understood biology about how
each of these antibiotics kill in their own unique and
independent way, but one of the things we were interested in
here was to better understand anabolic death physiology that
was common across different types of an aquatics, so we
chose to work with three aquatics instead of one, because
we expected that would improve our sensitivity for shared
biology, and we chose these three athletics over several
different other kinds of antivirus, because these are
types of an aquatic cente — of antibiotics that are known to
kill bacteria as opposed to the growing more.
There are several other antibiotics where these tests could be extended, and we are
excited to add onto that. Allison: We also did the same
thing for antibiotics, so we look at common metabolic
mutations rather than drug-specific targeting
mutations and identify the common practices between the
drugs. Jason: And let me just add one
last comment in relation to the previous question, this
might lead us to new drug targets that could be broadly
applicable and useful for secondhand antibiotics as
opposed to, you know, creating targets that would be very
specific to individual antibiotics, which might draw
out the drug pipeline. Aviva: We have another
participant question here who wants to know about how the
method is to other microbe species, and they are
particularly interested in human gut microbes. Jason: Yeah.
Allison: That is a project that
is actually currently ongoing. Presumably, the mechanism should
be maintained. The concept should all be
applicable. We do not know the answer
necessarily to gut microbes yet, but we know generally we cross
tested, for example, the very first that we validated a
couple of state pathogens, some of the most dangerous or
classified dangerous pathogens, and so from that standpoint,
this is generally applicable, and we are looking at ongoing
right now. Jason: And to add, you know,
again, on a technical level, these approaches that we are
developing, these should be generalizable to any biological
organism, and in addition to the bacteria that Allison is
ascribing, we also extended to other pathogens, such as
tuberculosis and active ongoing collaborations can have anti-TV
at a while experience — anti-TB antibiotics.
We would be happy to email and chat with you about the work
that you might be working on. Aviva: I want to raise up
another question from Sabrina, because Allison had some pretty
briefing #antibiotic resistance, and Sabrina once you know simplistically based on your
work, can you avoid certain
foods to increase efficacy? And I will just add what we you
say to others who are worried about antibiotics in general?
What are your thoughts on that? How do you talk to people about
that ? Allison: These are great questions.
I love this question from Sabrina. It is fascinating.
In fact, to not necessarily come and about what foods to eat or not he, that is for treating
physicians to do, but we can say certain metabolites are being
investigated, and one study showed that glucose paired with
specific ones can increase the
efficacy of antibiotics, so likely there are different
combinations of commonly found, commonly would be coming into
contact with that you would be eating, and hopefully some of
this work can help elucidate exactly how, you know,
understanding, so we can guide people better.
Regards you Aviva’s question, these numbers scare me every
time I thought about them, because it is an incredibly
concerning topic, and it is a very difficult one to
appreciate, especially the patient-physician relationship,
because individually, it does not necessarily feel like it is
the time to take the risk and, you know, hold off on
antibiotics if there is something pressing going on, and
yet at the public health level, every prescribed antibiotic is
contributing to this process. So I would say that even being extremely informed, I am still
well aware of all of the, you know, caveats that happen once
we step outside the lab and be mindful.
There is a lot of education that hopefully can increase
everybody’s understanding of making decisions that will be,
you know, beneficial long-term. Aviva: OK.
We have a few more that came in. One wants to know in terms of
purely machine driven Biologics, have you ever created something
that works in real life? Jason: Yeah, so we have not
directly been working on, for instance, biologic compounds as
their abuse, but we do know — as they are abuse, but we do
know there — as thereap apies,
but we do know there are some Machine Learning in that area.
Gina has several in that space, and to our knowledge, to our anecdotal knowledge, yes, there
have been compounds formed that have been observed the
efficacious in treating several different diseases, including
infectious diseases in mice, but of course there have not been
human trials , so we are
optimistic, but I think the date, the data is supported, but
not yet conclusive. Aviva: OK. Elizabeth wanted to know — can
this type of approach be used to address the side effects of
antibiotics during the development of gut microbiota?
Allison: Sure, yeah, another
interesting question. I think, again, this will speak
to developing antibiotics that are more targeted, more
specific, more directed toward the use processes, that when we
get to a point where we can actually do that, things like
avoiding collateral damage of a microbiolome — microbiome,
that is something certainly contributing to antibiotics
resistance, because as we treat and obliterate natural microbiomes , I am thinking of
in terms of as we develop new strategies in the future.
Jason: And I guessed at, you know, there’s new innovations
that are happening in the space of network models to simulate
what happens to, you know, gut microbiome ecosystems.
Similar to what we described here, which in this case is very
focused on one individual organism could be generally
applied to, you know, exploring and understanding different
effects that you have in the different microbiomes between
different individuals. You can envision this is also an
idea new to microbiology that can inform, predict, or even
treat things that are related to the gut. Aviva: We have a few other
questions about different ways to apply parts that you develop.
They want to know if you could see a future application and
probiotics use and selection. Allison: That is something that
we have not necessarily touched on, or maybe Jason slightly
touched on , it is a very known
and established thing, so capitalizing on that is sort of
one of those areas which I would characterize as an eco-/
/evo, or economical, evolutionary, not
that bio specific target but targeting rather higher level
processes, take advantage of these evolutionary dynamics that
we know are occurring. So certainly not yet, but
eventually in the world of probiotics, taking advantage of
metabolism specifically, because that is something that can be
easily programmed into cells, is a very interesting and a very
exciting area that is probably just now on the surface of being
explored. Aviva: OK. We have Laura Dunphy, who wants
to say “Hi, Jason,” and says creative use of bio plates, have
you guys tried the resistance, and in doing so, how can you
predict longer-term impacts of metabolism and antibody to fix
the? — antibiotic efficacy? Jason: Lore is asking a
technical question about how we did our screen, and in our case,
in our paper, we describe that we used compound libraries that
were purchased from a company called by a wall, — called
Biowall. We used compound libraries to do , experience, and understand
which ones were influenced him and Laura’s question is can we
or how do we use the same compound lotteries, and to answer your question very
specific leak of a Laura, no, we have not yet done these
experiences — experiments. That would be interesting and
would be very much in line with some of the questions that
Allison is exploring. There are fundamental questions
that have not yet been answered about understanding, you know,
what are the traits that constrain how resistance
develops and what forms resistance , specifically
developed, and approaches like the experiment you are
describing, metabolic compounds of changing environmental
conditions would be really informative for that.
So it is a fantastic suggestion. Aviva: All right.
Thanks for expanding that question, too, for some of us
here who may not be as familiar with that stuff.
The last question to close us out — and we have had a great
questions about applications about this type of work.
I was wondering if there was anything else you guys wanted to
add about how Machine Learning can be used, other works you see
going on out there now that get you excited or future
applications that make you excited about where this kind of
work can go. Jason: Mmhmm. I can share.
So I think that, like I said, part of the inspiration for this
work is to address the fundamental questions about how
we can make Machine Learning accessible to us so that we can
understand what a computer is doing when it is making
associations or going into different models.
And so one of the things that I am optimistic for is that newer
approaches will come online that will help us augment the way
that humans, as scientists, we
already approach and tackle the scientific method.
You know, I would refer some of our viewers to a really nice
book called “the book of Why.” And what you do your Perl — Judea Perl proposes is helping
us identify more systematically new mechanism that can explain a
phenomenal, and Machine Learning can augment and accelerate by
increasing heading back to this. Allison: Yes, and I sort of
touched on this before, but the way that worldwide ask — is that probiotics can be going,
next generation microbial’s which take advantage of the next
evolution, because we are one set behind these passages every
step of the way, and being able to implement clinical approaches
so we can predict revolutionary for directory of populations
exposed to many different environments, knowing that we
can develop these new strategies that leverage all of this
information and sort of be them at their own game.
— beat them at their own game. Aviva: Jason and Allison, I want
to thank you for doing yesterday, and bought on behalf
of the Alumni Association, thank all of you for turning into
today’s faculty forum. The alumni staff will answer
anything that you addressed to Jason Allison or anything you
want to bring back to [email protected] Thank you.
Jason: Thank you.>>Thanks for joining us. And for more information
on how to connect with the MIT Alumni please visit our website.