Anselm Levskaya

http://anselmlevskaya.com/

He completed his PhD in synthetic biology. He will tell you about that topic. In his spare time, he built and leads an amazing experimental community in SF called Langton Labs. It's amazing. Building an interface to cells and brains.

With electronics. Electrons respond to small voltage levels. You can use positive and negative feedback to get accurate signals. In the 60s, people started using them in an additional manner. There's also an additional semantics on the top of the network. There are designs by abstractions. Which allowed us to, in the 1970s, build the first microprocessors. We have radically increased the power. Your brain, your brain typically has 100B neurons in them. The neurons are not the fundamental unit of cognition. The more fundamental unit is the synapse, and you have 100T of those. What about circuits? Each one is more complex than a transistor. High state variable system. The major differences between electronic computation and biologic computation is that, for one, speed. Electronics is really fast. The brain is extremely slow. The signals are amplified ion waves. They run at the speed of a car or a bullet. The cycles per second are very different. You have GHz. In the brain, you have about 1 kHz. The major connectivity. The gates. They are really simple components and whatnot. You can stream the components up in a row. They are very fast and reliable. With biological components like neurons and synapses, every computational step is probabilistic. It's very noisy. Chaining these things in a serial manner. All of these logic gates in a digital context. Doesn't work that well. In the brain, all of the components are wired together in a highly parallel way, and not that deep serialy. Computation and data storage are connected all the way in the brain. There's no separation. The other major thing, the real reason that ai hasn't been as impressive to date, as we would have hoped, is that frankly the brain has a lot of fucking computational elements in it. Basically. There's about 300M transistors in the best microprocessor, however the brain has 100T synapses. You can estimate a quick guess as how big your soul is in terms of bit. And it's probably 1 to 100 petabytes. The other major difference is that ... your brain builds itself. These are some movies of neurons. Your brain builds itself. The notion of plasticity stops as an adult, it's nonsense. All of these synapses, some of them die, ..

Protein computation. How do cells compute information? How do they take in information from each other? How do they compute an appropriate response. I'm going to cover what we know these days. Very abstractly. This represents an archetypal enzyme that does something to something else. In this case, the way that information is processed is that you have an enzyme that adds a molecular tag to another protein. There are many different tags. There are many groups that are widely used. They use these flows of tags that are moving down from one protein to another to transmit information. This is important mainly because of allostery, originally discovered by some frenchmen in the 60s. How does biology compute? Proteins are not static. They have interesting dynamics. They are bistable, they have two shapes, two functional shapes. Schematically, here's a classical receptor, a sensor, it's meant to pick up information. It's inactive normally, but when a sugar molecule binds to it, it causes the receptor area to change, and it actually changes it to modify the cell and to get activated. You have a lot of internal signaling going on. Alterations to the proteins.

How do you get this information to propagate? Cross-talk would be deadly. If you responded to the presence of food by running away, or responding to toxins by eating, and you would die. Natural selection would not let you get away with that. You get wires in a cell. You have protiens that bind together like glue. There's a lot of binding domains that mediate this process. They have a shape. Imagine this little red guy. This guy wants to talk to the orange one. It does this because it conforms. You only get the interactions that you want, mediated by the complementary binding domains. You have rapid diffusion with these compatible sticky binding domains. You get virtual information channels. It's mediated by shape complementarity. It's a weird schematic for computation.

The problem is that if we wanted to extract information from this network, the way that people figured this out was breaking the cell. It was all inference. But what if we want to set the system in tact and operating in real time? We have to perturb the system and watch what it's doing. You can't just stick a wire in here. A lot of people have come up with using light. Light is a really awesome physical modality to get information in and out of cells. Cells are primarily transparent. Even whole organisms like zebrafish at the young stage are completely transparent. You can get light anywhere in an organism. The other reason is that we're really good with photonics, and microscopy and lasers are really well developed. When you want it and where you want it, super precisely. Nature has done most of the work. To try to use light as a control mechanism, nature has done this- nature has many eyes. We've had sunlight raining down on it. Any species here has protein sensors for detecting light. You do, you have retinal in your retina. But even tiny little bugs. There's a bacteria, matrommonus, plants are the same thing. It definitely senses it and has to put light into its own biological computation network. A lot of people are starting to sequence these things, and hacking out these genetically-encoded light sensors, and starting to splice it into the networks that we care about in the brain.

It's basically using a little protein called a phytochrome, taken out of green plants, and the phytochrome is a little light switch. Basically it has two allosteric states, two conformational states that flip from one state to another, based off of infrared or red light. This is how plants avoid shade. It's really useful in a general way because the phytochromes transmit this information by binding to a small protein only when it's activated. The way that it works in plants is this protein-protein sticky interaction that you can turn on or off with red or infrared light. All wires in biological computation are mediated by these wires. You can control a huge number of interactions by hooking up these wires the way you want. Here are some cells being hit by red and infrared light. Just by hitting the cell with red or infrared light. There's a stichy surface in the membrane. It falls apart, it goes back in the cytopalsm. You can actually paint a little picture on the membrane. This is bad resolution, this is the walker automaton image. We have really impressive control technologies. You can project using the same tech of the projector, on to any sample, at the cells you're looking at, and control the input light signal at any light or speed. You can watch what happens when, well, see what happens. Furthermore you can do more thins, like doing GFP fluorescence on the membrane. You can recruit not just a fluorescent protein, but one of these signaling domains, you have to bring it out to the membrane, and activate a local process. Chemotaxis and cell shape determination can be used, by targeting the system that causes the cell to grow in one direction, these actin polymers that make it grow forward, you can control cell shape. PhyB-mCh. DHPH-YFP. You can cause the cell to grow out in the direction you want it to grow. I'm not going to go into the details. You can control a lot of internal cellular variables. Even more important optical tech is light gated ion pumps.

Optogenetics. It's using another family of light sensitive proteins. Light-gated ion channels only let in certain channels. The two most useful ones are channelrhodopsin. When you hit it with blue light, it puts positive charge into the neuron that activates the neuron. Halorhodopsin can be activated with yellow light to turn off the neuron. You can program in a spike train by activating these. The brain signaling is sort of digital. The timing of the spikes matter. You can pipe information into the brain. This is done by Carl and Ed Boyden. They are persuing that. You can do some pretty crazy things with this. You can add the channelrhodopsin to a specific region of the brain. You do this with the motor cortex of the mouse. You watch the blue light come on in a second. It's running, it's running, it has to run and it doesn't know why. Blue light activates channelrhodopsin in motor cortex. There are lots of experiments.

Video of mouse running around when blue light is activated.

There is a comprehensive set of proteins. With some clever engineering, you can make an optical readout of a particular variable. Calcium sensor. GFP. An impressive one is from maybe a month ago. It's one of the best calcium sensors that has been made to date. The reason calcium sensors are really important is that when your neurons start spiking a lot. If you can sense calcium fast, you can get a second-resolution readout of neural activity. This is a mouse running. This is what the mouse is thinking. This flashes, the optical signals. The signal indicates more calcium. We don't yet have any good genetically encoded sensors that tell you the spiking code. This is recent work where they had voltage dyes. The key thing is that it can report. They look at the cell and look at the variations at the signals that are correlated with the membrane potential. You can see signals up to 1kHz, and that's amazing. If we can do this performance on a genetically-encoded platform. We can watch neurons and finally see how they transmit information between one and another. A neuron here. Each neuron on average has ten thousand connections on average. 100k to 100M neurons. You have to watch these to really see the way higher-order computations work.

Here are some videos that let you imagine in 3D in real time by using clever optics. Zebrafish brains are transparent. You can look at every neuron in a zebrafish brain, and watch every neuron in the brain in near real-time. You can also control where you want to throw light. There's holographic microscopy that allows you to control what things are happening. There are a lot of tricks that you can go into. They are conductive wires. Typically you don't see those kind of things in a non-metallic polymer. Optical storage is one story. There are lots of stories about how people are re-engineering the brain to be understandable. Mouse brains, zebra fish brains. As they are, they are difficult, but you need to add synthetic hooks into the system, which helps information come in and out.

Nature is a hacker.

Tools for low-hanging fruit in "how to get smarter".