WEBVTT 00:00:00.000 --> 00:00:02.540 Welcome back to the deep dive, everybody. Today, 00:00:02.600 --> 00:00:07.559 we're diving into a pretty complex topic. Multiscale 00:00:07.559 --> 00:00:10.480 modeling. Yeah. You know, it's kind of like trying 00:00:10.480 --> 00:00:12.679 to understand, let's say, a hurricane, right? 00:00:12.699 --> 00:00:15.279 OK. You could look at the swirling clouds from 00:00:15.279 --> 00:00:19.179 a satellite. But to really predict its path and 00:00:19.179 --> 00:00:21.859 get its power, you also have to understand like 00:00:21.859 --> 00:00:24.000 those. tiny little air currents and pressure 00:00:24.000 --> 00:00:26.600 systems. That's kind of what multi -skill modeling 00:00:26.600 --> 00:00:28.699 is all about. It's about using all these different 00:00:28.699 --> 00:00:30.859 lenses, right? Right. To try and understand this 00:00:30.859 --> 00:00:33.539 complex system at different levels of detail. 00:00:33.579 --> 00:00:36.159 And this is important for, well, tackling a lot 00:00:36.159 --> 00:00:38.939 of different things like biological systems and 00:00:38.939 --> 00:00:40.920 even designing materials. Yeah. And especially 00:00:40.920 --> 00:00:43.079 biological systems are so hierarchical, right? 00:00:43.079 --> 00:00:47.000 You have from the quantum mechanics of the atom 00:00:47.000 --> 00:00:48.939 and then the molecules and the cells and the 00:00:48.939 --> 00:00:51.439 tissues and the organs. And to really understand 00:00:51.439 --> 00:00:53.939 how life works, you have to connect those dots 00:00:53.939 --> 00:00:56.399 between all those different levels. Yeah, absolutely. 00:00:56.560 --> 00:00:58.939 And that's what multiscale modeling tries to 00:00:58.939 --> 00:01:01.880 do. And this is helpful for all kinds of applications 00:01:01.880 --> 00:01:07.140 in biology, from enzymes, how they work, to ecosystems. 00:01:07.540 --> 00:01:10.079 It's really amazing. It's like, if you're trying 00:01:10.079 --> 00:01:13.040 to appreciate a painting and you only look at 00:01:13.040 --> 00:01:14.819 one brushstroke, you'd miss the big picture. 00:01:14.989 --> 00:01:17.810 Totally. And this isn't just limited to biology. 00:01:18.290 --> 00:01:19.969 Right. I mean, chemical engineers use this a 00:01:19.969 --> 00:01:23.170 lot to design catalysts or to optimize industrial 00:01:23.170 --> 00:01:26.069 processes. So it really is this powerful tool, 00:01:26.290 --> 00:01:28.790 this interdisciplinary approach that is really 00:01:28.790 --> 00:01:32.359 fascinating. And a lot of the most important 00:01:32.359 --> 00:01:36.459 processes in biology hinge on these complex interactions 00:01:36.459 --> 00:01:41.000 between proteins, lipids, nucleic acids. These 00:01:41.000 --> 00:01:43.659 interactions drive everything from how our cells 00:01:43.659 --> 00:01:46.819 communicate to how genes are processed. And modeling 00:01:46.819 --> 00:01:49.359 these is a huge challenge because it occurs at 00:01:49.359 --> 00:01:51.780 such a huge range of scales. I mean, we're talking 00:01:51.780 --> 00:01:55.620 about events that occur at fractions of a nanometer 00:01:55.620 --> 00:01:58.859 in femtoseconds all the way to micrometer -sized 00:01:58.859 --> 00:02:01.140 structures change over seconds or even longer. 00:02:01.760 --> 00:02:03.980 OK, so let's unpack that a little bit. Sure. 00:02:04.159 --> 00:02:06.260 You're saying that these traditional computer 00:02:06.260 --> 00:02:09.000 simulations really have a hard time capturing 00:02:09.000 --> 00:02:11.219 those tiny little interactions and these slow 00:02:11.219 --> 00:02:14.340 changes simultaneously. It's kind of like trying 00:02:14.340 --> 00:02:17.240 to photograph a bullet train and a snail with 00:02:17.240 --> 00:02:19.400 the same camera settings. Exactly. You'd probably 00:02:19.400 --> 00:02:22.379 miss one or the other. Totally. But this is where 00:02:22.379 --> 00:02:24.639 machine learning comes in. Yeah, exactly. This 00:02:24.639 --> 00:02:26.400 is where things get really, really interesting. 00:02:26.500 --> 00:02:29.039 That's right. So that leads us to what we're 00:02:29.039 --> 00:02:30.560 talking about today, this new infrastructure 00:02:30.560 --> 00:02:34.259 called MumMI, Multiscale Machine -Verned Modeling 00:02:34.259 --> 00:02:38.159 Infrastructure. And our mission, really, in this 00:02:38.159 --> 00:02:41.879 deep dive is to explore how MumMI is tackling 00:02:41.879 --> 00:02:46.259 these challenges, specifically how proteins interact 00:02:46.259 --> 00:02:49.379 with the membranes of cells using the Rostov 00:02:49.379 --> 00:02:52.719 pathway as a key example. And this pathway, of 00:02:52.719 --> 00:02:54.599 course, is really important because it regulates 00:02:54.599 --> 00:02:57.319 cell growth. And when things go wrong with it, 00:02:57.129 --> 00:03:00.430 it due to mutations or other things, it can lead 00:03:00.430 --> 00:03:02.889 to cancer. Right. So let's start with the challenges 00:03:02.889 --> 00:03:05.909 then of bridging these incredible scales, these 00:03:05.909 --> 00:03:07.810 time scales, these size scales. You mentioned 00:03:07.810 --> 00:03:10.550 molecular dynamic simulations. And these are 00:03:10.550 --> 00:03:13.069 crucial, but they have limitations. They do. 00:03:13.210 --> 00:03:16.250 They do. Molecular dynamic simulations are really 00:03:16.250 --> 00:03:19.289 amazing. They allow us to see these movements 00:03:19.289 --> 00:03:23.870 of atoms and molecules. But the problem is to 00:03:24.139 --> 00:03:27.060 accurately represent the forces and interactions 00:03:27.060 --> 00:03:29.719 across those spatial scales and across those 00:03:29.719 --> 00:03:33.879 time scales from femtoseconds to seconds is computationally 00:03:33.879 --> 00:03:36.759 a huge, huge challenge. Yeah. I mean, it's like 00:03:36.759 --> 00:03:39.539 trying to map every single ant in a colony, but 00:03:39.539 --> 00:03:41.560 also track how the whole colony is moving all 00:03:41.560 --> 00:03:43.960 at the same time. That's a great analogy, yeah. 00:03:44.080 --> 00:03:46.449 It's kind of mind boggling. It is. And many important 00:03:46.449 --> 00:03:49.090 processes, like when proteins assemble or the 00:03:49.090 --> 00:03:50.710 rearrangements of membranes, involve these huge 00:03:50.710 --> 00:03:53.650 structures that span many orders of magnitude. 00:03:54.750 --> 00:03:57.689 And traditional MD simulations struggle to capture 00:03:57.689 --> 00:04:00.889 both the rapid little vibrations and the slower 00:04:00.889 --> 00:04:02.569 movements. At the same time. At the same time. 00:04:02.610 --> 00:04:04.710 And then it just becomes computationally way 00:04:04.710 --> 00:04:07.770 too expensive. So you're forced to choose exactly 00:04:07.770 --> 00:04:10.430 between this incredible level of detail and this 00:04:10.430 --> 00:04:12.669 broader perspective. Yeah, that's right. And 00:04:12.669 --> 00:04:16.370 what's missing is a truly integrated framework 00:04:16.370 --> 00:04:18.470 that can handle these different levels at the 00:04:18.470 --> 00:04:21.810 same time. OK, right. So you can have an atomistic 00:04:21.810 --> 00:04:26.149 simulation that tells you how two molecules bind, 00:04:26.149 --> 00:04:28.509 but it won't tell you how often that happens 00:04:28.509 --> 00:04:30.689 in a cell, which is what you need for those larger 00:04:30.689 --> 00:04:33.129 scale models. Right. And then there's this time 00:04:33.129 --> 00:04:35.910 scale difference. Huge. Yeah. I mean the vibrations 00:04:35.910 --> 00:04:39.430 in a molecule happen so fast compared to a membrane 00:04:39.430 --> 00:04:42.449 changing shape. Exactly. So it's not just running 00:04:42.449 --> 00:04:44.449 simulations at different scales separately. Right. 00:04:44.569 --> 00:04:46.509 You have to somehow transfer that information 00:04:46.509 --> 00:04:48.850 from one level to another. Yes. So it's like 00:04:48.850 --> 00:04:51.089 translating an engineering blueprint into a simple 00:04:51.089 --> 00:04:54.110 sketch. Yeah. Yeah. But you can't lose that structural 00:04:54.110 --> 00:04:56.209 information. Exactly. That's right. And even 00:04:56.209 --> 00:04:59.129 within methods that simplify like coarse graining 00:04:59.129 --> 00:05:02.269 methods, bridging different time scales is still 00:05:02.269 --> 00:05:06.120 a big, big problem. To create a truly robust 00:05:06.120 --> 00:05:09.839 workflow, we need to not just overcome the computational 00:05:09.839 --> 00:05:12.439 limitations, but also figure out how to transfer 00:05:12.439 --> 00:05:15.259 this information, integrate it. Okay, so these 00:05:15.259 --> 00:05:18.519 are some pretty big hurdles, but you mentioned 00:05:18.519 --> 00:05:21.709 machine learning could help. Yes. How so? Yeah, 00:05:22.089 --> 00:05:23.949 so this is where machine learning really comes 00:05:23.949 --> 00:05:27.709 in and helps us bridge those gaps. And machine 00:05:27.709 --> 00:05:29.790 learning is great at learning complex patterns 00:05:29.790 --> 00:05:32.829 from data. So it can really help connect different 00:05:32.829 --> 00:05:36.230 scales in this multi -scale framework. OK, so 00:05:36.230 --> 00:05:39.100 how does that work? Yeah, so think of it as like 00:05:39.100 --> 00:05:40.959 feedback loops between these different levels, 00:05:41.139 --> 00:05:43.199 okay? So machine learning can help with something 00:05:43.199 --> 00:05:45.879 called forward coupling. This is where you use 00:05:45.879 --> 00:05:48.540 information from a larger scale to improve things 00:05:48.540 --> 00:05:51.519 at the smaller scale. So for example, imagine 00:05:51.519 --> 00:05:54.060 a model that shows specific regions of the cell 00:05:54.060 --> 00:05:56.959 membrane changing a lot. That information can 00:05:56.959 --> 00:05:59.740 then be used to focus a smaller, more detailed 00:05:59.740 --> 00:06:02.300 simulation on that exact region. So this is like 00:06:02.300 --> 00:06:04.319 the bigger simulations guiding the smaller one. 00:06:04.480 --> 00:06:08.279 Exactly, exactly. backward coupling where the 00:06:08.279 --> 00:06:11.779 detailed simulations give feedback to improve 00:06:11.779 --> 00:06:15.019 the larger scale model. So for example, a detailed 00:06:15.019 --> 00:06:18.240 simulation might reveal some very specific atomic 00:06:18.240 --> 00:06:21.560 interaction. Then machine learning can use that 00:06:21.560 --> 00:06:23.920 to adjust the parameters of the larger model 00:06:23.920 --> 00:06:26.180 so it becomes more accurate. So you have this 00:06:26.180 --> 00:06:28.480 back and forth and things get more accurate. 00:06:28.579 --> 00:06:31.040 So it's this iterative process. Exactly. Can 00:06:31.040 --> 00:06:33.980 you give me some concrete examples of how ML 00:06:33.980 --> 00:06:36.860 is being used? Yeah, yeah, sure. So one example 00:06:36.860 --> 00:06:40.459 is in these QMMM methods, where you combine quantum 00:06:40.459 --> 00:06:43.120 mechanics with molecular mechanics. Machine learning 00:06:43.120 --> 00:06:45.480 can be used to handle these very complex interactions 00:06:45.480 --> 00:06:49.220 and reduce the computational cost. Another example 00:06:49.220 --> 00:06:52.600 is in material science, where you can train machine 00:06:52.600 --> 00:06:55.259 learning on simulations to predict how material 00:06:55.259 --> 00:06:58.319 might fail or classify defects. And then you 00:06:58.319 --> 00:07:01.079 can use that in these larger continuum models 00:07:01.079 --> 00:07:05.540 to go from very small to very large. really connecting 00:07:05.540 --> 00:07:07.540 those scales. Exactly. And then you also have 00:07:07.540 --> 00:07:10.779 things like ML potentials. These are ways to 00:07:10.779 --> 00:07:14.019 predict energy and forces between atoms with 00:07:14.019 --> 00:07:16.560 almost the accuracy of quantum mechanics, but 00:07:16.560 --> 00:07:20.079 much, much faster. Wow. And then ML is also used 00:07:20.079 --> 00:07:22.680 to go from these simplified representations back 00:07:22.680 --> 00:07:25.620 to the full atomic detail, like enhancing a blurry 00:07:25.620 --> 00:07:28.199 picture. That's a great analogy. Yeah. And it's 00:07:28.199 --> 00:07:31.139 not just simulations. It's also integrating it 00:07:31.139 --> 00:07:34.350 with experiments, with biophysical methods. So 00:07:34.350 --> 00:07:36.829 ML is helping us overcome those computational 00:07:36.829 --> 00:07:39.810 limitations, making things faster, focusing on 00:07:39.810 --> 00:07:42.610 the important parts. And these deep learning 00:07:42.610 --> 00:07:44.649 advancements are really making this accessible 00:07:44.649 --> 00:07:46.490 to more researchers. That's really exciting. 00:07:46.790 --> 00:07:49.930 So where does Mamma fit into this? Yeah, so Mamma 00:07:49.930 --> 00:07:52.910 Mai is a platform designed specifically to leverage 00:07:52.910 --> 00:07:55.649 machine learning for multi -scale modeling. And 00:07:55.649 --> 00:07:59.490 it focuses on those intricate biomolecular systems. 00:07:59.930 --> 00:08:02.790 And it does this by really tightly coupling the 00:08:02.790 --> 00:08:04.850 different scales using machine learning. OK, 00:08:04.850 --> 00:08:08.250 so what makes Mamma Mai stand out? Yeah, so one 00:08:08.250 --> 00:08:10.970 key thing is its ability to create these huge 00:08:10.970 --> 00:08:13.769 ensembles of micro simulations. We're talking 00:08:13.769 --> 00:08:16.829 tens of thousands even, all guided by machine 00:08:16.829 --> 00:08:20.110 learning. It also uses this thing called a machine 00:08:20.110 --> 00:08:23.889 learned latent space, which is a way to simplify 00:08:23.889 --> 00:08:26.889 complex data. So it's like taking all this complex 00:08:26.889 --> 00:08:29.060 information and boiling it down to its essence. 00:08:29.740 --> 00:08:31.920 And that allows us to see patterns and connections 00:08:31.920 --> 00:08:34.379 we couldn't see otherwise. And it focuses initially 00:08:34.379 --> 00:08:37.159 on these interactions between RAS -RAF proteins 00:08:37.159 --> 00:08:39.679 and the cell membrane. And as we said, this pathway 00:08:39.679 --> 00:08:42.139 is crucial for cell growth. And when it goes 00:08:42.139 --> 00:08:45.899 wrong, it can cause cancer. So Mamamai is trying 00:08:45.899 --> 00:08:48.360 to understand how these proteins interact with 00:08:48.360 --> 00:08:51.240 the membrane at a molecular level. And it can 00:08:51.240 --> 00:08:54.340 identify these lipid protein fingerprints, which 00:08:54.340 --> 00:08:56.840 are these patterns of lipid molecules. that can 00:08:56.840 --> 00:08:59.240 affect how these proteins work. Interesting. 00:08:59.419 --> 00:09:01.899 OK, so tell me more about this three -scale architecture 00:09:01.899 --> 00:09:05.779 that Mamai uses. Yeah, so Mamai has these three 00:09:05.779 --> 00:09:08.519 levels to bridge these gaps in space and time. 00:09:08.700 --> 00:09:11.720 The first level is the macro scale. OK. And it 00:09:11.720 --> 00:09:13.960 uses something called dynamic density functional 00:09:13.960 --> 00:09:18.730 theory, or DDFT. OK. And this allows us to simulate 00:09:18.730 --> 00:09:22.110 milliseconds of time over a fairly large area, 00:09:22.269 --> 00:09:24.649 like one square micrometer of the membrane. So 00:09:24.649 --> 00:09:27.289 it's the broad view, right? Exactly. Exactly. 00:09:27.429 --> 00:09:29.789 It doesn't show every atom, but it can capture 00:09:29.789 --> 00:09:32.350 the general movement of the membrane and how 00:09:32.350 --> 00:09:35.690 proteins like RAS and RAF generally prefer to 00:09:35.690 --> 00:09:38.149 sit on the membrane. OK, so that's the big picture. 00:09:38.149 --> 00:09:39.980 Right. And then we zoom in. Right, so then you 00:09:39.980 --> 00:09:43.419 have the coarse -grained or CG scale. OK. And 00:09:43.419 --> 00:09:46.919 this uses the martini bead model. And in this 00:09:46.919 --> 00:09:48.960 model, groups of atoms are represented as single 00:09:48.960 --> 00:09:51.580 beads, so we can simulate larger systems for 00:09:51.580 --> 00:09:55.279 longer. And this level focuses on smaller patches 00:09:55.279 --> 00:09:58.700 of the membrane, maybe 30 by 30 nanometers, areas 00:09:58.700 --> 00:10:02.200 that the macro scale has identified as interesting. 00:10:02.340 --> 00:10:05.639 OK, so the action spots. Exactly, yeah. And these 00:10:05.639 --> 00:10:08.970 patches often have proteins like RAS or the the 00:10:08.970 --> 00:10:12.029 RAS -RB -DCRD complex, which is part of RAF. 00:10:12.730 --> 00:10:15.110 And we use the martini force field, which are 00:10:15.110 --> 00:10:17.529 the rules for how these beads interact. And we 00:10:17.529 --> 00:10:20.830 can run these on GPUs, and we can get about a 00:10:20.830 --> 00:10:23.809 microsecond per day on a single GPU. And this 00:10:23.809 --> 00:10:26.789 gives us a good view of how proteins change their 00:10:26.789 --> 00:10:28.929 shape and interact with each other. OK, so we've 00:10:28.929 --> 00:10:30.610 got the big picture. We've got some more local 00:10:30.610 --> 00:10:34.129 details. What about the really, really fine details? 00:10:34.230 --> 00:10:36.710 Right, so then you go to the all -atom, or AA, 00:10:36.970 --> 00:10:39.049 level, and now we're back. to every single atom 00:10:39.049 --> 00:10:42.269 being represented. We use the CharM M36 force 00:10:42.269 --> 00:10:44.629 field, which is very detailed. And this allows 00:10:44.629 --> 00:10:48.370 us to see those very specific interactions between 00:10:48.370 --> 00:10:51.350 lipids and the amino acids in the protein and 00:10:51.350 --> 00:10:53.889 how those lipids might affect the protein structure. 00:10:54.129 --> 00:10:56.250 But this must be very computationally expensive. 00:10:56.549 --> 00:10:59.470 It is. It is. So we take snapshots from the CG 00:10:59.470 --> 00:11:02.370 simulations and we convert them back to all atom, 00:11:02.450 --> 00:11:05.529 a process called backmapping. And then we use 00:11:05.529 --> 00:11:08.710 energy minimization to to kind of relax the structure. 00:11:09.009 --> 00:11:11.429 And we run these simulations using things like 00:11:11.429 --> 00:11:14.789 AMBER or GROMACS. And these are much slower. 00:11:14.929 --> 00:11:18.950 We're talking maybe 14 ms per day on a GPU. But 00:11:18.950 --> 00:11:21.769 it gives us that really crucial atomic detail 00:11:21.769 --> 00:11:25.210 that we can't get otherwise. And so these transitions 00:11:25.210 --> 00:11:28.149 between scales, are there tools that help you 00:11:28.149 --> 00:11:31.230 with that? Yes. Yes. MUMMI has tools to help 00:11:31.230 --> 00:11:33.549 move between these different levels. OK. For 00:11:33.549 --> 00:11:35.870 example, there's a tool called Create Sims that 00:11:35.870 --> 00:11:38.029 converts the macro scale to the coarse grain 00:11:38.029 --> 00:11:41.009 scale. And then for going from CG to AA, there's 00:11:41.009 --> 00:11:43.889 a backmapping protocol to make sure we get an 00:11:43.889 --> 00:11:46.250 accurate atomic structure. OK. So all these tools 00:11:46.250 --> 00:11:48.990 help us move between these different levels consistently. 00:11:49.330 --> 00:11:52.340 OK. So. We've got these three scales, each with 00:11:52.340 --> 00:11:54.679 their own methods. How does machine learning 00:11:54.679 --> 00:11:57.600 tie into all of this? Yeah, so machine learning 00:11:57.600 --> 00:12:00.840 is not just an add -on here. It's really essential 00:12:00.840 --> 00:12:03.740 to Mamamai's ability to connect these scales 00:12:03.740 --> 00:12:06.620 to manage all this data. And it does this by 00:12:06.620 --> 00:12:10.559 coupling these scales, both macro to CG and also 00:12:10.750 --> 00:12:14.269 sort of indirectly CG to AA through choosing 00:12:14.269 --> 00:12:16.570 those starting configurations. OK. So it's really 00:12:16.570 --> 00:12:19.330 the glue that holds everything together. So let's 00:12:19.330 --> 00:12:21.769 talk about that coupling in more detail. How 00:12:21.769 --> 00:12:24.830 does ML do this forward and backward flow? Yeah. 00:12:24.870 --> 00:12:28.110 So in the forward direction, the ML models are 00:12:28.110 --> 00:12:31.029 trained on the macro scale data. And they're 00:12:31.029 --> 00:12:32.830 trained to pick up the interesting parts, the 00:12:32.830 --> 00:12:34.429 parts we want to look at in more detail. OK. 00:12:34.809 --> 00:12:38.970 So let's say protein starts to cluster or it 00:12:38.970 --> 00:12:42.169 really likes a specific lipid or the membrane 00:12:42.169 --> 00:12:45.250 changes shape in a specific way. The ML model 00:12:45.250 --> 00:12:47.950 will flag that and MMMI will then say okay let's 00:12:47.950 --> 00:12:50.409 look at that region with the CG model. So it's 00:12:50.409 --> 00:12:52.590 like a smart filter. Yeah, yeah, like a scout 00:12:52.590 --> 00:12:54.330 that says, OK, look here. This is where the action 00:12:54.330 --> 00:12:56.990 is. And then in the backward direction, machine 00:12:56.990 --> 00:12:59.710 learning helps refine those larger models based 00:12:59.710 --> 00:13:01.809 on what we learn from the smaller simulations. 00:13:01.970 --> 00:13:05.970 OK. So as the CG simulations run, we analyze 00:13:05.970 --> 00:13:09.190 in real time how those proteins and lipids are 00:13:09.190 --> 00:13:12.090 interacting. OK. And that information can then 00:13:12.090 --> 00:13:14.950 be used to adjust the macro model, make it more 00:13:14.950 --> 00:13:17.990 accurate. Similarly, the information from the 00:13:17.990 --> 00:13:20.350 all -atom simulations can be used to improve 00:13:20.350 --> 00:13:22.649 the CG. model. So it's this constant learning 00:13:22.649 --> 00:13:26.129 and refining. Exactly, exactly. It's fascinating. 00:13:26.350 --> 00:13:28.769 And beyond that, machine learning also helps 00:13:28.769 --> 00:13:31.929 with how MomMI manages all these simulations, 00:13:32.149 --> 00:13:34.850 how it decides which simulations to run. Okay. 00:13:34.909 --> 00:13:37.629 It uses something called dynamic importance sampling. 00:13:38.110 --> 00:13:40.470 Okay, so what does that mean? What makes a simulation 00:13:40.470 --> 00:13:43.419 important? So it really depends on what the scientist 00:13:43.419 --> 00:13:46.299 is looking for. OK. Right. So maybe it's a protein 00:13:46.299 --> 00:13:49.419 binding to a specific lipid, or proteins clustering 00:13:49.419 --> 00:13:52.220 together, or maybe it's the membrane changing 00:13:52.220 --> 00:13:54.980 shape. The machine learning model will look at 00:13:54.980 --> 00:13:57.039 all these configurations from the macro scale 00:13:57.039 --> 00:13:59.700 and say, OK, this one's interesting. This one's 00:13:59.700 --> 00:14:02.539 not so interesting. OK. And then MOMA -I uses 00:14:02.539 --> 00:14:05.139 that to decide which regions to simulate in more 00:14:05.139 --> 00:14:08.039 detail. So it's like a priority system. Exactly, 00:14:08.240 --> 00:14:10.519 yeah. And this is really important because it 00:14:10.519 --> 00:14:14.120 helps us use those computational resources efficiently. 00:14:14.460 --> 00:14:16.799 So we're not wasting time on things that aren't 00:14:16.799 --> 00:14:19.549 important. So this must require huge amounts 00:14:19.549 --> 00:14:22.350 of computing power. It does. It does. High resolution, 00:14:22.690 --> 00:14:25.850 long time scale models need a lot of power. Yeah. 00:14:26.570 --> 00:14:29.629 And MamaEye is designed to run on all kinds of 00:14:29.629 --> 00:14:32.149 high performance computing systems, from smaller 00:14:32.149 --> 00:14:35.110 clusters to super computers. Wow. It's been run 00:14:35.110 --> 00:14:37.690 on systems like Sierra at Lawrence Livermore, 00:14:37.909 --> 00:14:40.789 which has hundreds of thousands of CPU cores 00:14:40.789 --> 00:14:44.129 and thousands of GPUs. Wow. It's hard to even 00:14:44.129 --> 00:14:46.269 imagine that much power. How do you manage that? 00:14:46.450 --> 00:14:49.690 Yeah, so Mamamai is designed to use both CPUs 00:14:49.690 --> 00:14:53.230 and GPUs. OK. The macro scale model with those 00:14:53.230 --> 00:14:55.629 equations on a grid, it's really good for CPUs. 00:14:55.750 --> 00:14:58.690 And we can get almost a millisecond per day with 00:14:58.690 --> 00:15:02.129 thousands of cores. Wow, that's fast. And then 00:15:02.129 --> 00:15:05.330 the CG and AA simulations with all those particles, 00:15:05.370 --> 00:15:09.409 they're better for GPUs. OK. So CG. can get a 00:15:09.409 --> 00:15:13.429 microsecond per day on one GPU, AA, about 14 00:15:13.429 --> 00:15:17.269 nanoseconds per day. So by using both, we can 00:15:17.269 --> 00:15:19.350 handle all those different scales. Right. And 00:15:19.350 --> 00:15:21.509 with thousands of simulations running, how do 00:15:21.509 --> 00:15:24.210 you keep track of everything? Yeah. So we have 00:15:24.210 --> 00:15:27.379 the Workful Manager. or WM, and this is like 00:15:27.379 --> 00:15:30.120 the control center. It monitors resources, it 00:15:30.120 --> 00:15:32.460 starts jobs, it tracks progress, it restarts 00:15:32.460 --> 00:15:35.759 things that fail, and it uses these tools like 00:15:35.759 --> 00:15:38.860 Flux or Maestro to handle all the communication. 00:15:39.000 --> 00:15:40.480 So it's like the conductor of the orchestra. 00:15:40.740 --> 00:15:42.940 Exactly, yeah. And all these simulations must 00:15:42.940 --> 00:15:45.620 generate a ton of data. Oh, they do, hundreds 00:15:45.620 --> 00:15:48.220 of terabytes, sometimes even petabytes. How do 00:15:48.220 --> 00:15:50.799 you manage that? So we need special storage solutions, 00:15:51.039 --> 00:15:52.860 special analysis tools, and a lot of this data 00:15:52.860 --> 00:15:55.700 is made public through server. in a repository 00:15:55.700 --> 00:15:57.980 so other researchers can use it. That's great. 00:15:58.200 --> 00:16:00.940 Yeah. So how has MUMI actually been used to study 00:16:00.940 --> 00:16:04.139 this RAS -RAF system? Yeah, so studying RAS -RAF 00:16:04.139 --> 00:16:07.179 has really been a key focus for MUMI. OK. This 00:16:07.179 --> 00:16:09.820 pathway is so important for cell growth. And 00:16:09.820 --> 00:16:12.299 when it goes wrong, it causes cancer. Right. 00:16:12.379 --> 00:16:14.600 So we really need to understand how these proteins 00:16:14.600 --> 00:16:16.879 work and how they interact with the membrane. 00:16:16.899 --> 00:16:19.399 OK. And MUMI is perfect for this because it can 00:16:19.399 --> 00:16:21.659 handle those atomic details and those long time 00:16:21.659 --> 00:16:24.330 scales. OK, so what have you learned? Well, one 00:16:24.330 --> 00:16:25.889 of the biggest things is that we've been able 00:16:25.889 --> 00:16:29.070 to identify those lipid protein fingerprints. 00:16:29.789 --> 00:16:32.049 These are those patterns of lipids that surround 00:16:32.049 --> 00:16:35.549 the protein, and they can affect how it's oriented, 00:16:35.610 --> 00:16:39.169 how it binds to other molecules. OK. So by simulating 00:16:39.169 --> 00:16:42.710 both RES on its own and the RES -REF complex 00:16:42.710 --> 00:16:45.149 on the membrane, we've learned a lot about how 00:16:45.149 --> 00:16:47.149 they interact with the lipids. OK. Can you give 00:16:47.149 --> 00:16:49.629 me some more specific examples? Sure. So for 00:16:49.629 --> 00:16:51.950 example, we've looked at how a specific region 00:16:51.950 --> 00:16:56.179 of REF, cysteine -rich domain, or CRD, moves 00:16:56.179 --> 00:16:59.759 when it binds to RAS. This is important for understanding 00:16:59.759 --> 00:17:03.159 how RAF gets activated. We've also looked at 00:17:03.159 --> 00:17:07.460 how the orientation of the RAS rave complex affects 00:17:07.460 --> 00:17:10.440 how it interacts with lipids. And we've even 00:17:10.440 --> 00:17:13.900 looked at specific RAS variants like KRS4B, which 00:17:13.900 --> 00:17:16.200 is known to be really difficult to target with 00:17:16.200 --> 00:17:19.039 drugs. And we've been able to simulate all this 00:17:19.039 --> 00:17:21.920 on realistic membranes. So it's really relevant 00:17:21.920 --> 00:17:23.700 to what's happening. in a cell. Okay, so really 00:17:23.700 --> 00:17:25.660 connecting those details to the bigger picture. 00:17:26.200 --> 00:17:28.819 Exactly, exactly. And that's the power of this 00:17:28.819 --> 00:17:31.019 multi -scale approach to get that full picture. 00:17:31.400 --> 00:17:33.559 So what are some of the key findings from all 00:17:33.559 --> 00:17:36.839 this? Yeah. So one surprising thing is that RAS 00:17:36.839 --> 00:17:38.779 proteins are not just sitting there passively 00:17:38.779 --> 00:17:41.380 on the membrane. OK. They can actually change 00:17:41.380 --> 00:17:43.619 the lipids around them. Interesting. So they're 00:17:43.619 --> 00:17:45.960 shaping their environment. Exactly. Yeah. And 00:17:45.960 --> 00:17:48.799 we've also seen that RAS proteins like to cluster 00:17:48.799 --> 00:17:52.119 together. And the way they cluster is influenced 00:17:52.119 --> 00:17:54.720 by those lipid fingerprints. So the lipids might 00:17:54.720 --> 00:17:57.759 be controlling how RAS signals. So the membrane's 00:17:57.759 --> 00:17:59.920 not just a backdrop. It's an active participant. 00:18:00.140 --> 00:18:03.579 Yeah, absolutely. And we've also seen how RAS 00:18:03.579 --> 00:18:06.539 changes its orientation when it binds to REF, 00:18:06.940 --> 00:18:08.900 which then affects how it interacts with lipids. 00:18:08.960 --> 00:18:11.559 OK. And it turns out that RAS has different lipid 00:18:11.559 --> 00:18:13.180 fingerprints, depending on whether it's bound 00:18:13.180 --> 00:18:15.619 to REF or not. So it's like the lipids are telling 00:18:15.619 --> 00:18:19.079 us what state RAS is in. Yeah, exactly. And we've 00:18:19.079 --> 00:18:21.559 also learned a lot about how the race rough complex 00:18:21.559 --> 00:18:24.240 forms and how it moves. OK. The Continuum model, 00:18:24.279 --> 00:18:27.400 when we use data from the CG simulations to refine 00:18:27.400 --> 00:18:31.140 it, it can actually accurately show how RAF binds 00:18:31.140 --> 00:18:34.799 and unbinds to RAS. And the all -atom simulations 00:18:34.799 --> 00:18:37.519 have shown us very specific lipid -dependent 00:18:37.519 --> 00:18:39.859 changes in the complex, which we can then use 00:18:39.859 --> 00:18:42.359 to improve the CG model. So again, that back 00:18:42.359 --> 00:18:44.940 and forth. And the CG simulations have even hinted 00:18:44.940 --> 00:18:47.099 that you can get these larger clusters of RAS 00:18:47.099 --> 00:18:50.619 and RAS. So it's really a complex and dynamic 00:18:50.619 --> 00:18:52.920 picture. And this all gives us a much more accurate 00:18:52.920 --> 00:18:55.359 understanding of these interactions. Yes, absolutely. 00:18:55.519 --> 00:18:58.519 And by using both the continuum and CG models 00:18:58.519 --> 00:19:00.769 together, we can actually make the simulations 00:19:00.769 --> 00:19:03.750 run faster. Oh, really? Yeah, so the CG simulations 00:19:03.750 --> 00:19:06.230 reach equilibrium faster. Okay, so it's more 00:19:06.230 --> 00:19:09.309 efficient. More efficient, yeah. So all of this, 00:19:09.309 --> 00:19:13.009 I think, shows how powerful MUMi is and how it 00:19:13.009 --> 00:19:15.210 can be used to generate these hypotheses that 00:19:15.210 --> 00:19:17.990 we can then test in the lab. This has been a 00:19:17.990 --> 00:19:20.640 fascinating deep dive, so... What's the key takeaway? 00:19:20.920 --> 00:19:22.920 Well, I think the key takeaway is that mummini 00:19:22.920 --> 00:19:25.660 is a major step forward in computational biology. 00:19:25.839 --> 00:19:29.140 OK. Its design and how it uses machine learning 00:19:29.140 --> 00:19:32.140 lets us study these complex systems across these 00:19:32.140 --> 00:19:35.180 huge ranges of time and size. And what it's told 00:19:35.180 --> 00:19:37.420 us about Rostroff is really exciting, right? 00:19:37.519 --> 00:19:40.779 It is. It's giving us this really detailed understanding 00:19:40.779 --> 00:19:42.880 of how these proteins work, which is important 00:19:42.880 --> 00:19:45.400 for understanding cancer and potentially for 00:19:45.400 --> 00:19:48.779 developing new drugs. So looking ahead, what's 00:19:48.779 --> 00:19:51.210 next? for MummyMai? Well, I think we can expand 00:19:51.210 --> 00:19:53.450 it to study all kinds of other systems, other 00:19:53.450 --> 00:19:57.230 proteins, other signaling pathways. We can use 00:19:57.230 --> 00:19:59.130 even more advanced machine learning, like deep 00:19:59.130 --> 00:20:02.170 learning or reinforcement learning. And I think 00:20:02.170 --> 00:20:04.869 this could really change how we develop drugs, 00:20:05.390 --> 00:20:08.210 especially for targets like RASROF that are so 00:20:08.210 --> 00:20:10.450 difficult. It sounds like these machine learning 00:20:10.450 --> 00:20:12.250 -driven simulations are going to be essential 00:20:12.250 --> 00:20:14.910 tools going forward. Oh, I think so, yeah. I 00:20:14.910 --> 00:20:17.529 think with the way computing power is increasing 00:20:17.529 --> 00:20:19.910 and how these methods are getting better, this 00:20:19.910 --> 00:20:22.589 is going to be the standard way to study complex 00:20:22.589 --> 00:20:25.190 molecular systems. And maybe we'll even see fully 00:20:25.190 --> 00:20:27.490 automated simulations in the future. Oh, I think 00:20:27.490 --> 00:20:32.369 that's very likely. Thinking about all this complex 00:20:32.369 --> 00:20:35.369 dance of proteins and their environment makes 00:20:35.369 --> 00:20:38.089 you wonder, how will this change how we treat 00:20:38.089 --> 00:20:41.109 diseases? Maybe we'll go beyond targeting just 00:20:41.109 --> 00:20:43.410 the proteins themselves and start thinking about 00:20:43.410 --> 00:20:46.549 the whole context. It's really fascinating. Thank 00:20:46.549 --> 00:20:48.049 you so much for joining us. This has been a great 00:20:48.049 --> 00:20:50.450 deep dive. My pleasure. And for everyone listening, 00:20:50.970 --> 00:20:53.829 keep thinking about these questions. And we'll 00:20:53.829 --> 00:20:54.430 see you next time.