 So, thank you for the invitation, it's been a really interesting day and a half for me. I think for all of the big data people here, this is a small data talk, but I think that part of my job in participating is to make the point about how these interdisciplinary multidisciplinary collaborations can evolve to try to utilize everything we know about the brain to solve problems that affect individual patients and I want to talk about how these models have been built and how they've led to treatments and then how we can work and reverse engineer them to move into big data platforms to actually test some bigger hypotheses in other data sets. I'm going to be talking about work from basically two parts of the lab. There's a treatment selection biomarker group that and what I want to emphasize is that, you know, we have a very small group of people who do the imaging. We have a big group of people that operate on patients, recruit patients, treat patients, we have to have pharmacologists, we have to have psychotherapists, but the scientific group is really engineers and imaging scientists and statisticians, so hopefully that will be a familiar kind of group. So I want to start by kind of the question that's going to guide my talk and it's really about biomarkers and seeing a lot of these tools that might be extracted by either by the data reduction techniques we've been talking about is why do we want them and I think it's sort of an obvious question, but I think it's important because it will drive how you actually set up data acquisition and data analysis and I've learned that keeping this primary question in mind is important. If you have the wrong diagnosis then you definitely go down the wrong rabbit hole for treatment and that kind of seems basic, but actually picking the wrong treatment is an even harder problem because that really can have unambiguous consequences. So it's easy to diagnose that you're having a heart attack with an EKG, but you don't flip a coin as to how you treat it, you actually get a scan to actually determine what's blocked, what you're going to open up. Similarly, if you have a mass and are presumed to have a cancer, it's not enough to know if it's benign or malignant, but now actually you're treated by the type of receptor profile you have and that is not flipping a coin at all. And even the most basic things like if you have an ammonia, penicillin is cheap and available, but it's absolutely worthless if you have tuberculosis. So again, that diagnosis actually may not be the most important step and I would argue that for depression, we need some other diagnostic schema if we really want to do better than the psychiatrist. They're pretty good at the diagnosis, but the issue in depression or thinking about precision medicine and that's a big push is the idea that we have lots of treatments for depression and the problem is the likelihood that you'll get well first treatment somebody gives you is 40%. And I'm sorry, nobody wants to be ill and have trial and error when basically you feel bad, you can't concentrate, you aren't eating, you're not sleeping and actually you might want to be dead that the consequences of trial and error for that kind of problem are significant. And worse is that the wrong treatment may actually allow the illness to evolve, which may have maladaptive brain consequences, not otherwise specified, but people become resistant over time and we don't even know that actually getting the wrong treatment doesn't contribute to that. So we may be part of that problem and again we need markers of that. And right now we don't really have any kind of biomarkers of any kind that helps us to do that. So I want to talk about the idea that a brain scan, can it or can it not enter into this equation and that may seem like a huge leap. Oh, I want to make the argument that while if one thinks about depression in the brain and starts to look at the nuances of is this or is this not a identifiable circuit in the brain, what goes wrong with it when you get sick because obviously it's an event, you aren't born with it, you have an episode, you move in and out of episodes, those are different states that one can model. And you can say well it's a very heterogeneous disorder and well is it running your family, did you have a life event, what are your prenatal insults, what's your, are you abused. There are so many variables that you might argue from an information point of view, too many variables to model. I want to argue that you can take a first step and say can the brain inform me of in any way. And we have a lot of tools and we've had a lot of tools to do that and the ones that are readily accessible now started out with doing PET scans of resting state metabolism or blood flow. Now we can do fMRI or diffusion, I'll talk about all of these. There are a lot of ways in which one can approach this problem but I think judging from you know thinking about network systems neuroscience take a first pass of is there a resting state because this is a pervasive state of the brain. It isn't like you have to be provoked to get depressed once you're in the state you're in it, so a resting state ought to be able to identify it. And in fact if you go back 30 years you could throw depressed people into a first-generation PET scan and actually see with fair reliability and reproducibility low frontal metabolism, low frontal blood flow. You could see it in patients with unipolar, bipolar, obsessive-compulsive disorder and depression. You could see it in Parkinson's, Huntington's, Alzheimer's, seizure patients with depression. And so it looked as though there was a replicable finding in the frontal lobe until when Drevitz actually reported some years into it that no depressed patients equally depressed didn't have low frontal activity they had high frontal activity. So all of a sudden we have a conundrum and if you think about all the discussions about failure to replicate as somehow failure, no actually it's a clue. And I think the issue is that was an early clue that the community saw as it reflected different etiologies or reflected something about symptom heterogeneity, but I want to make the argument that actually it was a clue to the dynamics of the brain except we weren't looking at it right. So if you actually took depressed people who had low frontal activity and you put them in the scanner after you had treated them for six weeks and now they were well and just asked what's the change in metabolism, nothing fancy. What you saw is that people who got well corrected their frontal metabolism but they also had new changes in the brain in areas that weren't abnormal to begin with. And this was the first example of changes in limbic and paralympic areas decreases in the subclosal cingulate area 25 decreases in the anterior insula changes throughout the rest of the brain. But more importantly people who got the same treatment who started out with the same apparent pattern didn't change the same on the same treatment. So we were seeing variations in response to the treatment as a function of an unknown. And if you went back to the baseline scan in people and looked at what was different between people who got better and people who didn't at the beginning, it wasn't about their frontal activity. It was actually about activity in the rostral cingulate that could be in an upstate predicted that you would do well or a downstate in people who do badly. And the frontal cortex was equivalent. And that became one of the early and subsequently replicated finding as a predictor for drug response. The problem was is that it just had something to do with if you would or wouldn't get better on drug, but it didn't actually help you to know which drug and it had nothing to do with any other treatment. So the next step became how do you start to build these regional models that are evidence based and data driven to start to think about what brain regions not their dynamics yet, but just who should be at the table. And we started to study how the brain changes with drug, how the brain changes with cognitive behavioral therapy, you know, different evidence based treatments that get people well. And we started to see that different regions recurred over and over, but the direction of change wasn't necessarily the same. And it was different combinations of regions with different treatments, which really set the stage for saying, what's the source of the variance so that recovery is recovery moving to the same endpoint, or is recovery actually moving to different endpoints and different people. And again, the thing was how to decide what someone needs. So we started to think about how to do very rudimentary modeling at first with structural occasion modeling to actually look at the interactions more causally between these regions that were coming up as important to the change effect. And we started to notice that there was this recurrent pattern that stability of any of these models, whether it was for the therapy responsive type, the drug responsive type, or people who weren't getting better on anything, had something to do with the stability and the interaction of this area 25 and adjacent areas in the singleton. If you heard Konstantin's talk yesterday, this came out in his model with DTI data with his mania analysis. And this is old pet data that goes back 10 years. So again, these these different convergence zones were starting to emerge. And so something we needed to do next was to say, well, let's randomize people and actually look if we can toward our goal of, is there a pattern that will help us to decide you go down the road that you need therapy and drug won't touch you, or you need drug. And even though you don't want to take a drug, therapy is not an effective argument. So way experiments are set up is to actually enroll people, scan them, randomize them to one treatment over the other, spend three months with expert treatment to actually look at the outcome as your independent variable to actually run what was different about your scans at the beginning. And what we found when we ran a simple, you know, two way ANOVA on outcome, trying to figure out yes or no therapy or drug, we got six candidate regions. But the goal clinically was to have a biomarker that could categorically be used in both directions. Yes, yes or no to drug or therapy, and a choice between yes therapy or yes drug or no therapy or no drug. So all of those interactions. And the only region in this analysis that emerged that criteria was the anterior insula. And even at the individual level, you can see that if you had high activity above mean metabolism of the whole brain, you did great on drug. And if you had gotten randomized to therapy, it didn't touch you. Alternatively, if you had low metabolic activity in the insula, you did great on therapy, but drug was a bad choice. And if you added as a test, drug or therapy, the second the treatment you weren't randomized to as a add on treatment, and then look to see did your starting metabolic pattern predict your ultimate response. Actually, it did pretty well that drug added to to therapy. If you had the hypermetabolic pattern got most of those patients well. So the next step is now that there's a putative biomarker is to enroll new patients, scan them, type them and treat by brain type. With the bar being you better do better than chance. Or why would you do a scan? But I just want to kind of set up the logic that this this could have been done with machine learning. But but that with machine learning, you wouldn't have known the biology of necessarily of what was driving your signal. So again, I just want to kind of start with the idea that in many ways, this just kind of lays foundation for the fact that the brain can be the same brain regions can be in different configurable states. And I think that should set up some ideas about how to think about network dynamics, and to find the nodes that you might want to use. Well, we also learned that there were patients who didn't get better even on both treatments. And so what characterize them? Well, they tended to be the drug type. Not a big surprise. Once you're kind of sure someone you've eliminated sending them to a therapist, then you get to the hard work of what drug are they going to need electric convulsive shock? I mean, you're starting to get into a different domain of trouble for a depress patient. But if you look at is there any signal that could characterize the people who never get better versus the people who do get better, a signal was seen again in an area of the brain that certainly I pay a lot of attention to this subcolosal siguline that overactivity there was predictive or tracked with high activity in the insulin people who were non responders. Well, this area 25 has been a recurrent theme in depression and in negative affect. It goes up with intense negative mood. It down regulates across most treatments. It's been identified as an interloper to the default mode with resting state fmri. So our next experiment was to actually try to see if it is always part of the equation with negative affect. How are other members of the rest of the brain tracking with it? And so we set up that same experiment that I said again, except we got a new group of people. We did a resting state fmri scan this time instead of a PET scan. Again, randomized them this time to two drugs, one of two drugs and the same cognitive behavioral therapy for three months. And again, did an analysis to look at can we discriminate with another metric? This idea that there are actually different brain sub types of depression. And again, because I want to move to another theme, the bottom line is we found three regions through an elaborate permutation analysis to really try to ensure that we weren't getting spurious results. But we found that 25s connectivity to the peri aqueductal gray to the ventral lateral frontal cortex and to medial frontal cortex, again, each discriminated. And actually, if you added them together in terms of their functional connectivity profile, did quite well. And we could use a sliding scale accuracy coupling analysis to actually find a level that you could use with this connectivity score and have actually a fairly respectable effect size and fairly respectable accuracy to actually predict response or non response to both to all four groups, which in the end is what you want. A patient wants to come in and not see the psychopharmacologist and have them have to decide no, I'm going to send you to someone else or to the psychotherapist who is going to want to try to treat you and keep you want to come through a door and actually be slotted to the appropriate treatment based and everyone wants to know that what I have to offer won't be good for you. And I think that's what's incredibly important about these kinds of findings. And not that we will send people to have scans, although that would certainly be one possibility, but it actually does get out of the religiosity of thinking that a physician just knows what to do that you could you could actually see this illness as different types at the brain level and not necessarily at something clinically relevant. So in thinking about, you know, the the talk from you yesterday and and the talk this morning on the details of the connectome, you know, again, when you see these regions that come out of whole brain functional connectivity, you want to ask yourself, well, are these regions relevant? Well, it turns out that those three regions that showed a functional connectivity in that analysis for outcome are actually three regions that have strong and monosynaptic connections with area 25. So there's a functional disconnection or a functional hyper connection in actually regions that are wired together. And I think that again, it shows the value of, you know, these these pictures are are from old, very old studies in non human primates that really were at the core of how we interpreted all of our putative network analyses, we didn't have the luxury of a human connectome. And now we do, but that actually working at whole brain instead of isolated regions, we're actually recapitulating these connections that and hardwiring that seem important. So I want to switch gears to just spend the last five minutes. What time do we start? Okay, that to to actually move to I think what a lot of people thought I might talk about, which was the brain stimulation. But I wanted to give that background because it kind of shows with, you know, starting it's sort of, I don't know if it's bragging right, or just a sign of age. You know, the early pet scans were three slices, we never luxury of even doing a whole brain analysis, we had to actually use models whether they were from basal ganglia physiology from the early 80s to make some educated guesses. And now, you know, the the new wave is to say, let a machine tell you where the answer is and in in a machine is smarter than you are anyway, which I'm sure it is. But that at the same time, there's a certain fun factor, but responsibility factor that when you don't have the tools, can you actually set up a hypothesis, but then now work backwards and see what the machine have told you are more analytic tool, give you the same answer. But prior to that, knowing and kind of trying to describe the details and whether or not it makes biological sense, seeing that change in area 25 over and over again was part of our our main rationale as to why we thought that that might be a target for deep brain stimulation in people who were intractably ill, that if you need to down regulate it, whether it's with different drugs, tuning it with psychotherapy, it even changes with placebo response. If you can't talk it, drug it, shock it down, maybe just go in there and directly tune it with an electrode, which is what we do did leveraging the technology from deeper in stimulation for Parkinson's disease. So we actually took this putative model and actually directed electricity to affect change, hoping that we would down regulate locally, but impact remotely on the rest of the network. And it worked. And what we've spent the last 10 years trying to figure out is how it works and how to do it better. And that comes back to actually now invoking the tools and neuroinformatics, but also the high resolution and patient by patient imaging tools, because what we learned is even for resistant patients, there's variants. So maybe different patients need to be implanted in different places, depending on the type of treatment resistance they have. That's one possibility, but they're much smaller samples. But we now have another variable that's more controllable. And that's a surgeon. Surgeons can miss. It's a small wire stuck into a small place. And again, after watching two micron analyses, I hesitate to act as though, you know, a couple millimeters is like the universe to all of you. But at the level of the human brain in an electrode, a contact separated by a millimeter and a half may be the difference between getting better and not getting better. And the question is, is can that explain the variance in outcome? And by looking at tractography, and what is driving a group response? Is there a common track map? Can you look at what are the common features that when you model the volume of activated tissue, what white matter tracks does it hit in common to people who get well? It turns out that it's area 25, and the singulum bundle, and the forceps minor, and the striatal fibers and brain stem fibers, and the unsinnet fasciculus. And if you miss on either side, you only have a partial effect. And if you get, but that common blueprint is seen in everyone who gets well. So we've prospectively targeted the blueprint and never changed the location and stimulated for six months and have a 75% response rate instead of a 40% response rate, controlling one variable and never getting into tuning the parameter space, even though listening to this morning, one would like a readout of that. But now we can, because we no longer have the variable of where to worry about. That I think we've optimized that at the current level of precision. You know, if we have hardy scans, if we have higher level resolution, maybe we can do better. We at least know that that's necessary, but we don't know if it's sufficient. I want to kind of just end with where I think that this has been kind of a linear step-by-step approach. But as you kind of eliminate variables that you don't want to deal with so much, you start to make some observations that actually totally disrupt your thinking. And that's kind of a point that the team is at now. And I think that relationships with bio informatics and big data and physiology and control theory and all the things that require collaboration well out of any area of my expertise, comes back down to a clinical observation. There is a phenomena when you're in that right spot and you turn the current on and you reach a threshold of adequate currents where the patient basically pops out of this local minimum. These people are in pain and they cannot move. And you get into that right spot. They feel a lifting and they want to do stuff. And this is predictable by the tractography and we are 17 for 17 patients of knowing which contact will do it. But then you'd say, okay, well that's what's that signal you're done. No, you have to actually leave the stimulator on and the evolution of the classic antidepressant response is variable. Some people are better in a week. Some people take, you know, three, four months. There's an instability element to recovery. But it's all dependent on that first reset component. And this is reflected if one looks at the PET scan signals over time taken at epics. You come back a month after surgery and look at the PET scan change in 25 also in the insula. Whatever we did, we changed it. And even though it's been off for a month and the patients still don't feel well, the state has changed. Whereas areas remotely in the network actually require stimulation on and they're changing over time. And they don't necessarily stay in the state you put them. They bounce around. So where should the control signal be? So what we're doing now is we're hunting. We now have access to a device that allows us to measure local field potentials. It's a really crude, really awful. Any of the physiologists would just roll your eyes. But it actually doesn't take building the machine, doesn't take DARPA, doesn't take special permission. You can make the recording off the electrode that's commercially available. And it'll actually record the LFP and store it and can actually now listen and follow to what the time activity curve is of the recovery at 25. And then do other measures that might give us alternative control signals that might be in facial expression, that might be an actigraphy, that might be an autonomic function. And we're also doing high density EEG. And I just want to kind of end with the idea that this is where I'm realizing that machine learning approaches, more big data approaches are going to be important. So there's a phenomenon in these wildly ill patients, their facial expression is is disturbed. You know, just like all of us can tell when someone is down. When people are depressed, and I think when they're profoundly depressed, I found this little avatar cartoon on the web. Actually, patients are aware that their face doesn't work right. They're aware that their mimetic facial expression does not match their internal state. And they worry about it. And so, you know, we said, well, you know, there's commercially available software to look at facial expression, we'll just do videos. And you can use the automated measures, you know, they've been well validated with Ekman's, you know, 60 years of work. Everyone knows the corrugator and the zygomatic muscles are very important to smile and frown. But when you actually look at how these patients are changing, it's not capturing it. You can see it, you can hear it, the metric is not capturing it. So, we engaged Shamim Nammadi in the bioinformatics group at Emory, and we gave him these videos. And he and his computer science student, Sahar, basically looked at those same videos without sound and performed machine learning set of algorithms. And we're given three time points of where the patient is clearly well, patient is clearly ill, and an intermediate point where unless you're talking to the patient, nobody in the lab, not one psychiatrist without the sound could actually tell if they were partially better or still sick. So again, suddenly beyond talking to someone where they may be in a period of instability, can their brain and plastic changes tell you the state? And it turns out this becomes a clue beyond what we thought we could do ourselves. To actually see that at a place that the psychiatrist picks that they think they know where they are, the brain says no, some people are still in the sick place and some patients are now in the well place. So we're actually seeing that facial expression itself, which is probably driven by midline singulate, which is one of the bundles we must change. It's actually maybe a better readout than actually talking to the patients, which doesn't make the psychiatrist happy, but certainly is moving in, I think, a big science kind of way. So I just want to end with the idea that I think the discussion is the bidirectionality of the synergies and the opportunities that we have that, you know, as, you know, I'm not a psychiatrist, but I obviously study and take care of these ill people. So I have to have a very pragmatic point of view, even though I would like not to. I'm looking for partnerships to do reverse translation. I don't want to wait till someone in a mouse tries to model this disorder, which is nowhere close to the human condition. But now that I know which aspects of the network that I'm in, where the comparative neurology intersections are, then one can do reverse engineering to actually model break systems and try to understand why what we're doing is actually working. We've looked at how to do our PED and MRI with machine learning and Cameron Craddock was a graduate student in the lab and published on that work. We've collaborated with statisticians to try, you know, different kinds of graph theories, different kind of hierarchical models. They all work. You can discriminate across a data set with different points of view. We've even built computational models to actually understand the reset between what is this nonlinear dynamic as someone gets sicker thinking about the excitatory inhibitory balance between 25 and prefrontal cortex. But I think that we have to start thinking big about you know, single modality versus multimodality, how to actually have the structural connectivity inform the functional connectivity, how to now look at if you break successively different aspects of this network, how does it impact the whole brain? And I think there's just a lot of opportunities, many ideas I've gotten today as to how to think about that. And so I think I just kind of end with the idea that think part of collaborating with clinicians is the idea that it's the opportunity to test a hypothesis that has real world implications. So I'll stop there. So let me kind of expand just briefly. One of the thing that happens is you follow these people weekly, and you're following their Hamilton depression rating scale is it's going down, it's going down, you get to about eight weeks and suddenly out of nowhere, they sort of have this emotional irritability reactivity. When we first started to think it, you think you need to change or alter the current. But as you actually listen to the content, you actually realize that they're actually feeling and having more feedback and their emotional bandwidth just seems to be alterating and they're starting to get it. And that the team started waiting and not making any parameter changes. And it would actually smooth out. So as we started looking at and you can tell when you're on the table, you can see a change in facial expression as soon as you turn the stimulator on that maybe there was a clue because we would all laugh that we could see how well they looked and they didn't get it. Then they suddenly had this emotional reactivity. It was almost as though emotion was too loud while they got used to it. And then it smoothed out and they resumed actually having insight. So I mean, I think there are there are some connections to your work that could be tested as well. We were just trying to see if there was a readout there. So it's in essence, a bigger picture of how to kind of stop the fighting between the clinicians who think they know and are often wrong. People not twiddling the dial and maybe without a control signal actually moving it off of the place it needs to be. And I am envisioning that that will be tested as our independent variable to see how the LFP signal is going. So I would like to see how the network dynamics at the LFP or EEG level is related to where the face property is. And they're getting fancy to where they can they can put it into 2D space. They can click on, you know, voxels and see which part of the face is actually driving the discriminator. And maybe we'll have an idea that will help guide some biology. But it's it's more that I'm trying to kind of reconcile the patients don't know what's happening to them. Do you see the same thing in activity monitoring? Do they start to get more active at that time too? I'm sorry. Do you see the same thing in activity monitoring? Do they start walking around more or they start moving around actually more very early. So the first thing they do is it's almost as though as soon as you apply the current, you release the break. And they're if they weren't bolted into the headframe in the operating room, I think you could actually pick it up. It's just a very hard way because it's a truncal aconizia. I mean they feel weighted down and then they feel as though they can move. So I think that activity happens early. This emotional reactivity happens in an intermediate and actually autonomic reactivity is also changing late. So I think that different pathways, different systems are correcting and they're not coordinated and and there and the question is is how to get a read out so that one can get a sense of of relative changes in different parts of the network, which is why I would argue instead of trying to do that with the whole brain and just see what happens, one might be able to have some more tenable hypotheses to track the behaviors that actually have a known biology and try that first. Thank you. Hi, is the the location of treatment where you placed the electrode? Is that the same throughout the patient's life? Or is it allowed to variable? Is it allowed to be variable? Is it would it be tough luck if it's in a bad place such that the person can be treated? So good, good question. And I probably didn't explain that adequately. We always have very minor anatomical variability. So the surgeons all want just give us a coordinate in standard space. Bad plan because the variability is the branching of the tracks. So what we did is we said, let's take a person's track map of and model what we're stimulating, binarize that person in native space, then put everybody in standard space and see what is common to everyone and kind of define that. And even when you know that you it's there's no Euclidean difference in terms of the location. But you can drive in an individual person in their deterministic track map in native space and find the spot that has that. And you you have some play. And we have to figure out if rate of changes, maybe you need proportionally singular and forceps minor and equal degrees versus, you know, but when when you drive, and if you ever take a deterministic track, you can drive and you reach a point where you see these four bundles just converge. And so missing with the tractography has not been a problem. But missing when you do it by anatomical eyeballing is the location. It stays the same throughout. Yes. So so with the surgery, you make a burr hole, you put a guide wire in canula and you put it in and it doesn't move. You anchor it in at the skull and it doesn't. There's there's some minor things and we've just finished looking at whether or not air in your head from being in the operating room moves it along. But once it's in, you can verify sometimes, you know, where you put it in, it can bend or when you but we haven't actually had that problem. The angle is fairly straight and it it seems sort of magical, but the surgeons are good at it. And that hasn't actually been the problem. I think getting people to use that it's conectomic surgery. It's not nodal surgery. And I think that's what we've kind of learned and now the details of the connectome that we're in is where we're at. Thanks. The nomination of my public measurement and the resting state functional connectivity after treatment in 25 area. So did you see the relationship between the two measurements? I mean, those medial analysis. So we're trying to now look at how metabolism, functional connectivity and actually white matter, disintegrity are related to one another. And that and I am sorry, I didn't include it. We actually see an overlap, particularly in the DBS patients where we have all the data where actually you can see points of white matter disintegrity in the bundles that we're stimulating that are on that are adjacent to areas of functional disconnectivity. So one of the things we're really interested in is is the process of recovery and the time activity curve recovery, a function of how much white matter damage you have. And there's evidence in animals that high frequency stimulation in short bursts will actually induce oligodendroglia changes and will grow new new oligos that make myelin that will migrate across the corpus colosum. And I think that that becomes a best place to kind of look mechanistically at what this might be about. But I think the biomarker of trying to decide is the pet better than the fMRI. How are they related? That's kind of what we're doing now. OK.