Stephen Wilson 0:05 Welcome to Episode 15 of the Language Neuroscience Podcast. I'm Stephen Wilson. My guest today is Rodrigo Braga who is an Assistant Professor in the Department of Neurology at Northwestern University in Chicago. He's a cognitive neuroscientist who uses fMRI and ECoG to study large scale networks of the brain, with a particular focus on association areas that are expanded in the human brain. He published a very interesting paper about the language network in 2020 and that's going to be the topic of our discussion today. It's called "Situating the left-lateralized language network in the broader organization of multiple specialized large scale distributed networks" and it's published in the Journal of Neurophysiology, an excellent journal when you want to explore something in depth and you don't want to be worried about pesky page limits. This beautiful paper comes in at 34 closely set pages. Okay, let's get to it. Hey, Rod, how are you today? Rodrigo Braga 0:56 I'm good. Thank you, Stephen. Thanks for having me. Stephen Wilson 0:58 Yeah, I think this might be one of the might be the only podcast I've done where my guest is in the same time zone as me. Right. It's also 11 in the morning for you, right? Rodrigo Braga 1:08 It is yes, yeah. When your guests are circling in on you. Stephen Wilson 1:12 Right. And it's a very beautiful spring day in Nashville. About 65 degrees. How about where you are? Rodrigo Braga 1:19 It's also really nice here. Yeah, we've had a bit of rain, but right now, it's really beautiful. Not too hot. Stephen Wilson 1:25 Cool. A good day to talk about language and the brain. Rodrigo Braga 1:29 Yes, every day is. Stephen Wilson 1:31 Yeah. So well, you're not really I mean, most of the guests that I have on the podcast are, I would say like language people. I'd say you're not really a language person, right? You're a cognitive neuroscientist that just happens to have written a really interesting paper about language, or is that... Where does your heart lie? Rodrigo Braga 1:47 No, I'd say that's pretty fair. Yeah. When you invited me on, I was like, Oh, um, you know, I, I can see why and the paper's relevant and I'm more and more getting into language as it's obviously a very interesting process. But no, my background has not been in language at all and I you know, kind of avoided it. Kind of on purpose. (Laugher) for most of my career. Stephen Wilson 2:10 OK, cancel the interview. (Laughter) Rodrigo Braga 2:14 But uh, but you know, I was in, my PhD was in a language lab with Richard Wise. Stephen Wilson 2:18 Yeah, I realized that. Yeah, so you have actually like, quite a pedigree. Um, yeah so, I Googled you a little, before we met today and I learned that you were born in Brazil and then you move to the UK, like, how old were you when you moved to the UK? Rodrigo Braga 2:35 I was seven and I had my birthday, like a week after, my eighth birthday the week after I moved to England. Stephen Wilson 2:41 Aha, and I also learned that you're a musician with numerous albums. Rodrigo Braga 2:45 That's right, yeah. After my undergrad, I took a couple of years out to be a musician full time and that's what I was gonna do. That was my path. Stephen Wilson 2:55 Yeah, I took a listen to some of your stuff, it was pretty cool. You definitely got the Brazilian kind of guitar influence there. I could, I could pick that up and then kind of some more sort of ambient stuff. I don't know how you would describe it but I'm not musicologist. Rodrigo Braga 3:08 Yeit depends which album I guess I was with a progressive rock band. We call it progressive funk, because it was kind of like funk rock. But then, when I went went solo, later on, I sort of took in some of those Brazilian influences and did some Bossa Nova, more Samba rhythms. Yeah. Stephen Wilson 3:27 Yeah. Very cool. But what made you decide to become a scientist instead of a musician? Rodrigo Braga 3:34 Yeah. It's a weird path because I only went to university to do music really. You know, I decided I always wanted to be in a band and so I thought, well, I'll go to university that's a better place to meet people and meet other musicians. But I didn't really want to study music formally so I thought well, what do I, what's cool to study? I like philosophy and psychology. You know, I was always fascinated by the mind and thought I wanted to study those things in more detail. But then slowly, that brought me into neuroscience, because that was the stuff in psychology that I found most interesting. And so after a couple of years of being musician, and having some success, but also just seeing that the path is going to be quite arduous, sorry. I just realized, you know, I was living in my parents house and thought, how many gigs? How many weddings am I gonna have to play to to actually support myself and I thought, I just need a job. So I started looking for bad jobs and then I thought, why am I getting a bad job? I have a degree in Neuroscience. (Laughter) I just go back to university and then when I started, I did a master's first Stephen Wilson 4:36 so you see, you're like, oh, academia, that's like, you know, that'll be an easy life. Rodrigo Braga 4:43 Pretty much. Yeah. I was like yeah, you know, my plan B side gig, but I was still doing music during my PhD and masters but basically, I just realized that I really loved it and super interesting and met some cool people have had the privilege of working with really cool mentors like Richard Wise and Robert Leech as well. Stephen Wilson 5:02 Yeah, really legendary language, guys. But you didn't catch the light language vibe from them? Rodrigo Braga 5:09 No, I guess not. Yeah. So when I started my PhD, Richard saw on my CV that I was a musician and he was like, why don't you study music? And so, I did my PhD was on auditory processing. I did look at some of the distributed networks and the default mode network, the posterior cingulate cortex, in particular, a couple of experiments that Rob Leech had devised to look at that I, you know, worked with him on those. But yeah, so until then there was no no language at all. Stephen Wilson 5:38 Right. And then you move to the US, for your postdoc and that's where you wrote the paper that we're going to focus on today. But um, just going to ask you in general, like, what do you think about science in the US versus the UK? Like, what have you noticed about, you know, cultural differences and whatnot. Rodrigo Braga 5:56 Very different, very different environments and people warned me before I came to the US that, you know, people work really hard there. The hours are really long and so those are kind of scary things but it's also the science in the US is really valued. So training in the US is really valued and I really thought I would do a good postdoc and go back to the UK. But then, I kind of really enjoyed the, doing science here. You know, at least the labs have been working on very well funded experiments really ambitious, they don't have I mean, it seems to me at least some of the funding limitations that I hear more about in the in the UK and so yeah, and also the PhD programs are longer, and you know, people can go deeper into the topics. So that's been something that's been fun to be a part of, and have people really go deep, deep into questions rather than Stephen Wilson 6:49 From the mentoring side of things you mean? Rodrigo Braga 6:51 Ah for both for the student and for the mentor. Stephen Wilson 6:54 Yeah, I mean, but you, you as the mentor, you've enjoyed having being working with us, actually, we've got a longer study. Rodrigo Braga 7:01 Exactly Stephen Wilson 7:02 Yeah. Rodrigo Braga 7:03 Yeah, and just to see how that training is prioritized early on, rather than is more of a rush for papers. I think the three year PhD is quite first, when you're expected to get three or four papers out of it. Stephen Wilson 7:16 Yeah, well, I mean, talk about Americans working hard. I mean, that would require some work wouldn't. Rodrigo Braga 7:20 That's true. Very true. Yeah. Yeah. I didn't mean to say that the Brits don't work hard. I, you know, we will I work too hard for PhD. Yeah. It's different, I guess. Stephen Wilson 7:28 Yeah. I remember, I was visiting London, my friend, Fred Dick there couple of years ago and I remember he was showing me around and he pointed up at Cathy Prices' office, which we just happen to be walking past and the light was on and he was like, that's Cathy Price's office. And I was like, oh, no, she's probably like writing something really amazing. (Laughter) Rodrigo Braga 7:47 On a Sunday morning. Stephen Wilson 7:51 Yeah. No, it was like, I think it was like 7pm or something. Anyway, um, and so yeah, you got you got yourself a K99. Right? K99/R00? Rodrigo Braga 8:03 That's right, yeah. Stephen Wilson 8:04 Is that in progress, still? Rodrigo Braga 8:07 No, that's finished so I'm on, well, I'm on the R00 phase, which is okay. Stephen Wilson 8:11 So it's sort of in progress. Rodrigo Braga 8:14 Yeah, I guess so. Yeah. So the they split it but it's basically the same grant, when you get a faculty position, the K99 is a mentoring as a training ground, you're supposed to train a new technique, learn some new skills, apply for faculty positions and then once you get faculty position, you get the R00 funding, which is the three year funding. So it's a great, great grand. Stephen Wilson 8:33 Yeah. And when did you have a position at Northwestern? So congratulations on that. I think any position in this current job market is incredible achievement. When did you start? Rodrigo Braga 8:44 So I got the confirmation email that they offered me that, I accepted the offer and they said it went through. I think it was it was shortly before the pandemic was, I think it was like January or something where that email came through, Jan 2020. And then obviously, everybody started talking about closing down or shutting down job openings and also, taking back some offers. I heard some horror stories about that. I didn't experience that myself. But I definitely sent an email to say, just want to check (Laughter) will I still be coming to Chicago? Stephen Wilson 9:17 Yeah. Rodrigo Braga 9:17 Yeah and so we had to move ahead. We had to move during the pandemic and then have to start the lab during the pandemic and we had to do everything remotely for a while. Stephen Wilson 9:26 Yeah. What was that like? Rodrigo Braga 9:30 It was a few months into it, so we sort of more used to it. I imagine it was hard for I had one trainee at the start and we did everything remotely but managed to make it work. We didn't shut down scanning for too long, which was quite helpful because we were able to go in and test different sequences. Stephen Wilson 9:47 Okay. Yeah, to me, yeah, I mean, I think everybody's been forced to do things just that didn't seem like they would be possible. Rodrigo Braga 9:56 Yeah. Stephen Wilson 9:58 And you know, it's been a It's been rough for everyone. I mean, certainly it's been rough in my lab in some ways, but I have a lot of empathy for people like, you know, that are in transition at that time, like, you know, starting a new lab just seems like a whole nother level of, you know, challenge. Rodrigo Braga 10:13 Yeah. Stephen Wilson 10:17 Okay, so let's talk about this really interesting paper that you wrote that I contacted you about. It's about essentially identifying the language network using functional connectivity and I was hoping we could start by just kind of, if I could ask you just a basic question, which is, you know, what is a network? As, you know, as defined in this sort of functional connectivity sense? I mean, just bearing in mind that, you know, our listeners range from, you know, students to postdocs to faculty with different areas of expertise, I want to try and get everyone on the same page of knowing what it is that we're talking about when we're talking about functional networks? Rodrigo Braga 10:59 Yeah, it's a good side question, because there's a lot of different interpretations of that, different uses of that word and my understanding is that it comes from graph theory, and how you can have different nodes that are connected by some measure. Usually we, in brain imaging we look at correlations and so you have two brain regions, those would be your nodes, and then you'd have a correlation between them would be the link between them. So if you're doing full brain imaging, some people consider that the whole network the whole graph, as sorry, the whole graph as a network. So every brain region is involved in that network and then you have subsets of those nodes that form systems or sub networks or whatever term you want to use. In my training, I've always used that term slightly differently and in Randy Buckner's lab, they also use that word slightly differently. We talk about a network as being more regions that are, that shows strong anatomically connected patterns and so we sort we, it's almost like a subsystem of the full graph that is what we call a network and so if I, if I do a correlation analysis of a given region of the brain, and I show a correlated set of regions, that's what I would call a network. And so and there's some evidence of anatomical connections if you look at tracer studies in the macaque, so that's also what we call a network and maybe you wouldn't see the full the whole brain being connected using a single tracer injection and so there is that difference between what we're calling a network and what I think graph theory, what they consider a network to be. Stephen Wilson 12:41 And when you say regions that are correlated, what is correlated with what exactly what do you mean by that? Rodrigo Braga 12:47 Yeah, so we study the brain using functional magnetic resonance imaging, and that measures, changes in blood oxygenation, a signal called the blood oxygenation level dependent signal, BOLD signal, and what we do is put someone in the scanner and just have them stare at a plus sign or just do something very low level. In our case, they're just staring at a plus sign. They do that for seven minutes and we record the activity of multiple brain regions during that whole, the whole of that seven minutes and then we can look at correlations of signal from different brain regions that have taken place during that time. So, yeah, does that answer your? Stephen Wilson 13:28 Yeah, I'm just trying to get that real groundwork in, so, brain regions are kind of fluctuating in ways that seem almost random, but probably aren't and then they they will have a signal that's fluctuating in time and then other brain regions that belong to the same network will have a somewhat similar fluctuating signal and when those time varying, blood oxygen level dependent signals correlate, that's what leads you to put those regions together in a network. Rodrigo Braga 13:54 Yeah, that's exactly right. When you look at the spatial pattern of correlations, they aren't random. There is a very, there are very strong, reliable, reliably observed sets of regions that show high correlations with each other and those you can look at the anatomical tracer studies in the macaques and there is some homology there, which builds evidence that these correlations are actually tapping into anatomically connected networks. Stephen Wilson 14:21 Yeah. I mean, I think that the first discovery of it was in 1995. Right? There is this is paper by Biswal that I'm sure you're familiar with, where they put seeds in some motor region and then, you know, look at regions that are correlated with that, and what they end up is like mapping out this really nice set of motor areas, like, you know, it's, you know, the homotopic part of motor cortex, there's, like, supplementary motor area on the midline, maybe even some brainstem motor areas. So, it's sort of became apparent that like, you know, these spontaneous fluctuations were meaningful. But you know, and then that continued in, there's 1000s of papers about it as as we both know. But the precious little, I mean there are these well known networks, but like the language network never really popped up much, right? I mean, we've got these well known networks like default mode, you know, you can get motor and visual and frontoparietal and salienceground whatever you want to call them all. But never language. Almost never. So why? What was the initial observation that made you think you could somehow parse out a language network from this functional connectivity approach that no one had really done before? Rodrigo Braga 15:27 Well, just before I answer that, I'm curious to know what your perspective was. Because as someone coming from a language perspective, you would see these resting state maps, network maps and you'd, how was that for you? We wouldn't see a language network. Did you just think this is all? This isn't true? (laughter) Oh, no. I thought it was true. I mean, going into some, I mean, I guess, like, there this is paper by Steve Smith, I think from 2009, that does show this, like, left lateralized and the right lateralized variants of the frontoparietal network, that they speculate that they're seeing language there, like that, you know that one of them is language network, and one is not. And then, you know, I guess there is Jeff Binder's work on the relationship between the default mode network and the semantic network. So I guess my prior, prior to reading your paper, was that language regions were getting kind of, I'm looking for the right word, like, dragged into other related networks, such as default mode, frontoparietal, maybe even ventral attention and that we weren't seeing a coherent network emerge using most of the approaches that were commonly used. But no, I never, I never thought that resting state was, was bunk. Oh, I mean, I think that most of the studies that are done with that are bunk. But the fundamental concept is great. Right. Yeah, I mean, I was just curious, because I came at it from the resting state world. So, I sort of was curious as to what your perception of it was. But so a couple of things. One is that even within the, the field of language, I listened to your interview with Ev Fedorenko and she had been making this point for a while, that to look at, to study language, and to study specialization, we need to look within individuals, obviously, training with Nancy Kanwisher, who had been doing this and others in the field of vision, they've been looking at within individuals for a while showing that you can actually see much more detail, and you can separate things by looking within individuals that you don't see in the group average. So there were hints already that to study language you need to do, you need the high resolution imaging, you need to focus within individuals, you can't do group averaging. But we, I came at it from a completely different, I came at it from a completely different perspective. So, I was interested in the default network and so we were trying to understand that organization, and the language network sits right beside the default network and so as we explored the anatomy of the default network with the new techniques that we've been developing, another network just popped out of the data that was like, oh, this looks like it could be a language network, and it's so reliable, we can see it within multiple individuals. Stephen Wilson 18:19 Right. Rodrigo Braga 18:19 It is kind of an accident, in a way. Stephen Wilson 18:22 I guess I'm just gonna note for our listeners, that the default network is this network of regions that, deactivated to a very wide variety of tasks, as originally observed in 1997, by Shulman et al. and the most prominent ones, prominent nodes of it being the bilateral Angular gyri, and as well as several midline regions, maybe anterior temporal. And yeah, so the adjacency to language is clear. So yeah, so you, you kind of notice the language network popping out and then this paper is, you know, you kind of deep dive into exploring the language network in relation to the adjacent networks. So can you tell us, and the first analyses you do you kind of have the seed based approach? Can you tell us about how you selected the seeds? And what you saw when you put seeds in those regions? Rodrigo Braga 19:10 Sure, so a little bit of background is that like everybody else in the field of functional activity, we had been doing things at the level of the group, which is when you scan multiple individuals who align their brains, and then just average maps across individuals to get better signal. Just when I started my postdoc, the paper had come out by Russ Poldrack and Timothy Laumann in the Petersen group, Steve Petersen's group, where they had basically collected so much data from an individual that they can make the same maps but from at the level of that individual. So that didn't need this group averaging step which is actually quite catastrophic because it tends to blur. You're just basically blurring across anatomical boundaries because every, everyone's brain is different. And so, this paper came out that was, that was showing that you could define these networks with precision within individuals and we basically took that idea and just ran with it basically. And so, we what we did is we collected lots of data, lots of high quality data, from lots from a few individuals, just from four individuals. We had them come in 24 times for 24 different MRI sessions. We collected hours of data for resting state analysis, but also loads of different tasks. We didn't really, you know, quite ambitious projects that I coded these tasks with postdoc called Matt Hutchison, in Randy Buckner's lab, and we had, you know, I can't remember how many in our 15 different tasks, one of it was actually a language task, which we just collected. It was Ev Fedorenko's language task. We had functional localizers for motor cortex for instance, that became really relevant later on, but basically had all of this data. And then the idea was just to take a look and see whether we could observe any new features, upgrade organization. So that's the context and so then, then we decided to how we're going to actually study the anatomy of the default network, well, a great way to do it is to do the seed based functional connectivity approach. Basically involves picking a region of the brain and then looking at the correlation map. So what we set up this experiment where I would just pick different seeds within an individual's brain, and then just look at the resulting correlation patterns to see if we can detect a robust network. Now, there are certain issues with that, not sure if you know if your listeners will care about, but Stephen Wilson 21:39 They might. Rodrigo Braga 21:40 Basically anywhere you select the seed, you're going to see something, right? And so we have to determine a couple of rules to sort it, to narrow down the search space. And one was that we wanted really high correlations, we wanted the seed based correlation map to show really robust correlations as opposed to the more diffuse correlation, which happens when you're in between two functional regions. And the other tasks that Randy assigned to me was, show me two networks that are side by side, both of which are robust correlations, but, but occupy distinct regions of the brain. And so that's basically the the full task was that and so I just started selecting seeds and trying to find weed seeds actually captured some of the networks that were evident in that person's brain. And some of the first few images that I showed him were two networks that looked really similar, but they weren't spatially shifted along the cortex. And both of those networks fell within the bounds of the canonical group defined default network. And so that was the first inkling that we got that, oh, we actually see new features here when we look within individuals. Stephen Wilson 22:53 So, was that a subdivision of the default mode network into what you call in your paper, DN-A and DN-B? Or are you talking about parsing out the language network from the default network? Rodrigo Braga 23:04 So right now I'm talking about splitting the default network into two networks, default network A, default network B. Stephen Wilson 23:10 Okay. Rodrigo Braga 23:10 Conveniently named. And so, so that was the first insight but while I was doing this seed-based approach, occasionally, I would move my seeds more eventually, into the lateral frontal cortex. And I would see this really striking, robustly correlated network that had large regions in the inferior frontal cortex, and the lateral temporal cortex. Here, I was just looking in the left hemisphere, typically, but when when when I looked at the right, it would also have smaller regions on the right hemisphere, but they were shrunken in size in comparison. Stephen Wilson 23:44 Right. Rodrigo Braga 23:45 So it was left lateralized, which to me was like this, this is clearly a lot must be the language network. Stephen Wilson 23:50 Yeah. You'd spend enough time with Richard Wise to know the language network when you saw it. Huh? (laughter) Rodrigo Braga 23:58 Yes, for sure. But also like, you know, neuroscience 102 is like Broca's area and Wernicke's area. Stephen Wilson 24:05 Right. So where were your seeds for these default mode networks? Like what brain region were you poking around in exactly? Was it like in the MFG or something? Rodrigo Braga 24:14 It was close. It was actually more dorsal in the superior frontal gyrus. For the default network, it was roughly around there, but some subjects I couldn't find a good seed, so I'd move it around and go a bit more ventually and yeah, if I, if you go ventral to the middle frontal gyrus and then posterior towards the central sulcus, there's a really strong language region there, which in Matt Glasser, his Nature paper, they talked about it being region 55B, which they also attribute it to language network. It was one that I didn't know about until coming across it. David Summers has also looked at that region with regards to auditory working memory, I think, to something that had been studied, but just when I was searching around kind of naively just popped out that that was connected to more inferior frontal regions and lateral temporal regions, Stephen Wilson 25:07 Right. So that's where you end up placing your seeds in your paper, right, in your first seed-based approach like that's where you kind of have your starting seeds. Rodrigo Braga 25:15 In the language that work, exactly, that region 55B. Stephen Wilson 25:19 Yeah. And so in the paper, you, you know, you show that when you put a seed there, you basically light up not only Broca's area, which is central to that but something like Wernicke's area in the left Superior Temporal sulcus-ish, and both sides, as well as several other smaller kind of that rather reproducible language regions, right? Rodrigo Braga 25:45 Yeah, and even the, you know, so called Broca's area, we typically see multiple regions there, that circle around the Operculum and sometimes you have three, what looks like three distinct regions in the IFG. So I don't know what to call all of those Broca's areas, or, you know, but just say what we just described, what we see, we see multiple regions there, we see that area 55B, one that's more dorsal and then along the lateral temporal cortex, you have multiple regions that often just, you know, a large swath, swathes swath of cortex that seems interconnected, extending all the way to the temporal pole, which again, we know is important for language from semantic dementia patients that have atrophy there. But as you say, we also saw these other regions that we didn't expect, we saw along the midline, in the dorsal posterior cingulate cortex, there was quite reliably a small language network region there. Similarly, similarly, in the middle cingulate cortex and other small region, in the inferior temporal lobe, we'd see a tiny region of the language network, and sometimes even in the ventromedial prefrontal cortex, which is a hard image a region to image, we'd also see this tiny region of the language network, consistently enough across individuals. So it wasn't always there but it was there consistently enough that we thought, oh, I think this is part of this extended network and what was interesting is that when we get my initial initial super excited, like, Oh, we found all these new things that have never been described. But then when you look in the literature, there are instances of people describing a language that has those regions. Famous, famously in the paper by Lee in 2012 and Hacker in 2013, they use the more advanced task map informed way to parcellate the brain using functional connectivity and if you look at those maps, they do have these tiny regions, and we saw them. Stephen Wilson 27:41 Ah that's cool. Rodrigo Braga 27:42 Which is cool. Yeah, it was just reinforce that those are probably real as well. Stephen Wilson 27:48 Yeah. And then you found that if you put the seed in other nodes of your language network, you would also be able to kind of produce the network, right, wherever you started, right? You didn't need to start only in 55B, you could start anywhere and kind of see the same thing. Rodrigo Braga 28:05 Exactly, yeah, you could put it in the inferior frontal gyrus and the lateral temporal cortex, and I think we put the seed in speech SMA, pre-SMA region, not what you always could seed to find the network. Stephen Wilson 28:17 Yep. Okay, so that's like your seed based approach and then you have this whole other approach that you use extensively in this paper too, which is this K-means clustering, right? So can you explain how that works and what you see with that approach? Rodrigo Braga 28:32 Sure, yeah. So one limitation of the seed-base approach is that I had to manually go in and select seeds to define the network and somebody could argue that I just found very unique seeds that don't really represent the organization of the brain. You know, the unicorn seed that I pick just so happens to look like language network and there are other that that's a fair criticism. There are other criticisms, like, when I select a seed in a given region, I actually bias all the correlations there because there's a spatial bias. Regions that are close to each other, just tend to have high correlations. Stephen Wilson 29:06 Whatever seeds you choose, the nextdoor voxel is always going to be the most interconnected voxel in the brain. Rodrigo Braga 29:10 Exactly. Yeah and so, so that's it, you know, these are good things to consider when looking at seed-based connectivity maps. So another way to do it is to just do a data driven format that doesn't involve me actually selecting anything manually and there are different ways to do this. There are more advanced ways to do this. We just did a K- means clustering approach, which basically takes the connectivity pattern of each voxel or each vertex if you're on the surface and then it clusters, that connectivity map. It clusters those connectivity maps from each vertex. So at the end, you get a map that's which vertices have similar connectivity patterns. Stephen Wilson 29:49 Oh, okay. So you're not clustering like the resting state time courses, you're clustering, the conductivity maps that would be generated from each voxel when using it as a seed. Rodrigo Braga 29:59 That's right. Stephen Wilson 29:59 Okay. Right. And so that kind of produces the same thing. More or less, right? You replicate your language network? Rodrigo Braga 30:08 You do. You basically replicate it. Some regions change in shape because of the spatial bias, but yes. Stephen Wilson 30:13 Yeah and then reading your paper like that to the point where you suddenly show us the right hemisphere for the first time and I was very excited because like, you know, up till then I was skeptical. I was like, Okay, you're showing me something, it looks like a language network but like, you're only showing me and left hemisphere, like, is this, you know, does this have that hemispheric bias that we would expect? And then you do your K-means and we see that it clearly has a leftward asymmetry. Rodrigo Braga 30:39 Right, but I mean, that's exactly right. So we cluster the whole brain, right and left hemispheres, and it still produces a map that is left lateralized in the sense of having larger regions on the left compared to the right and that, that's cool in itself and because if you use a parcellation technique, given the spatial confounds, sometimes it can split a network into a left hemisphere version and a right hemisphere version. And so it was was cool to see that in this case, it didn't do that it was still there still bilateral, much larger on the left and on the right and when we then took after seeing that, I then took the seed-base approach for the scenes on the right hemisphere and it again, showed larger readings on the left for some individuals than on the right. So they also have supported each other, which was good. Stephen Wilson 31:26 Yeah, cool. Okay, so then you do something very cool, which is you kind of connected to task activation data, and you had, I don't know whether it was wisdom or good fortune that you got a good task from from Ev, and because you know, not all tasks are as good as each other, that's for sure. But you had a good one and you looked at the correspondence between your putative language network and the regions that are activated by the language task, which is listening, it's reading sentences versus reading pseudowords and what did you see? Rodrigo Braga 32:05 Yeah, so we saw a really striking correspondence I wasn't expecting when we look and we compare the two maps. And first of all, credit to Randy Buckner who decided to collect that task. That wasn't a decision that I made and credit to Ev Fedorenko, who devised the task and you can see in the resulting maps how well controlled that task is. But yeah, so when we looked at, when we compare the functional connectivity map derived from parcellation, with the task activation map, they were just, it almost seemed like it was the same map in a way. The boundaries of regions that showed task activation, overlapped in striking detail with the regions that showed connectivity using correlation. And it wasn't just the case that you'd see that in the main language regions like IFG and inferior frontal gyrus, and middle and superior middle gyrus. You'd see that throughout the whole brain, you'd see that and even in the smaller regions that are described in the posterior cingulate cortex, middle cingulate cortex, you'd see evidence of task activity and that's a good call because it's a completely different data set different analysis, it's a contrast, rather, you know, task activation contrast rather than a correlation map and yet, they're showing you the same map that tells you something about something. (Laughter) Stephen Wilson 33:28 Yeah. It tells you that it's real. How did you feel when you saw that correspondence between those two datasets? Rodrigo Braga 33:36 I felt excited, but I was like, I must have done something wrong. (Laughter) Stephen Wilson 33:40 Because it was too good? Rodrigo Braga 33:41 It was too good. Yeah. Yeah. So I checked other subjects I rechecked the data made sure I had the contrast right and but yeah, was it didn't expect it to be such a close match was really impressive. Stephen Wilson 33:51 Yeah, I thought so too. And we talked a bit about, like, the lateralization, how it tends to be in you know, left greater than right. But that's not what you saw for every subject, right? Rodrigo Braga 34:04 That's right. Yeah. So most subjects were left lateralized, and but to varying degrees, it's important to say that in every subject, we saw language network regions on the right hemisphere, even if they were really, really small, there was still there are some hints of them. But most objects had were bilateral in the sense that they clearly had regions on both hemispheres, even if they weren't slightly larger on the on the left and we did try to quantify that as a percentage of the total cerebral hemisphere that was in the language network. Most objects were left lateralized and then a couple were really bilateral, you can say one was larger than the other and we had one individual was completely, very strongly right. lateralized Yeah. And the cool thing there was we had the task activation maps because well, two things one, we initially identified the language network in the left hemisphere of this subject, just by looking at the distribution of regions. It's like, okay, we think this is the language network, but the regions are really small. That's frustrating. And then when we look at the right hemisphere had really big regions on that same network. But then when we looked at the task activation map, that again, seemed to match the correlation map. So this person also had stronger right hemisphere language task activity than the left. Stephen Wilson 35:23 Right. So you have concordance between laterality as determined by the more conventional approach of a task versus your novel approach of I mean, I guess it's K-means based at this point, at the paper right? So that's pretty cool. Yeah, I'm sure. And it wasn't even just that one subject who was right lateral eyes when you were lucky to get a right lateralized person, because they're pretty rare. You know, like, you have 15 subjects or something. How many siblings do you have? Rodrigo Braga 35:50 18 subjects. Stephen Wilson 35:50 18 subjects, I mean, yeah, I mean, your, your chances of having a right lateralized person would not be high within 18 random individuals. You got one, and then you also have those bilateral people and they also kind of replicate right, um, between the task based and connectivity based approaches. Rodrigo Braga 36:08 Yeah, that's yeah, they these are already good to see the right hemisphere. lateralized person was the exception that proves the rule that there was a correspondence between them between the functional connectivity maps and the task activation maps, the bilateral ones were also just good validations. And but again, not all the subjects had some degree of bilaterality. So that was good to just see that replicated across subjects and just to emphasize, the maps weren't perfect. If you look at the maps, you can see regions where it doesn't match. So it's not like maybe you don't want to oversell it, there were definitely regions that didn't correspond across the the two types of analyses. But if you look across enough individuals, that the pattern is that there is a really strong correspondence. Stephen Wilson 36:52 Yeah. What's an example of a finding that you saw, that was where you'd see something different in the activation map and in the connectivity map? Rodrigo Braga 37:03 Those were largely not as consistent when we saw discrepancies that weren't as consistent across individuals. But there were regions. For example, sometimes in the Angular Gyrus, Inferior Parietal Cortex, we would see a region in the task activation map that wasn't usually there in the correlation map. And then in some individuals, you'd see hints of other networks, like the default network, or the dorsal attention network that's involved in attending to the outside world. Stephen Wilson 37:34 I mean, wasn't that bit of the Angular Gyrus that you often saw, kind of, I mean, that was part of the default network, right. In those individuals? Rodrigo Braga 37:41 Yeah, it could it could have been, it could have been, we didn't look at it in enough detail. But Stephen Wilson 37:47 I mean, don't you think it's not that surprising that you wouldn't have perfect overlap? I mean, because I mean, it's not like a task could be expected to identify only one brain network, right? It's gonna draw on multiple, you know, cognitive and linguistic processes. You wouldn't, you'd expect to get a bit of something else, right. I mean, let's say for instance, suppose there had been a button pressed as part of the task there wasn't controlled out in the way that it is, you know, then you would expect to see like a task would activate language network and some motor network, right. I mean, so like, we wouldn't really expect a one to one correspondence, right? Rodrigo Braga 38:21 Yeah, that's exactly right and that's, I think, speaks to the strength of the task that we did see that much correspondence. But for instance, if someone found the the nonwords condition very difficult, then you expect to see other regions that would be recruited. Initially, when we piloted one good thing about the approach of focusing on individuals is that you can test scan someone a couple of times, and if they're not good candidates, you just recruit somebody else, rather than scan them 24 times and get bad data and so initially, when we piloted, like one individual particular, worked out the timing of the task, and when the button presses were going to be and so they would just close their eyes. And like snooze for the 715 seconds, whatever it was in the wake up and press the button and so we we monitor their eyes and make sure that that wasn't a factor in the in the actual data. But that's definitely something something similar can be happening that we aren't able to track. Stephen Wilson 39:14 Right. Yeah. So this correspondence that you have between the latterality, raises the possibility of clinical application, right. So, you know, part of my job, we do language mapping and pre-surgical patients and there's two questions that you often want to address, which is, you know lateralization like which hemisphere is language lateralized to and also, where's language localized within the dominant hemisphere. The fact that you can do this from resting state data has a lot of potential because there's a small subset of patients that might not be able to comply with a task. Have you thought about clinical development of this finding, or do you think you'd leave that to others? Rodrigo Braga 39:58 Yeah, it's definitely it's something I find fascinating and I think they definitely see the utility of it. Typically, in hospitals, they have a language worker, but I guess, you know, they use different tasks to try to identify these regions of, yeah. I think it needs a rigorous large study to actually show the validity of it in terms of improving outcomes. I think that's something that's, that's necessary important can convince clinical teams to, to base their decisions on the functional connectivity maps. Stephen Wilson 40:32 Yeah. Rodrigo Braga 40:32 I think that's yeah, that's a question, definitely it's interesting. Stephen Wilson 40:36 Yeah, I mean, I think I will certainly plan to acquire some data to test this. Because I feel very confident about the language maps that we can generate with tasks. But, you know, there is a small like, especially with kids sometimes, or with patients who are very impaired, like you do sometimes get, like, the inability to comply with the task instructions, and then you don't get good language maps. And the question would be whether you might get something, you know, valuable from resting state and those people? Rodrigo Braga 41:05 Yeah, absolutely. I mean, I think that's super exciting. I think that one of the cool things that we found is that it didn't really matter too much what they were doing, obviously, we talked about the resting state, but it's an unconstrained state, somebody could be doing multiple things like remembering the past, thinking about the surroundings, feeling very uncomfortable and angry are the people scanning them, putting them through this. And so, we don't really know what's happening in the resting state. But it's assuring reassuring that if you look at a different task, so we looked we did we compute the correlations during the language task, we computed correlations during the motor task, which involved making the finger movements and tongue movements, for instance. And the maps, the correlation maps look very similar throughout that to my eyes are very, very similar. And so, you know, we've even collected data during movie watching, because it seems to be better for the participant is less boring, than seeing him sitting in the scanner for an hour, but yet, we can still see the same map. So I think, especially for kids, that's something that could really be useful. Stephen Wilson 42:04 Yeah, I mean, I think it'd be very cool to just be able to put the kid in the scanner, show them a movie and just collect functional data incidentally, and then get what you need from that. That'd be really useful. I mean, it seems it seems quite plausible to me that that could be developed. Okay, so you know, you in your paper, you relate the language network to several other networks, default mode A, default mode B, frontal parietal. Is there another one that I'm forgetting? Rodrigo Braga 42:36 The salience network. Stephen Wilson 42:38 Oh, yeah, salience network, also called Cingulo-Opercula. How does the language network kind of lie in relation to these adjacent networks? Rodrigo Braga 42:48 Yeah, that's something that is, you know, one of my favorite findings in that in that paper is just we talked about how, when we looked at the organization of language network regions, it wasn't just the key Broca Wernicke's areas that were coming out there also these other smaller regions that came out like in the posterior cingulate cortex, for instance. Now, the reason that was cool is because when we had been studying the default network, and we managed to fractionate, that into two separate networks, both of those networks looked like copies of each other in the sense that they had regions that were side by side, in multiple parts of the brain. So it looked like you just took one network, copy and pasted it, and then just shifted it spatially along the cortex, right? And so, there are very cool ideas. Randy's written about this quite a lot about how that might come to emerge during development, that you have a basic wiring pattern and as the brain is expanding during infancy, different parts of the cortex start to specialize, but they inherit this broad organization and that's why the networks look similar to each other. So that's, that was super cool. But when we looked at the language network, and we included those smaller regions, the language network just looked like another example. So another copy paste of the default network distribution, just now it's shifted towards classic language regions, like the IFG in the lateral temporal cortex and so, that just show that the language network is another instance of these distributed Association networks. It also raises questions about why did these regions specialized for language because the most prominent regions of the language network in terms of their size, were in Broca and Wernicke's areas, right near to and those regions are right near to auditory cortex, they're right near to tongue samatomotor motor regions and so to me, that implies that nearby Association cortex has specialized for language because of the influence of nearby sensory motor regions that are important for language. So, so those were that was the sort of the coolest thing for me which really has sparked questions for me about what to research next and that's something that we're doing in my lab. And so that yeah, that sequence of networks that goes from sensory regions all the way up to the language and beyond to the default network, where they all have very similar shape, but just spatially shifted along the cortex looks like this idea of a broad gradient from unimodal sensory to the default network and the language that works very well within that gradient. Stephen Wilson 45:26 Right. So you kind of see it as like an intermediary between sensory representations and like the ultimate a modal, semantic conceptual representations of the default mode network. Rodrigo Braga 45:38 I mean, at least spatially this, it sits right in between those, those brain areas and, you know, I will leave that, you know, better than I do about linguistic theories, and how that maybe is exactly what you need for a brain region to be specialized for languages to sit somewhere in between those. Stephen Wilson 45:54 I mean, I yeah, I do think that language sort of plays a mediating role. But I mean, you know, meaning has to be translated into a transmissible form, which involves, you know, the auditory modality or the visual modality in the case of sign. But, yeah, so does language have a particular, does the language network have a particularly special relationship with the default mode network? Or does it also kind of have a similar juxtaposition to the frontal parietal network, for instance, which is another network that I think is in very much analytical language territory. Rodrigo Braga 46:28 Yeah, they were both they both had regions that were side by side, all three of those networks was, you know, had a very complex, detailed organization of regions that were hard to separate, unless you look within individuals that have a lot of data. It's possible that the organization, the spatial layout of the language network better matches what we call default network be in the sense of where just where the regions are and so but I think it just, it will vary so much from individual to individual that, you know, I think it would be easy to mix all three of those together, if you're doing your group average approach where you're blurring things together. Stephen Wilson 47:10 Yeah. Rodrigo Braga 47:10 So yeah, I think there will be evidence, that would be a good rationale for why all three of those would be mixed together in past work. Stephen Wilson 47:18 Okay. So, you know, one of my favorite papers, prior to yours on on resting state parcellation is the paper by Yoe et al. which from also from Randy Buckner's lab and I know that that paper must have been influential on your paper. But in that one, what, they don't really get a language network, even though they're using a rather similar approach, like this came in on conductivity maps. Do you have any idea why you get this language network? Is it because of the individual approach that you get it and the Yoe paper doesn't? Is that the key difference? Rodrigo Braga 47:56 I think so. Yeah. I mean, basically, they, you know, completely landmark study was hugely influential. They scanned 1000 individuals, I think, more like 3000, but that those maps in that third in that paper are from 1000, individuals, using cutting edge techniques at the time, the paper was published in 2011, so cutting edge techniques just in the years before that, but that still relied on this idea that you have to average data across individuals, you know, aligning them as best as you can, and atomically, but you're still averaging across individuals and just the shape of, if you look at our our papers showing the language network in different individuals, the shape of the regions will change, the exact anatomical location of the region will change. So even if you align it as best as you can, when you average across individuals, you're just blurring across things. And I think that's ultimately, the main the main thing, it also, perhaps because we're doing a resting state task, where people are introspecting, and doing things that the default network likes to do, or shows increased activity for like remembering. The default network is one of the easiest things to detect in functional connectivity analysis of resting state data and so it's possible that that high signal there just completely overwhelms the language network that was really closely juxtaposed with the default network. And just to say that, more recently more advanced algorithms by Thomas Yoe's group, they've been able to show the language network, even in group average data, which has a slightly higher quality. So it is possible, I think it was just just beyond the edge of the detection limits to separate those things. Stephen Wilson 49:34 Right. Cool. Yeah, it's funny, like I'm looking back at his paper with your paper in mind, in the, he has the seven network parcellation, as I'm sure you know, and in that, you know, the default mode network is very clear, it's very, I mean, not everything in this seven network version is is symmetrical, right? And he's got the default mode as one of the seven networks and then he goes to the 17 network version, and then it actually does fractionate and then in Interesting way, so the default mode network seems to kind of fractionated into two, one of which is left lateralized. But it's less clearly a language network than it is in your paper like it just more like parts of the default mode network in the left breakoff. And in the right, only the ATL breaks off, which is exactly the right hemisphere region that is most easyto, which is the most, you know, language involved. node of the right hemisphere network. Rodrigo Braga 50:29 Right. So definitely hence that Stephen Wilson 50:31 There's definitely hints if you go back and like I've looked at this, I looked at Yoe's paper, like, dozens of times, if not hundreds and I've never really appreciated this before. So I think it's like hints that you see once, once you've seen that the final resolution version that you Rodrigo Braga 50:46 Yeah. Stephen Wilson 50:47 put together. Rodrigo Braga 50:48 Well check out Ruby cons paper, I think is 2019. Who were they? Do they do show the language network? I think in the 2011 paper, there's also an auditory network that's quite broad, if I'm going from memory, but I think it goes beyond primary auditory regions. And that's probably also taking up some of the language regions. Stephen Wilson 51:07 I think it is, especially in the Seven Network version, but I think it's more auditory than okay, because I think that because I unlike I mean, I think that most of the STG is just auditory, apart from the dorsal STS. In other words, apart from the ventral aspect of the STG, I think it's, you know, really auditory rather than language. But that's just my opinion. (laughter) Well, this is great. Do you think we covered all the things that are most important about this paper? Rodrigo Braga 51:40 I think so. Yeah. We talked about gradients, we talked about the association networks and how they have a similar distribution. Yeah, I think so. Stephen Wilson 51:49 We didn't talk about the intermediate network, that you identify and that's kind of my deliberate decision, because there's a limit to what you can talk about in the podcast and it's very complicated and I'll just kind of alert, interested listeners that there's this other aspect of the paper that would be best appreciated by reading it. (Laughter) Rodrigo Braga 52:08 Sure and seeing it. Stephen Wilson 52:10 Yeah, sometimes you need to see things. Yeah. Okay. Well, this was great. Thank you so much for taking the time to talk with me about this really cool study. Rodrigo Braga 52:20 Thank you Stephen. It has been a real pleasure to be among such illustrious guests. Thank you for the invitation. Stephen Wilson 52:27 You fit right in and I really enjoyed it. So thanks, and have a good rest of your day. Rodrigo Braga 52:35 Yeah, and before I go, I just wanted to do a quick plug, if that's okay. Yeah, we have we have our new library, a year old at Northwestern University in Chicago and we have positions open for PhD students and postdocs. So anybody out there who's interested in some of the stuff we've been talking about, please do get in touch. bragalab.com is where to go. Stephen Wilson 52:55 All right, cool. Yeah, that sounds like a really great opportunity. You know, for either a student or a postdoc. I'll link to our website on the podcast website so that people have that link to if they want to follow it from there. Rodrigo Braga 53:06 Perfect. Thank you so much. Stephen Wilson 53:07 All right. Take care. See you later. Rodrigo Braga 53:10 Bye. Stephen Wilson 53:11 Okay, that's it for episode 15. As always, I've linked the paper we discussed and rods Lab website in the show notes and langneurosci.org/podcast. Special thanks to Yev Diachek who presented and led discussion of this paper in my lab meeting, which made me realize that this would be a great paper to cover and thank you all for listening. See you next time.