Scientist Stories: Viviana Gradinaru, Machine Learning Assisted Directed Evolution

Presented by Caltech. Yeah, so the next series of talks will be on applications of AI in different areas, and we're pleased to have Viviana Gradinaro talk about machine learning for directed evolution of viruses. So, here we are. Great. Thank you for choosing to spend the morning with us. And I will try to convince you in less than 30 minutes that biology is not as wet and mushy and unproductive as it might seem to those trained in the exact sciences. And I came to appreciate that when I moved from physics to neuroscience in undergraduate at Caltech. I was an undergraduate here early 2000s. And I thought that the more exact training that I got in engineering and physics was actually a wonderful starting point to address some very difficult challenges in the biological sciences, and especially in the neurosciences. So within my group at Caltech, we study deep brain stimulation and also perform deep brain stimulation of the best kind on this wonderful team. We have a lot of fun tackling difficult challenges. Mainly we focus on how deep brain stimulation can affect behavior and improve outcome in disease. And we focus on motor circuits, circuits that when tampered with can affect locomotion. And as a base for that, we are using the learnings broadly in neurological and psychiatric indications. However, every time we try to do biological experimentation, we bump into wall after wall. And many times this has been technological, the lack of appropriate tools, but more so recently, The walls are of computational nature. We generate a lot of data, so it's one challenge just to plain analyze the sheer amount of data. But the other one has to do with the low throughput experimentation that actually doesn't allow us to map the full space, the full space of neuronal circuits, the full space of proteins. And I have to say, I started as a skeptic of machine learning, and I converted out of necessity. So I will show you with concrete success stories, how it helps some of our difficult problems, and how I think it can actually— it's just the beginning. So what we do in the lab, we focus on 3 core technologies that address key challenges in the neurosciences. One is to control— it's not— can you see the cursor? Great. One is to control neuronal circuits with very high precision. An electrical stimulator, a deep brain stimulator, It's indiscriminate, affects inhibitory circuits, excitatory circuits in the same manner. It's highly inadequate and can cause side effects. So what we've developed over years, starting with my graduate work at Stanford, was to develop light-sensitive molecules that can control neuronal activity with very high temporal dynamics. And this is known now as optogenetics, and we use it to turn on and off neuronal circuits and influence behavior. To do that though, we need to know those circuits. So in order to fully map those circuits in the human brain is quite difficult because we usually map it by optical methods, but the human brain is highly scattering due to the fat content. So what we had to do was to add yet another technology that has to do with tissue clearing. We remove the fat content and we replaced it actually with a transparent hydrogel, and we were able to make transparent brains or transparent organs, and then we can map this circuitry. So we know the highways, then we apply optogenetics to control the traffic. However, for both of these problems, there is a serious limitation. We need to deliver genes. We need to deliver the genes that paint the morphology of the neurons, or we need to deliver the genes that are gonna have a control on neuronal activity. And this is a very difficult problem. And usually the way we deliver genes is either through transgenic methods, which in the rodent world you can do that. What I'm showing you here is deep brain imaging of activity of neurons in an area called striatum. It's involved in your motivation, reward, and also locomotion. And we are able to do that by delivering activity indicators, either through direct invasive injections or through transgenic means. We had to build our own two-photon microscope in the lab for this, so we built in the lab a lot of hardware and software to be able to access and analyze this data. However, as you can appreciate, this is a very limited field of view, and if anything, it's inadequate in terms of actually capturing the full circuits, the full distributed circuits in the brain that mediate behavior. This has been identified as a problem across neurosciences, and a while ago we had this workshop at the National Academy that pointed to the fact that in order to effectively tackle neuroscience, we need ways to deliver genes to the brain, to the entire brain, ideally, in a non-invasive way. And this is very, this is highly necessary because you wanna be safe, and if you can imagine one way to deliver genes is to be rather invasive. You do have to drill the skull, and you have to do a needle insertion. And to have enough coverage in the large brain, this can have some deleterious effects that confounds even your experimental results. And also for efficiency, you wanna cover large enough volumes in order to have a handle on behavior, and you also wanna be compatible with the newer technologies such as gene editing, editing. You want to be able to package these genetic editors and distribute it throughout the brain in a meaningful way. Where are we now in the neurosciences? We use viral vectors, safe non-replicating viral vectors. We modify them and we take away all of the deleterious viral genomes that are responsible for toxicity and replication, and we replace that with useful genes for our purpose, whether they're genes to excite neurons, inhibit, measure activity. And the way we do that is through direct invasive injections. And there are a few problems here. You get very limited coverage, as you can see here. And there's one more, non-uniform multiplicity of infection. With spot injections, at the injection site you get too many copy numbers and this can be toxic. In a way, it's too few to be effective. You're not gonna have enough of the product that you need for the function. This is a big problem and it points to how we are using these vectors. We got pretty good at changing the cargo, what we package in the virus. We have a few available serotypes, but we haven't explored delivery to its full extent. And the reason why this is quite a pressing need, not only in the research space but also preclinical and clinical, is that I have to disclose this big concern. What I'm showing you here is a mouse, mouse rendering, so rendering of the mouse brain where we can inject to some extent. This is the coverage that we get. This is what we are dealing with in our efforts to make our research relevant for the larger brains. So you can see it doesn't scale. And the only way to scale it is to now move from one injection to maybe 6 or 8 to cover a human hemisphere. Is there a better way? Can we do something about it? And if you are relying on the fact that we already have this embedded microfluidic system in our brain that we can use as delivery, it's the vasculature, right? And the vasculature reaches all parts of the brain and in theory could be an excellent venue to access these neuronal cells that we are interested in. However, there's a catch. The brain is protected from circulation from your vasculature and blood flow, by the blood-brain barrier. It's a good reason for it, it protects us from disease and pathogens, but it also prevents research tools from accessing the brain and also therapies. One example that you might be familiar with is in the case of Parkinson's disorder, the patients are getting L-DOPA rather than dopamine because L-DOPA can cross and convert and be effective in restoring their locomotion. So what can we do about it? How can we actually use the vasculature bypass the blood-brain barrier in order to reach brain-wide. And here is where actually an engineering approach was extremely influential. In the biological sciences and neurosciences, many times you are trained to have so-called smart hypothesis. Read enough, think enough, write the two-liner that makes sense, and then execute it, right? However, this falls short when you don't have enough prior knowledge of the system, and many times that's true. And then you're stuck unless you embrace the different thinking paradigm, right, which comes from engineering. And in engineering, at least the way I was taught, is you have starting points and versions and iterations. You start with something that's good enough for your goal, and then you quickly prototype and you quickly advance it. And you don't overthink it that much and have that beautiful hypothesis that you work at for 5 years to find out it's it's wrong. You quickly change course as the evidence shows that to you. So this is actually a principle behind directed evolution. And at Caltech, we should all be very familiar with the method of directed evolution that our colleague here, Francis Arnold, invented decades ago. And it's been so influential across many areas of research—chemistry, biology, virology. And the two core principles of directed evolution is that you can pretty much design anything you need for a particular purpose if you have a good enough starting point. If from that starting point you can generate many offspring, you can generate a lot of diversity. Sometimes you can have many starting points and breed them, right, to generate diversity. And then the second main point is to have a way to curate that diversity and select the ones that overperform.. And then you take this through rounds of evolution and from a good enough starting point expand the space, sample the space, reduce it, then you end up with a handful of possibilities. Whether they're enzymes or opsins for optogenetics or viral vectors for gene delivery and you see, are you indeed better? And in the biological sciences we can put it to test. So this is exactly what we did. We started with a good enough starting point, which is in the form of adeno-associated viral vector 9. AAV9, I encourage all of you to know about it. And the reason why you should know about AAV9 is that children with spinal muscular atrophy now live longer and better because of gene therapy developed by Avexis Novartis and delivered through AAV9. So this is a safe vector, is a vector that's in, in people and creates a better outcome. We started with AAV9, however, there's a catch. I mentioned it's in the clinics for children because the blood-brain barrier matures, and as it matures, it's not as permeant to this gene delivery vehicle. So in the, in our brains, would not work as well. As we grow older and have more needs, will fail to work at all. So what we wanted to do was to take the starting point, AAV9, make lots of offspring, and put them to the test. However, there is a catch. The viral vector has a, the one we work with, has a beautiful icosahedral symmetry. 6 faces, triangulation number of 1. And to, and it's a complex moiety, right? To steer it away enough from its tropism, you need to change the surface to an extensive degree. And if you only make minor changes, you have not created enough diversity and you didn't steer away enough. So why is this important? I'll give you a number of examples. If you just change 4 amino acids, so these are the 60 units are repetitive, right? So they're symmetrical. If you change 4 amino acids, for each one you have 20 combinations, then you have 160,000 possibilities. With the 160,000 possibilities in the lab, we can have 100% coverage. We can make it, we can measure it, we can rank it, no problem. They're not distinct enough and they're not useful enough. So what we have to do in order to generate this diversity is to make many more changes. And you don't have to go too far. If you go from 4 amino acids to 7 amino acids, now your space to sample is more than a billion. In the lab, we cannot have a billion coverage. It's just impossible. So what we end up doing is partially sample, and it's very likely that in this dark unsampled space, there's better solutions, right? So this is where machine learning comes in, and this is where AI comes in to help us navigate through this dark space and make predictions based on our measurements. We can only sample a few percentages. Based on those percentages and the features of the amino acids and positions, we can generate models of this dark space and then pick out handful of variants and ask, are you better than what we measured? And the answer is that many times that can be true. So I'll show you examples from opsins and what we're doing now in viral vectors. So what could just plain wet squishy biology and data analysis, high-level data analysis, but no learning and no eye, how far could it get us? Okay. So this is what we were able to do with very partial coverage and not looking at the dark space. I'll spare you the molecular details here. So you start for AAV9, which I told you is very efficient in children. It fails to work in adult brains. So this is, it should be a brain, you don't see much because there's not much there. And this is what we are able to do with directed evolution of viral vectors. We are able to generate vectors that cross the blood-brain barrier and now deliver these proteins that can be very useful. They can be anything you want, colors for neurons, activity modulators, activity readout, or gene editing tools. And as good engineers, we didn't stop at generation 1, we extended. And I have to give a big shout out to Paul H. Patterson here. Some of you might have crossed paths with Dr. Patterson, a professor in biology. He was my undergraduate research mentor and a brilliant mind that bridged neuroscience and immunology together. And he had as a goal to cure genetic disorders. And I trained with him in the biological sciences and motor disorders. And these variants are named in his honor because unfortunately he passed away 2014 due to glioblastoma, a very fast progressive and aggressive disorder, unfortunately. However, he lives through many of his trainees and tremendous work that he did. So if you have, if you're curious about neuroimmunology, his reviews are something to read and learn from. What can we do with these vectors? And I'll show you, so you see a mouse that doesn't really move, it's just chilling. It's a healthy mouse that's just chilling. And then you see one that's very exciting moving around. We can do that in a highly precise fashion on only the motor circuits. And we can do this in a non-invasive way because the gene delivery system crosses the blood-brain barrier. And because we package activity modulators that respond to small molecules, small chemicals that also cross the blood-brain barrier. So what you have here is a paradigm that can get closer to non-invasive brain modulation. And we also apply, we deploy this technology in the context of mood control, whether it's reward or anxiety. And I am very excited about this because the level of coverage you need and the level of invasiveness is something that you can see carrying through. However, full disclaimer, these models, this circuits, they're dopaminergics, they also innervate the heart. And when they innervate the heart, you can get this confound of causing arrhythmia. So then the really wish list increases to be not only I want the circuits, but I want to stay away from liver because of liver tox, I want to stay away from heart because of arrhythmia. So how can we do this? And the way to do this is to refine one delivery for positive and negative selection, and I will elaborate the most recent viral work we do towards the end. And the other way is to be very refined with your stimulation or your modulation. What I showed you was a small molecule. It's on for as long as the lifetime of the small molecule is. However, in the neurosciences, we have developed this very high precise temporal control, the options, and we've used them in in many ways as activity modulators to read out voltage as well. They have some capability. However, we were not able to push them towards non-invasive deep brain modulation because they are not highly potent. So in optogenetics, and I want to give you a bit of background because it's very important, these proteins respond to light. They absorb photons to undergo isomerization, and that energy is required required for the channel to open and have ion flow, and this ion flow can shut off neuronal activity or can turn it on with millisecond precision. They're extremely powerful. However, there's a catch because the human brain and any tissue really is scattering. You get scattering of photons and absorption, and what this means is that you're not going to recruit enough volume. So we wanted to bypass this problem. And we wanted to generate sensitive enough opsins that will work with very little photon count in order to be able to recruit larger volumes and in order to be able to pair it with systemic delivery of these genes. And systemic delivery would cover the entire brain, and then in theory you could put your light source distant from the actual opsin area to modulate neuronal activity and behavior. And again, we resorted to directed evolution because we know the structure of channelrhodopsins. And you can take different parents and you can basically chop it into pieces and mix and match. Imagine proteins where you, like Legos, you destroy them and then you recreate from different parents. And then you have a lot of diversity. And that diversity is something extremely fun to play with because now you can change their light color sensitivity. Conductivity, you can change their ion conductance properties. And that's what we did. And here is where machine learning had to come into place out of necessity. In order to measure conductivity of ion channels in this particular case, you need to perform this extremely low throughput technique called patch clamping, wall cell patch clamping. Take a glass pipette with a very thin tip, you poke the cell and then you measure it. And you have to do this every single cell. And for variants, you have to do this for dozens of variants and to have statistical power for dozens of cells. It's extremely low throughput and it allows you to go only through a few variants at a time. So that was the method we were applying in the lab. But fortunately, like many of you in the audience, my students were just not patient enough and they said, no, I'm not gonna spend 3 years patching all of these to get an answer. We need to do better than that. And out of necessity, then we developed this different platform. We said, could we use learning? Could we patch 10% of these variants? Because we make the variants, they're diverse. Could we patch 10% and then feed those data and then predict on the ones we didn't patch and then take them out and measure them? In this case, you need enough data and you need high-quality data. So just this approach alone didn't work without the pre-filter. And the pre-filter was membrane localization. So in this case, you have actually multiple levels that can curate your data. And one intermediate step was to use optical methods, which can be high throughput. So you use a high throughput method to curate and sort away the bad localizers, because the bad localizers are not going to work. It needs to be in the membrane to work. We toss those out, and And then so basically build diversity, measure the properties. Then once you measured, train these datasets and see which ones work. So I'll show you what we are able to do. So use learning. I want to give you a few numbers here. So mainly Gaussian regression. We measure the properties, we plot them. I want to give you some numbers we made after this pre-filtering for membrane localization. Which is high-throughput fluorescence measurement. We sorted it to 120,000 variants that were good localizers and had therefore a high chance of performing. And then we used for training data 100 variants. We fed a few different ones that we had from prior work, 61, got the models, and then we iterated this and we sorted from the dark space that we have not touched experimentally 28 variants, and then we measure those. And much towards our, well, my surprise, because I was skeptical of the method. I wasn't sure, given the complexity of opsins, they're very difficult proteins to handle. Even one amino acid change breaks them. So what we were able to do, and I'm showing you just a few examples, was to generate opsins that are much more sensitive. With a low photon count, they can flux enough ions And we can use them with the systemic delivery to achieve the following. You can, in theory, cover the entire brain. Cortex is involved in many activities of relevance to us. And then you can put your light source on the skull. You don't have to drill the skull, you don't have to be invasive, and you can actually cause meaningful behavior results. So by pairing learning and biology and measurements, we are able to both achieve non-invasive delivery and to achieve highly fluxing ions. If you are interested in both the biology and the machine learning parts, the paper is out in Nature Methods just, I think, this month. And all of the software is on GitHub for our lab. So if you look GitHub, Grudenauer Lab, Francis Arnold Lab, you can find all of the software and the measurements. And you can learn how we did all of this and feel free to reach out by email or otherwise if you want to know more. What do we do now in the lab? We use the same framework for the viral vectors. And we use this because in the gene delivery, many problems are not solved. And we are integrating it in the pipeline knowing that at the end we can measure. And a very big ask is the following. Use this directed evolution pipeline for viral vector engineering in a way that can give you cell type specificity and tissue specificity. And one big ask is variants that are strong in the brain but weak in the liver. The reason is physiological. Anything that you ingest, whether by need or want, it will go to your liver, and most times it's going to mess up your liver because the liver is the filter of the body. If we can design molecules that bypass the liver when that's not needed, it's much better off in in terms of a research tool and also in terms of eventual therapeutic. And let me show you, can we do this? The question is, can we do this, right? Similar paradigm, diversify, rank. So in this case, it's just data analysis, no fancy AI or anything. Rank them by enrichment in different tissues. Red means— so this is log enrichment. Red is high enrichment, blue is not.. And then you can have a lookup table. I want something that's high in the brain but low in the liver. And you can see if you scan the liver data set, there's quite a few that are low in the liver. So what we can do now in the lab, we take them, we measure them. Is it true? And it is. So you can see here sections sagittal this way through, through rodent brain, very high expressing all of the brain sagittal, liver here underneath. This more like flute-shaped, leaf-shaped, you see at least a few variants that are very low in the liver. And again, this comes from directed evolution and disciplined data insight. It can get pretty hairy and complicated because what we get from the data set is not the best. So once we have one, we know because of this dark space that I told you about, we don't know about more than 90% of our data is spit. When we have a good variant, I know for sure that that's not the best of its kind. So what we do experimentally is to take that and then generate diversity around it. But now it turns out that you can generate that diversity in an artificial way. You can generate in an artificial way and plot. So here the size of the circle is the enrichment and the thickness of the bars is their similarity at the molecularity level. And then you can find patterns. So you generate it in an artificial way, and you curate it, and you add protein structure information because some of those combinations cannot exist. So you need to add some sanity check if the chemistry matches. Because in AI, you can make anything you want, right? But it needs to make sense for the chemists. So you add protein structure, you add minimizing energy levels, and you combine this. And we can do this because we modify, as I mentioned, we create a lot of diversity, but we don't modify that many sites. It's just a few spikes on the virus. So we can have some guidance from the structure. And we embed that information. And then the hope is that we can advance and use machine learning so we can go beyond finding a variant that's only in our measured dataset. We can find variants that are not in our datasets and variants that are better, and we expect them to be better. So I will close highlighting a few key folks. Claire Bedbrook did pretty much all of the work on opsin engineering in collaboration. She was co-advised by myself and Frances Arnold. She's now a postdoc at Stanford with Karl Deisseroth. Just an amazingly difficult project. This was started before directed evolution was applied to options and before machine learning came into picture. So she had to build all of these modules and bring them together. It's beautiful work. For Vialar vectors, Ben Deverman now leads a group at the Broad Institute working with many folks in the lab. And I want to highlight David Brown for pushing the machine learning algorithms and framework within our group. So David Brown is a great person to chat with if if you have an interest in the learning part of it. Then of course folks such as Xinghong and Priya that have developed this molecular method. So Priya Kumar, senior graduate student in the lab, she developed the biological platform. And many other folks that contribute to different projects in the lab that I didn't talk about. We work on dopaminergic circuits, serotonergic circuits, sleep, a lot of sleep research happening. So I'm gonna close and just leave this here for your thoughts and notes, and if there's questions, I'm happy to take any. Yes? I was just wondering if you could maybe explain more deeply the mechanism of how the viral vectors that you made, these are like 11, 15, and 17 genes. How do they deliver? How do they deliver? Yes, yes. So you can think of it, let me repeat to see if I, how do the viral vectors deliver new genes? And how do the opsins create new physiology? So first question, because they're separate. So the viral vector, you can think of it as an empty shell. So it's an empty shell and we modify the surface of the shell. Now there are instruction elements to stuff the shell with anything you want, and this happens in the biological part. There are two sequences that recognize different genomes. If you it's like a key and lock mechanism. And if you design the key and lock appropriate, you can stuff the shells that we modify with any DNA. So the way we deliver is to stuff the shells with DNA. And that DNA could code for anything you would want. So that's the delivery part. Now the opsins, the DNA for the opsins could be one type of DNA that goes in the shell. So they're complementary. So the opsins generate proteins. It's a bit confusing because the shell is protein and you stuff it with an instruction for protein. So there's a lot— I'm sorry, there's a lot of jargon there. But the opsin, you could think of it as the cargo for the shell. And the cargo is in form of a DNA that codes for a protein that you want. Now, how the opsin controls the physiology, the opsin is a channel, and when you insert channels in the membrane of neurons, they affect the input resistance, the conductivity, right? If you add channels, then you can cross that membrane easier or harder. And this is the mechanism behind neuronal excitability and action potentials that are the ways the neurons communicate, or neuronal depression when we just silence everything. Did that answer the two? Okay. Okay.